What shipped in AI.

Ling 3.2 Flash is a hybrid reasoning model from inclusionAI that accepts text and image input, with a 1M-token context window and up to 200 images per request. Thinking can...

MetaEncoder v1 produces embeddings for images, video, documents and text that suit search, clustering and comparison. Its decide function also answers closed-set questions about a state with probabilities.

Gemini 4 Hydrogen is a Google DeepMind model built for fast execution in agentic workflows and everyday coding. The announcement cites a score of 51 on the Artificial Analysis Intelligence Index (Argon: 50) and a price of $1.50 per million input tokens.

Gemini 4 Carbon is Google's strongest model for multi-step coding, debugging and open-ended scientific research, using more tokens than Argon to reason and verify. API pricing is introductory at $2 per million input and $10 per million output tokens.

ERNIE 5.5 Flash Preview is the latest foundation model in the ERNIE family, with comprehensive improvements in coding, agentic capabilities, knowledge work, and deep research. It has 280 billion total...

ERNIE-Light-DT-1.0 is based on open-source GLM models and includes research findings from Baidu Search's Lightning large model. It features Baidu Sparse Attention (BSA), a unique sparse attention mechanism developed by...

Muse Ember is the smallest and most efficient model from Meta that’s designed for complex reasoning execution and the work agents repeat all day: tool calls, coding agent subtasks, routing,...

Compact open-weights text-to-speech model with about 100M active parameters. It clones voices from an audio sample, lets you design voices by age, accent and vocal texture, and runs in real time on phones, laptops and in the browser.

Anthropic's Claude Fable 5.2 is reported to rank first on Artificial Analysis and is available through the API at 8 dollars per million input tokens and 40 dollars per million output tokens. Its highest reasoning level, max, failed to return an image in the author's SVG test.

Cola2 is a base text generation model that compresses text into semantic spans, generates continuous latents with a diffusion transformer, and decodes them into tokens. The model card reports a 60.10 average over eight tasks after training on 346.03B source tokens.

Gemini Omni 1.2 Flash is Google's native multimodal video generation and editing model, the successor to Gemini Omni 1.1 Flash. It generates clips up to 15 seconds long with native...

Images 2.6 brings faster image generation, GIF creation, more precise editing, and new API controls across ChatGPT, Codex, and the API.

Gemini Omni 1.1 Flash is Google's native multimodal video generation and editing model, the generally available successor to Gemini Omni Flash Preview. It is suited for text-to-video, image-to-video, first and...

Multimodal model that accepts text and image input and returns text. The model card positions it for agentic and tool-use systems, coding assistants and RAG. Open weights allow fine-tuning and integration into third-party products.

LongCat-2.5 is a native multimodal MoE model with 1.6 trillion total parameters and about 48 billion active per token, built for long-horizon agentic work. It pairs visual understanding with agentic coding for codebase development, full-project builds and 3D tasks.

Open-weights MiniMax model that generates images from text and edits up to 16 reference images in one checkpoint. According to the model card, it outputs images of about 16 megapixels without an upscaler, using 6 sampling steps.

Crouton Alpha is a lightweight model built for scalable coding, agentic, and multimodal workflows. It's ideal for executing agentic sub-tasks, tool use, and document analysis in high-volume, repeated work.

GLM-5.3-FlashX-Legal (GLM-Legal) is Z.ai's legal-domain variant of GLM-5.3-FlashX, a native multimodal model that accepts text, image, and video input and produces text. It adds legal-domain mid-training and post-training with reinforcement...

Mage-VL-Cowork A Full Training Recipe to Learn Hybrid Computer Use from Evolved Harnesses

The supplied sources contain no information about GPT-6.1 Terra. The announcement covers Kimi K3.1 Pro and mentions GPT-6.1 Sol only as a price comparison, so no capabilities or use cases can be derived.

Baidu vision-language model for OCR and parsing of very long documents. Its SR-SWA attention keeps per-step decoding latency and memory nearly constant regardless of document length, according to the model card.

Kakao's decision model answers choice, yes/no and scale questions about text, JSON, chat histories or images, returning a probability for each answer without generating text. It scores 72.50 on Kakao's in-house Korean decision bench.

Open-weight, natively multimodal MoE model with 2.8T total and 104B active parameters, aimed at agentic tasks such as coding, 3D creation and financial research. Artificial Analysis scores it 55 against 58 for Claude Opus 5.5.

5-bit MLX build of LFM3-1.7B for Apple Silicon, run through mlx-vlm 0.7.11 or later. Optional thinking mode via a flag, with 12 listed languages. Loading in mlx-lm is still pending upstream.

MiniCPM-OCR Jev is a 2B multimodal model that scores OCR candidates for text, formula, and table regions and returns a quality score from 0 to 1. The sources describe it as a scorer of OCR output rather than a plain text recognizer, and they give no benchmark results.

WeVisOne-VL is a vision-language model built on Qwen3-VL that unifies open-vocabulary detection, referring localization, counting, pointing, OCR, referring segmentation and alpha matting in one model. The release is intended for research in visual grounding and image perception.

Hy-SimulMT is a real-time simultaneous translation model for Chinese and English speech. It translates audio while the speaker is still talking and can output speech with voice cloning when a TTS service is configured.

Ling3-Tiny-UBTree-B16 is a UBTree speculative-decoding draft model for inclusionAI/Ling-3.0-tiny. It combines a five-layer DFlash backbone with a learned bigram selector to propose multiple candidate continuations for target-model verification.

LFM3-Omni-2B is Liquid AI's first omni-modal model in the LFM3 family: a single end-to-end model that understands text, images, video and audio, and answers in text and speech. It is built for on-device use: one model replaces a separate ASR model, vision-language model, LLM and TTS model.

ERNIE-5.5-Flash-Base is a sparse MoE language model with 280B total and 14B activated parameters, pre-trained on 33.5T tokens. It is the checkpoint before mid- and post-training and serves as a starting point for domain adaptation, instruction tuning and reinforcement learning research.

OCR model with 1.3B parameters for documents such as text, tables, formulas and charts. The model card reports 93.68 on Real5-OmniDocBench and 54.7 on OCRBench v2.

LFM3-VL-2B is a vision-language model for on-device visual understanding. It takes text, images, and video as input, and combines the LFM3 language model with a new in-house vision encoder.

