Model releases
Last scanned 26 September 2026, 18:54 UTC.