Cosmos3-H-Dreams is a real-time, action-conditioned surgical world model for the CMR Surgical Versius robotic system. It is fine-tuned on CMR Surgical Versius clinical cholecystectomy data and distilled into a causal, few-step student for interactive, streaming video generation.

Native multimodal mixture-of-experts model for coding, agentic workflows, computer use, and cybersecurity. It accepts text, image, and video input with up to 1M tokens of context and returns text. The vendor reports 76.2 on BankerToolBench and 82.2 on CyberGym.

NVIDIA-Nemotron-Labs-Dory1.0-4B-MXFP8 is a looped (latent-recurrent) large language model (LLM) trained by NVIDIA. The model employs a normalized-GPT architecture with interleaved hybrid sliding window attention (SWA) and full attention layers.

Mercury 3 is a diffusion-based language model from Inception, and the lab's most capable model to date. It is suited for agentic workflows and real-time applications where latency compounds, such...

Meta coding model for agentic workflows with 816B total and 42B active MoE parameters, a 1M token context window, and text, image, video and audio input. NVFP4 checkpoints are offered for download on Hugging Face.

Moondream 3 Preview is a vision-language model from Moondream with 9B total parameters and 2B active (mixture-of-experts). It handles visual question answering, captioning, OCR, and visual reasoning over images, with...

Hy-Memory-30B-A3B is a compact long-term-memory writer model (~3B active parameters) developed by Tencent Hunyuan to execute the write path inside external memory frameworks.

Timestral-123M is a time series foundation model for zero-shot forecasting. It supports univariate and multivariate inputs with past-only and known-future covariates and outputs 21 quantiles. The model card reports MASE 0.659 and CRPS 0.448 on GIFT-Eval, and an SQL skill score of 50.18 on fev-bench.

No description yet — the model card or announcement hasn't been summarised.

Model Summary: Granite Switch 4.1 8B is a modular LLM that composes a base model, multiple adapters and an encoder together so that a single hosted endpoint can move between natural-language responses, general and specialized decisions, and speech-to-text.

A compact speech recognition model for Chinese and English that can use hotwords or preceding dialogue to recognize domain-specific terms. The modelcard also reports evaluation on 15 regional dialect test sets, including Cantonese, Minnan and Shanghainese.

Ling3-Flash-UBTree-B16 is a UBTree speculative-decoding draft model for inclusionAI/Ling-3.0-flash. It combines a five-layer DFlash backbone with a learned bigram selector to propose multiple candidate continuations for target-model verification.

Gemini 4 Argon produces strong SVG illustrations and a Blender model, according to Simon Willison's hands-on review. Launch pricing is 2 USD per million input and 10 USD per million output tokens, set to double on December 1.

No description yet — the model card or announcement hasn't been summarised.

Apriel Voice 27B Preview turns spoken requests into text responses and tool calls on a shared Gemma 4 backbone. Speech synthesis comes from an external component, and turn-taking is handled by the application.

Meta announces Muse Wonder and Muse Wonder Full, which it calls its most capable models for personal and professional agentic tasks such as insurance claims, leases, and building apps. Muse Wonder is available via the Meta Model API, Muse Wonder Full only to Muse Code subscribers.

NVFP4-quantized vision-language model for reasoning, coding, tool use, and document and video understanding, distilled from its BF16 version. The model card reports GPQA-Diamond and HLE scores close to the BF16 release.

Micro-World 2.0 generates 352×640 video from a first frame, a text prompt and keyboard/mouse actions. The distilled student denoises each chunk in 4 steps, and the Wan2.2-TI2V-5B base weights must be downloaded separately.

Qwen-Image-2.1-Turbo is an accelerated checkpoint of Qwen-Image-2.1 for text-to-image generation and image editing with 8 denoising steps. It uses the same 7B visual generation architecture and loads directly with QwenImage21Pipeline in Diffusers.

Youtu-Parsing-Omni turns document pages, images, charts, geometry figures, audio and video into one structured JSON output, and the task prompt selects the parsing type. The model card reports 96.96 Overall on OmniDocBench v1.6.

Voxtral Mini 4B Realtime Arabic is a streaming speech-to-text model for Arabic dialects and Modern Standard Arabic. It is designed to handle challenging real-world audio and code-switching, where speakers alternate between languages within the same conversation.

Step 5 Preview is StepFun's flagship model for agentic work, built on a sparse Mixture-of-Experts architecture (27B active / 600B total parameters). It performs strongly in software engineering and professional...

Fast, low-cost model with reasoning effort levels from low to max that cannot be turned off. It costs the same as GPT-6 Luna up to 100,000 tokens, and the price rises above that.

Nemotron 3.5 Transcribe is an open-weight, offline speech and audio model for transcription in 40 languages, with support for recordings up to one hour and a 65,536-token context.

Nano Banana 2.1 (Gemini Nano Banana 2.1) is Google's image generation and editing model on the Flash tier, succeeding Nano Banana 2 and Nano Banana Pro. It improves product recontextualization,...

MiniCPM-V 4.7 is the latest model in the MiniCPM-V series, available in a 35B-A3B variant, with 35B-A3B being the first MoE model in the MiniCPM-V series.

NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB is a vision-language model (VLM) designed to run everyday agentic tasks locally, using a native 2-bit quantization to fit in 24GB of memory and run alongside other applications. It optimizes intelligence and speed per GB.

d1-omni-600M is a 600M parameter decision model built on LFM2.5-Encoder-350M. You give it a state (text or JSON, with images or a voice clip) and a set of named questions.

d1-3B is a 3B parameter decision model built on LFM2.5-VL-3B. You give it a state (text, JSON, images, or a mix) and a set of questions.

DiarizationLM is a Large Language Model framework designed to post-process, correct, and optimize automatic speech recognition (ASR) and speaker diarization outputs.

Ling-3.1-flash-1019 is an upgraded Ling-3.1-flash with 560B total and 25B active parameters, aimed at reasoning and agentic workflows. The model card reports the Artificial Analysis Intelligence Index rising from 41 to 44.

ALoDLM-1.7B is a diffusion language model for text and code. It repeatedly applies shared transformer layers and allocates computation adaptively across token positions. The model card limits it to noncommercial research on math, code and inference-time compute, and reports 77.7 on ARC-Challenge.

Mistral Large 4 is a natively multimodal Mixture-of-Experts model with configurable reasoning and native function calling, aimed at coding agents and cybersecurity work. The NVFP4 variant reportedly matches FP8 behavior with a smaller memory footprint.

No description yet — the model card or announcement hasn't been summarised.

GPT-6.1 Sol Pro is the same underlying model as GPT-6.1 Sol, served with reasoning.mode set to pro for higher-quality responses on complex tasks. Cost note: pro mode spends far more...

The source is a release note for the llm-anthropic 0.30 plugin that lists Claude Sonnet 5.5 as supported. The sources say nothing about the model's capabilities or intended uses.

Flash variant of MiMo-V2.6 with an MOPD upgrade that distills several domain-specialized teachers into one model and aims to reduce repeated tool calls in agent workflows. Sparse MoE with 309B total and 15B active parameters, accepting text, image, video and audio.

pplx-embed-v2-context-9b-preview is a contextual embedding model for document chunks in RAG systems. A document is passed as a list of chunks; the chunks are encoded together, so each chunk's embedding reflects its surrounding context, and one embedding is returned per chunk.

Gemini 3.8 Live with Live Avatar pairs live speech dialogue with near real-time video generation, so an agent can listen, see, and speak through a visual persona. Tool calls run in the background while the conversation continues.

GLM-5.3-Prime is the high-speed variant of Z.ai's GLM-5.3, inheriting its full capabilities while delivering 1.5–2× the output throughput through inference acceleration. It supports text input and output with a 1M-token...

The provided source covers Qwen3.8-Flash-Next, a different model from the Max Prime record, so it cannot describe Max Prime's capabilities.

Aion 3.5 Mini is a multi-model roleplaying and storytelling system from AionLabs, built on the GLM family of models. It is the smaller, lower-cost sibling of Aion 3.5 and uses...

Solar 4 is a model family by Upstage. The Pro variant supports a context window of 512K tokens and is available via OpenRouter under the ID upstage/solar-pro4.

NV-Reason-CT is a vision language model for 3D chest and abdominal CT that generates structured reports with step-by-step reasoning and answers follow-up questions. It is an open research foundation for post-training, not a cleared clinical product.

Space Bunny Alpha is an anonymous large model with blazing-fast inference, strong coding capabilities and native multimodal input support. It delivers adjustable reasoning effort, and a 1M-token context window.

GPT-6 Luna Pro is the same underlying model as GPT-6 Luna, served with reasoning.mode set to pro for higher-quality responses on complex tasks. Learn more in OpenAI's docs: https://developers.openai.com/api/docs/guides/reasoningreasoning-mode

Anthropic's Opus 5.5 costs $4 input and $20 output per million tokens, 20 percent below Opus 5.0, with cache reads 60 percent cheaper. Anthropic says it communicates more clearly and is more token efficient, but a blogger's SVG test at max effort hit the output limit twice.

Speech synthesis model for seven languages that clones a voice from a short reference clip (zero-shot) or designs one from a text description. It also supports style control, inline non-verbal tags and Sichuanese and Cantonese dialect control.

Text-to-speech model with zero-shot voice cloning from a short reference clip, voice design from a text description, and inline tags for breath, laughter or cough. It covers seven languages plus Sichuanese and Cantonese. Inference requires the separate GitHub repository.

The sources only give Grok 4.7's pricing: 2 USD per million input tokens and 6 USD per million output tokens. They do not describe what the model can do or what it is suited for.

Multimodal mixture-of-experts model for agentic coding, document processing and tool-driven workflows with long context. Three of four layers use Gated DeltaNet, one layer uses Qwen Sparse Attention for retrieval across the full context.

The sources give no information on this model's capabilities or intended use. The announcement only notes that MiMo v2.6 Flash and Pro were released the day before.

MiMo-V2.6-Flash-RL is the efficiency-focused RL checkpoint of the MiMo-V2.6 series, handling text, image, video and audio in one model. According to the model card it was trained on coding, general agent and visual tasks, and supports 1M tokens for long repositories and agent runs.

Switchyard routes each request among the models you list in models, using a routing decision from NVIDIA's open-source Switchyard library running on OpenRouter's edge. Pass model: nvidia/switchyard plus two or...

GLM-5.3-FlashX is the high-speed variant of Z.ai's GLM-5.3-Flash, a native multimodal model delivering inference speeds of up to 200 tokens/s. Built on the same hybrid sparse and linear attention architecture...

Text-to-image model with 6B parameters for UI, infographics, posters and other text-rich designs. It generates complete compositions and can output RGBA images with transparent backgrounds.

Decomposes a flattened design image into a requested number of transparent RGBA layers, guided by an input image and a layer plan. The recommended setup is 1024 working resolution, 12 sampling steps, and one CUDA GPU with 80 GiB VRAM.

FrogNano is a compact 4B coding agent from Microsoft, trained with reinforcement learning on about 1,500 synthetic software engineering tasks. It generates structured tool calls for the Leaf harness to work on repository issues, and its patches still need human review.

Realtime-Venus is a 9B audio-visual model for full-duplex voice dialogue. It keeps listening and watching while it speaks, handles interruptions, and can start a response when an event in the video or audio warrants one.

WeVisDoc-4B converts page images directly into structured Markdown with LaTeX formulas and HTML tables. The paper reports an Overall score of 95.38 on OmniDocBench v1.6, ahead of the compared end-to-end parsers.

Qwen-Image-2.1 combines text-to-image generation and image editing in one model. According to its model card, it outputs transparent RGBA images, accepts up to 10 reference images, and aims to preserve identity for people and products.

Open embedding model from Google DeepMind that maps text including code, images, video and audio into one shared 768-dimensional vector space. Designed for on-device use in search, RAG, classification and clustering.

Fugu Ultra v2 is the higher-performance model in Sakana AI's Fugu family. Rather than a single monolithic model, Fugu is a learned multi-agent orchestration system: a language model trained to...

Fugu Max is the cost-performance model in Sakana AI's Fugu family. Rather than a single monolithic model, Fugu is a learned multi-agent orchestration system: a language model trained to route...

Agentic instruct model for research, software building, report generation and security analysis in authorized settings. The model card targets tool use and multi-step tasks with failure recovery.

Multimodal mixture-of-experts model for images and text with up to one million tokens of context. According to the model card, only 8B parameters are active per token during prefill, aimed at input-heavy agentic workloads.

ChatGPT Images 2.5 helps turn your ideas, sketches, and reference photos into more personalized, polished images that better reflect your ideas.

Mercury 2.5 is the fastest reasoning LLM, and the latest diffusion LLM (dLLM) from Inception. Instead of generating tokens sequentially, Mercury 2.5 produces and refines multiple tokens in parallel, achieving...

A multimodal model that reads screenshots at native resolution and outputs normalized coordinates or structured GUI actions for mobile, desktop, and web. It decodes block by block via diffusion instead of token by token.

Mixture of Experts diffusion language model aimed at agentic work such as tool calling and multi-turn use, with up to 128K tokens of context. It activates 1.4B of its 16B parameters and uses Levenshtein editing with DELETE and INSERT tokens for error correction.

GPT-6 Astra is OpenAI's flagship model for demanding end-to-end work. It is suited for advanced analysis, software engineering, deep research, scientific work, and document creation, with particular strengths in long-horizon...

Multimodal model with 124B total parameters and 5.5B active per token. It accepts image and video input and is aimed at agent tasks that involve reading web and software interfaces. The model card reports 42 on the Artificial Analysis Intelligence Index v4.1.1.

Ling 3.0 Flash Sante is a health and medicine-focused mixture-of-experts model from InclusionAI, built on Ling 3.0 Flash with 5.1B active parameters out of 124B total. It is designed for...

EVIE-8B is a visual document retrieval model that stores 4096-dimensional vectors per token and scores relevance with MaxSim. The model card reports nDCG@10 of 66.75 on ViDoRe V3 and says the model serves as teacher for EVIE 4.5B.

No source describes Qwen3.8 Max itself. The available announcement covers the separate Qwen3.8 Flash Next model, so the capabilities and intended use of Max remain unverified.

WeatherNext 3 is Google DeepMind's AI weather forecasting model. It uses real-time satellite data for hourly updates with higher resolution and more precise precipitation forecasts, and is integrated into Search, Gemini, Maps and Google Cloud.

Muse Spark 1.3 Contributor is the cost-efficient contributor tier of Meta’s multimodal reasoning model for experimentation, learning, and early-stage agentic, multi-agent, and coding workflows. It is designed to track information...

Gemini 3.8 Live with Live Avatar pairs native live dialogue with near real-time video, so the persona listens, sees and speaks with lip-sync. Tool calls run in the background while the conversation continues. Available in Gemini Enterprise.

Streaming ASR model that transcribes speech as it arrives and attributes each utterance to a speaker (who said what). Users can supply custom hotwords such as names and technical terms, and the card lists ten languages.

No description yet — the model card or announcement hasn't been summarised.

No description yet — the model card or announcement hasn't been summarised.

Claude Fable 5.1 improves on Claude Fable 5 across the board, with the biggest gains in agentic coding, long-running agentic workflows, and knowledge work: long code refactors, front-end and visual...

Jina OCR V1 is a document parsing model built on DeepSeek-OCR that uses FastMTP speculative decoding for faster page transcription. It scores 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR-Bench.

A single model for text-to-image generation and instruction-guided image editing, with Chinese and English text rendering. The Turbo variant is a distilled version that the model card says runs in 2 to 4 sampling steps.

Qwen-Drive-1.0 retains the architecture of the pretrained Qwen3.5 vision-language model and integrates 3D perception, visual question answering, and motion planning within a unified framework.

Hy4 preview is a Mixture-of-Experts model aimed at software engineering, office and data analysis, game prototyping and hard research questions. In an internal blind evaluation it came out slightly ahead of GLM 5.3 and Kimi K3.

Ling-3.0-flash-Fin is a finance-tuned mixture-of-experts model for multi-step financial research, source-grounded search, and spreadsheet and valuation work with tool use. The model card reports competitiveness with similarly sized models, with particular strength in source selection and tool-intensive financial tasks.

MiniCPM5-2B is a dense 2B Transformer for local use on resource-constrained devices. The vendor claims strengths in coding, math, long context, tool use and agentic tasks, but the truncated card gives no figures.

Natively multimodal model in the GLM-5 series with 320B total and 18B active parameters. The authors say it approaches Claude Opus 4.8 on coding and agentic benchmarks, and it targets efficient long-context serving through a hybrid sparse and linear attention design.

Multimodal mixture-of-experts model aimed at long-context agentic coding and document workflows, with open weights for the Flash-Next preview. It combines Gated DeltaNet with Qwen Sparse Attention for efficient inference over up to 1M tokens.

Open-weight speaker diarization model that determines who spoke when. It supports streaming and offline inference for up to eight speakers, with input buffers as low as 80 ms.

Experimental multimodal model in the DeepSeek-V4 family, built on DeepSeek-V4-Flash with added visual modules. The model card reports gains on multimodal agent tasks over V4-Flash-0731, while text agent performance stays comparable.

- [2026/09/09] 🎉 AuK is now open-source. Code and model weights are publicly available.

GLM-5.3 is a text model aimed at complex coding and long-horizon agent tasks, built on the same base model as GLM-5.2. The model card reports a 50% gain over GLM-5.2 on the in-house Z.ai Code Bench and 84.5 on CyberGym.

No description yet — the model card or announcement hasn't been summarised.

Google describes Gemini 3.7 Flash as its most intelligent workhorse model for coding and agents. Vendor-reported results show gains over Gemini 3.6 Flash, for example 65.3% vs 49.0% on DeepSWE v1.1.

Research checkpoints for causal image-to-video generation that produce videos autoregressively in chunks. Variants cover short and long clips plus camera control, built on Cosmos-Predict2.5 2B.

Grok 4.6 is a model developed by xAI, featuring a context window of 500,000 tokens. It is available in a standard variant.

Seed 2.1 Turbo is a multimodal model from ByteDance Seed for coding and long-horizon agent workflows. It is suited for end-to-end software delivery, multi-step task execution, and understanding visual and...

Seed-2.0 is a model developed by ByteDance Seed, designed for code-related tasks with a large context window.

Namazu is a model from Sakana AI with a 256K token context window (262144). No information is available on parameter count, licensing, weight openness, intended use cases, or capability tier.

Solar 4 is a model family by Upstage. The Pro variant supports a context window of 512K tokens and is available via OpenRouter under the ID upstage/solar-pro4.

Compact 2.4B vision-language model that processes images at native resolution and covers visual question answering, captioning, grounding, OCR, charts and documents. Intended for prototyping and task-specific fine-tuning.

SingProbe is a lightweight streaming guardrail for Ling-3.0-tiny. It scores query intent, response unsafety and hallucination risk per token from the base model's hidden states, with under 0.5% decode overhead.

No description yet — the model card or announcement hasn't been summarised.

A dense 30B open-weight model from Meta for local agentic work, covering multi-step reasoning, tool use and image understanding. According to the sources it runs fully on a single GPU without cloud access.

Dense decoder-only language model from IBM for multilingual use, coding, RAG, tool use and structured JSON output, with a built-in thinking mode for extended reasoning. The sources frame it for enterprise use and customization.

MiniMax Music 3 generates complete songs up to five minutes long from lyrics and a style description, using an 8B global LLM for song structure and a 0.6B local LLM for acoustic detail.

LFM2.5-VL-3B is a multimodal variant of LFM2.5, a family of hybrid models designed for on-device deployment. It builds on LFM2-VL-3B with further mid- and post-training.

Muse Spark 1.2 contributor tier is a reasoning model from Meta designed for developers who want to start building at an even lower cost. It’s meaningfully cheaper than Muse Spark...

Dense 27B vision-language model with native image and video understanding, adjustable thinking depth, and a focus on coding and long-horizon agentic tasks.

Model Summary: Granite Speech 5.0 TurboCTC Non-commercial is a compact 470 million parameter English ASR model with very high inference speed that is released for research and noncommercial purposes only. We invite users to refer to granite-speech-5.0-470m-turboctc for other use cases.

Base checkpoint of the Nemotron 3.5 Lightning family, without SFT, RL or distillation. It is a hybrid MoE model with 30B total and 3B active parameters, intended as a starting point for custom post-trained models.

LongCat-Flash-Lite-Sparse is a non-thinking MoE model with 69B total and about 3B active parameters per token, supporting up to 1M tokens of context. It reports gains over its dense predecessor in agentic coding and tool use, e.g. 68.20 vs 54.40 on SWE-Bench Verified.

▶ Hear it first. Natural turn-taking, barge-in and live tool calling.

Multilingual 750B parameter model (37B active) for reasoning, agentic workflows and long context. Per its model card it is especially strong on long-context retrieval and safety. It covers ten languages and supports speculative decoding (MTP, DSpark) for faster inference.

Intern-S2-Mobius separates knowledge storage from reasoning modules, which the model card says yields shorter reasoning traces and higher throughput. The card reports large gains on scientific tasks, but exact scores appear only in figures.

UEmbed is a multimodal embedding model on a Qwen3.5 backbone that produces dense and sparse vectors for text, image and video in a single forward pass. It targets retrieval, multimodal search and visual document retrieval.

MiniMax-H3 is an omni-modal system that understands text, images, video and audio and generates videos with native stereo audio up to 2K and 15 seconds. The base stage outputs 768p, and a regeneration stage produces the 2K version.

Qwen3.7 Flash is a vision-language reasoning model from Alibaba. It is suited for multimodal agents, visual coding, search, and computer interaction, with strengths in object recognition, spatial understanding, and real-world...

Inkling-Small is an openly available multimodal model that takes text, image and audio input and outputs text. According to the model card, it targets developers building agentic, tool-use, coding assistant and RAG systems.

Mage-ViT is a vision encoder that uses codec signals to keep only motion-salient patches, which makes video processing efficient. This repo ships only the ViT checkpoint without language model training and can serve as an encoder for your own multimodal training.

Mage-VL is a 4B multimodal model for image and video understanding. It pairs a visual encoder trained from scratch with a Qwen3-4B language backbone. The modelcard reports over 75% fewer visual tokens and up to 3.5x faster inference than uniform frame sampling.

The source lists only Claude Opus 5's price of $5 per million input and $25 per million output tokens. It does not document Opus 5's capabilities or intended uses, because the article covers Opus 5.5.

Compressed speech recognition model derived from VibeVoice-ASR for real-time transcription on CPUs without a GPU. The model card reports RTF below 1 from 3 threads at a total size of 1.58 GB.

No description yet — the model card or announcement hasn't been summarised.

Hy-MT2 is a family of multilingual translation models covering 33 languages that follow translation instructions given in several languages. The 1.8B variant can be compressed to 440 MB with 1.25 bit quantization for on-device use.

SingProbe is an intrinsic streaming guardrail for Ling-3.0-flash that scores query intent, response unsafety and hallucination risk at every generated token. It reuses the base model's hidden states and adds under 0.5% decode overhead.

No description yet — the model card or announcement hasn't been summarised.

Open-weight MoE coding model with 35B total and 3B active parameters, built for agentic coding and text-only use. The model card reports 69.40 on SWE-bench Verified, a vendor figure.

INT4 weight-only quantized (W4A16) SFT variant for CPU inference on AMD EPYC with vLLM and ZenDNN. Only dense and shared-expert layers are quantized, while routed experts stay BF16, so the size drops only from 29.5 to 28.9 GiB.

Upstage open-weight MoE model (250B total, 15B active per token) aimed at agentic tasks such as tool calling, document work and coding. The model card reports 70.4 on SWE-Bench Verified, behind MiMo-V2.5 (73.0) and DeepSeek-V4-Flash (73.8).

Workhorse model for agents, coding and knowledge work. According to the Artificial Analysis Index it uses 17% fewer output tokens than 3.5 Flash. Priced at $1.50 per 1M input and $7.50 per 1M output tokens.

Google's fastest and most cost-effective 3.5-class model, aimed at scaling AI agents. The Artificial Analysis Index reports 350 output tokens per second, and the announcement says it significantly outperforms earlier Flash-Lite generations in agentic workflows.

Cosmos-H-Dreams is a real-time, action-conditioned generative surgical world model that lets a human operator or a learned surgical-robotics policy act inside a synthesized surgical scene and observe the interactions live.

Motif 3 Base is a pretrained base checkpoint without instruction tuning, intended as a starting point for fine-tuning, continued pretraining and research. The mixture-of-experts model has 314B total parameters, with 13.2B active per token.

Gemini 3.5 Flash Cyber is a lightweight model fine-tuned from 3.5 Flash to find, validate, and patch software vulnerabilities. Google says it outperforms its mainline Flash models on these tasks at lower cost than larger cybersecurity models.

Muse Spark 1.1 is a multimodal reasoning model from Meta, built for agentic tasks. It accepts text, images, video, audio, and PDF documents and returns text, with a 1M-token context...

Shieldstral is a compact 3B-parameter, policy-adaptive multimodal safety classifier. Instead of predicting a fixed set of moderation categories, Shieldstral evaluates content against a safety policy expressed in natural language and returns a single continuous safety score.

Agent-oriented diffusion language model with Levenshtein Editing for tool use over a 128K context, multi-turn interaction and error correction. The model card reports 49.28 on SWE-bench Verified and 80.33 on τ²-Bench.

PII-Tracer detects personal data in chat conversations. It tags nine PII categories with BIOES labels and adds a conversation-level sensitivity score. The repo ships weights only, so an MLX implementation of the encoder and heads must be provided separately.

Multimodal model from Shanghai AI Lab for scientific reasoning and long-horizon agent tasks. According to the model card, it learns directly from raw scientific literature pages and was trained with reinforcement learning across more than 20 scientific domains.

The sources contain no description of the model, only a download pointer. The name and pipeline class suggest animation, but the sources do not confirm it.

Inkling is a multimodal mixture of experts model that accepts text, image and audio input and returns text. Its model card targets agentic workflows, tool use, coding and chat applications, and it is released with open weights.

Hy-Embodied-VLM-1.0 is a vision-language model for embodied agents with about 3B active parameters out of roughly 30B total, focused on action planning and physical-world understanding. The model card reports first place on 19 of 38 embodied benchmarks.

RxBrain pairs language reasoning with visual imagination: it answers questions about images and video, predicts near-future frames, and plans subgoals together with goal images. Image generation additionally requires the external FLUX VAE.

NVIDIA-Ising-Calibration-1.5-31B-BF16 is a dense multimodal vision-language model built on Gemma 4 31B.

Wan-Dancer generates long, rhythmic dance videos from a reference image and a music track. It plans global keyframes first and then refines locally, using the full track to keep the choreography coherent.

KAT-Coder-Pro V2.5 is a flagship-level Agentic Coding model that can directly hand over an entire issue or an entire business workflow to it, allowing it to autonomously locate and make...

According to the announcement, GPT-5.6 Luna Pro belongs to the GPT-5.6 family, which is available in the Kiro coding agent for planning, building and reviewing software. The source gives no Luna-specific capability or benchmark data.

Grok 4.5 is a model from SpaceXAI with frontier performance on coding, knowledge work, and STEM.

Aion-3.0 Mini is a multi-model roleplaying and storytelling system from AionLabs, built on the DeepSeek family of models. It uses a collaborative generation process in which multiple specialized models each...

Generates a temporally coherent MP4 video from one or more input images and an optional text prompt, aimed at Physical AI uses such as robotics, autonomous driving and smart spaces. It is a 4-step distilled version of Cosmos3-Super-Image2Video.

Text-to-image model from NVIDIA, distilled to 4 sampling steps via DMD2 without classifier-free guidance. The modelcard reports up to 25x fewer diffusion evaluations per image than the 50-step base model with CFG.

Compact dense audio-text LLM for audio understanding, speech recognition and translation, text-to-speech and speech-to-speech generation, with thinking and instruct modes. Text abilities are said to stay close to its backbone.

Large MoE language model with 1.6 trillion total parameters, about 48 billion active per token, aimed at coding and agentic tasks with a 1M-token context. The model card says it works with harnesses such as Claude Code.

PAR checkpoint list: a 400M model for the main benchmark, a PDB fine-tune, and a 60M model for zero-shot motif scaffolding. The excerpt does not describe the tasks in detail.

MLX-compatible export of Laguna S 2.1 with NVFP4 4-bit quantization. The supplied sources do not describe its capabilities; the main model card is referenced for details.

Leanstral 1.5 is an open-source code agent model designed for Lean 4, a proof assistant capable of expressing complex mathematical objects such as perfectoid spaces and software specifications like properties of Rust fragments.

A 4B-parameter NVIDIA world model that generates video, images, actions and text from text, image, video and action trajectory inputs, aimed at robotics, autonomous vehicles and smart spaces. The sources contain no benchmark results.

Sonnet 5 is Anthropic's most capable Sonnet-class model, with frontier performance across coding, agents, and professional work. It supports adaptive thinking with selectable reasoning effort levels (low, medium, high, max,...

The source describes Google Pics, an image tool in Google Workspace built on the Nano Banana model. It lists text-prompt image generation, object isolation and text editing inside images, but does not mention the Lite variant.

Turns document images into structured text with layout classes, bounding boxes and reading order. Aimed at OCR, table and chart extraction, and data preparation for RAG pipelines.

Vision language model for captioning, visual question answering, open vocabulary detection and pointing, all returning structured output. It runs locally via Photon on NVIDIA Ampere or newer GPUs and Apple Silicon Macs without an API key.

TabFM is a zero-shot foundation model for classification and regression on tabular data with mixed numerical and categorical columns. Training examples are passed as context in a single forward pass, with no fine-tuning, and the model card reports it outperforming tuned gradient-boosted tree baselines on TabArena.

Remote sensing foundation model that processes Sentinel-1, Sentinel-2 and Landsat images and time series. It produces embeddings for downstream tasks such as segmentation and similarity search, and the model card describes fine-tuning for these uses.

Sakana AI's higher-performance Fugu model — not a monolithic LLM but a learned multi-agent orchestration system that routes requests across multiple models and tools. Multimodal (text+image input), 1M context, $5/$30 per MTok on OpenRouter.

Nemotron-Labs-3-Puzzle-75B-A9B is a deployment-optimized large language model developed by NVIDIA, derived from Nemotron-3-Super-120B-A12B.

Qwen-AgentWorld is the first language world model to cover seven agent interaction domains within a single model. It simulates agentic environments via long chain-of-thought reasoning, predicting the next environment state given an agent's action and interaction history.

GLM-5.2 is an open-source model from Z.AI for long-horizon tasks, with a stable 1-million-token context and strong agentic coding. It combines code generation with tool use and supports adjustable thinking-effort levels. It scores 62.1% on SWE-bench Pro, sits close to the frontier on AIME and GPQA-Diamond, and improves markedly over GLM-5.1. The MoE architecture with IndexShare cuts compute per token by a factor of 2.9.

Laguna XS 2.1 NVFP4 MLX is a 4-bit MLX build of a mixture-of-experts model with 33B total and 3B active parameters, aimed at agentic coding on local machines. The model card marks this build as experimental and not intended for production use.

Baidu's OCR vision-language model parses document images, PDFs and multi-page documents into structured text. The model card positions it as a step beyond DeepSeek-OCR and lists Transformers, vLLM and SGLang for inference.

Nano Banana Pro is a Google DeepMind image model. The announced Google Pics Workspace tool is built on the Nano Banana image model and generates and edits images from text prompts, including isolating objects and editing or translating text inside images.

No description yet — the model card or announcement hasn't been summarised.

Google's latest Gemini Flash generation with native image input/output. Available on OpenRouter ($0.50/$3 per MTok). Gemini 3.x is a standalone series alongside Gemini 2.5.

Microsoft's specialized code-search model for coding agents — powers the 'Explore' subagent in SWE-FastContext. Not a general-purpose LLM, but relevant for agentic setups.

North Mini Code is an open-weight sparse MoE model (30B total, 3B active) specialized for code generation, agentic software engineering, and terminal tasks. It supports tool-use through chat templates and was post-trained with supervised fine-tuning followed by reinforcement learning with verifiable rewards focused on agentic coding.

A 4B vision language model that forecasts the 3D trajectories of user-selected 2D points from a short frame history and a language action instruction, covering about two seconds at 15 fps. It is intended for motion forecasting research and as a base for finetuning, such as robot pick and place.

Kimi K3 is an open-weight multimodal model with 2.8T total parameters (104B active) and a 1M token context window, built for long-horizon coding and agentic knowledge work. Moonshot claims about 2.5x better scaling efficiency than Kimi K2.

Kimi K2.7 Code is Moonshot AI's coding-specialised agentic model. It accepts image input and tool calls and can handle complex software-engineering workflows end to end. It scores 62.0 on Kimi Code Bench v2 and uses about 30% fewer thinking tokens than its predecessor K2.6. It trails GPT-5.5 slightly on coding benchmarks but beats Claude Opus 4.8 on MCP Mark Verified with 81.1.

VISTA-9B is a vision-language model for GUI grounding that maps a screenshot and an instruction to a click coordinate in a 0-1000 frame. The model card reports 68.1 on OSWorld-G, against 63.1 for its Qwen3.5-9B base model.

LFM2.5-Encoder is a family of multilingual bidirectional encoders built on the LFM2 architecture, available in two sizes:

No description yet — the model card or announcement hasn't been summarised.

GGUF weights for ZONOS2, a text-to-speech model with voice cloning that runs as a native C++ pipeline without Python or PyTorch. Q8_0 is the recommended default, Q4_K the smallest option when VRAM is tight.

EvoQuality scores image quality without a reference image and without human quality labels. It compares image pairs and rates single images from 1 to 5. The model card reports a weighted average PLCC of 0.770 and SRCC of 0.726 across benchmarks.

UltraX is a refinement model for web pre-training data. It predicts structured edit calls, such as removing line ranges or replacing substrings, which are applied deterministically to the original text.

Anthropic's new top flagship — replaces the Opus tier as the strongest generally available model. Arena Elo 1508 (rank 1). 1M context, Adaptive Thinking always on, $10/$50 per MTok. Fable 5 + Mythos 5 (invite-only) launched together on June 9, 2026.

Open-weights multimodal model from Google DeepMind that takes text, image and video input and outputs text. It denoises 256-token blocks in parallel via diffusion instead of generating token by token, aimed at speed. The model card reports most benchmark scores below those of Gemma 4 26B A4B.

SCAIL-2 animates a reference character from a driving video, supports character replacement and multi-character scenes, and works without intermediate pose representations such as skeleton maps.

PP-OCRv6: From 1.5M to 34.5M Parameters, Surpassing Billion-Scale VLMs on OCR Tasks

PP-OCRv6: From 1.5M to 34.5M Parameters, Surpassing Billion-Scale VLMs on OCR Tasks

NVIDIA's powerful open-weights MoE model: hybrid LatentMoE + MTP layers, 1M context, 550B/55B active. Free on OpenRouter. Focus: complex multi-agent workflows, code, math, science.

NVIDIA Nemotron 3.5 Content Safety is a compact 4B-parameter multimodal guardrail model from NVIDIA, fine-tuned from Google Gemma-3-4B. It moderates both inputs to and responses from LLMs and VLMs, accepting...

ArtiFixer refines and extends renderings from imperfect 3D reconstructions with a few-step causal diffusion model. It is released for research and development only.

397B model from the Shanghai Innovation Institute — open source, Apache 2.0, free on OpenRouter. Agentic-focused, with papers on a 'Unified Ecosystem for Large-Scale Environment Construction'. 7.87K HF likes despite barely any Western coverage.

Qwen's latest Plus tier — 1M context, $0.32/$1.28 per MTok. Qwen3.7 Max (stronger) and Plus (more efficient) form the top of the Qwen3.7 line.

LFM2.5 is a hybrid model designed for on-device deployment, featuring 230 million parameters and optimized for fast edge inference. It is suitable for general-purpose text generation and agentic tasks, but not recommended for reasoning-heavy workloads.

MiniMax is barely known in the West — but its M3 model on OpenRouter ($0.30/$1.20 per MTok) is one of the cheapest 1M-context multimodal providers. Successor to Text-01 (4M context).

No description yet — the model card or announcement hasn't been summarised.

DeepSeek's heavyweight: 1.6T total, 49B active, 1M context, FP4/FP8 mixed precision. Codeforces 3206 Elo beats GPT-5.4. MIT license, fully open weights. $0.435/$0.87 per MTok on OpenRouter.

Video renderer from ByteDance Seed, fine-tuned from Wan2.1-1.3B for video generation and editing. The modelcard reports performance close to the 14B variant on style transfer, subtitle or watermark removal, and local editing, but weaker results on human generation.

Current Opus-tier model (until Claude Fable 5). Arena Elo ~1490. 1M context, Adaptive Thinking, knowledge through Jan 2026. $5/$25 per MTok direct, similar on OpenRouter.

201B multimodal open-weights model from StepFun — underrated in the West. Image+text input, strong reasoning. Successor to Step 3.5 Flash (199B). $0.20/$1.15 per MTok on OpenRouter.

Strongest Qwen3.7 variant: 1M context, $1.25/$3.75 per MTok. More capable than Plus, but pricier. The Qwen3.7 line shipped late May 2026 as the successor to Qwen3.6.

xAI's coding-focused model — 256K context, $1/$2 per MTok. 'Build' implies software development as the main use case. A separate coding line alongside the general-purpose Grok 4.3.

Google's current Flash generation: 1M context, all modalities, $1.50/$9 per MTok on OpenRouter. Gemini 3.5 Flash is the speed tier of the Gemini 3.x series that replaces Gemini 2.5.

Multimodal predecessor of K2.7-Code — Visual Agentic Intelligence with 1.1T parameters. 2.66M HF downloads. Still relevant for multimodal tasks (images, video) since K2.7 is code-only.

Direct predecessor of GLM-5.2 (June 20). Still interesting for local deployment. The GLM-5.1 FP8 variant has 1.33M downloads — massive interest from the CN community.

Google's cheapest Gemini 3.x option: 1M context, $0.25/$1.50 per MTok. The 'Lite' variant is the first choice when token costs are critical and multimodality is needed.

Poolside's strong coding-agent model: 226B MoE, Apache 2.0, SWE-bench Pro 49.2%. Free on OpenRouter! Specialized for agentic long-horizon coding tasks. Poolside is a little-known US AI company.

Poolside's lean coding model: 33B MoE, 3B active, for local deployments. Free on OpenRouter. Sliding-window attention for very fast inference. 231K downloads on HF.

xAI's current frontier model on OpenRouter ($1.25/$2.50 per MTok). Grok 4.3 positions itself as a strong all-rounder. Also accessible via Grok.com / X Premium.

Mistral ironically calls 119B 'Small' — Apache 2.0, with instruction-following, reasoning and coding in one. Mistral positions itself as the strongest European open-weights challenger.

Mistral Medium 3.5 is Mistral's first flagship to handle instruction-following, reasoning and coding in a unified way. 418K HF downloads, EAGLE-acceleration variant available. $1.50/$7.50 per MTok on OpenRouter.

IBM's Granite 4.x generation: 8B, Apache 2.0, $0.05/$0.10 per MTok (one of the cheapest). Enterprise-focused, strong on structured tasks. IBM ships Granite models consistently with open weights.

Moonshot's 'Visual Agentic Intelligence' — 1.81M HF downloads. Basis for K2.6 and K2.7. One of the first large models with real multimodal agentics at 1T parameters.

Cohere's first multimodal open-weights flagship: 218B MoE, 25B active, 128K context, 48 languages, Apache 2.0. Reasoning tokens, tool use with JSON schema. Not on OpenRouter yet.

NVIDIA's small multimodal reasoning model: 30B MoE, only 3B active, text+image+audio input. Free on OpenRouter. Designed for on-device, 4x faster than its predecessor, reasoning ON/OFF mode.

Fast variant of the Qwen3.6 family: 1M context, $0.19/$1.13 per MTok. Qwen3.6 shipped as Flash, 27B, 35B and Max-preview variants — all on April 28, 2026.

OpenAI's current frontier model: 'A new class of intelligence for coding and professional work.' 1M context, knowledge cutoff Dec 2025, $5/$30 per MTok. GPT-5.5 Pro as the stronger variant ($30/$180 per MTok).

Xiaomi's reasoning model on OpenRouter ($0.435/$0.87 per MTok). MiMo is Xiaomi's first serious LLM push — barely noticed in the West. 1M context. V2.5 (lighter) and V2.5-Pro available.

Tencent's Hunyuan 3 in preview on OpenRouter ($0.063/$0.21 per MTok). One of the cheapest models available. Tencent is a heavyweight that gets little Western attention in AI.

Ant Group's (Alipay's parent) 1-trillion-parameter MoE model — Apache 2.0. Extremely cheap on OpenRouter ($0.075/$0.625 per MTok). 472 likes on HF despite barely any Western coverage.

OpenAI's second image-generation iteration on the GPT-5.4 base: $8/$15 per MTok (input/output). The current standard route for professional AI image generation via OpenRouter API.

First release of the GLM-5 generation (April 2026). 2.1K likes on HuggingFace. Zhipu's development rhythm is remarkable: one major release per month.

Google's on-device multimodal model: handles text, image, video AND audio at just 4B effective parameters. MatFormer architecture enables sub-models within the same checkpoint. Designed for edge deployments.

A few generations before Fable 5 — but in thinking mode Arena Elo 1502, rank 3 worldwide. Still relevant for setups where Extended Thinking is desired and Fable 5's Adaptive Thinking isn't enough.

Predecessor of GPT-5.5, but still very widely used ($2.50/$15 per MTok vs. $5/$30 for 5.5). Mini variant (400K context, $0.75/$4.50) for coding agents and subagents. IMOAnswerBench 91.4%.

Alibaba's strong open-weights model from April 2025 — basis for many downstream variants. Qwen3.5 and 3.6 followed in 2026. Often overlooked in the West despite AIME 2025 85.7%.

Meta's multimodal MoE flagship: 17B active, 128 experts, 402B total, 1M-token context. Scout variant (16E, 109B) for lean deployments. Llama 5 is expected in H2 2026.

Google's open-weights flagship: 27B (also 1B, 4B, 12B), 128K context, text+image, 140+ languages. Gemma 3 is the basis for many community fine-tunes. Superseded by Gemma 4.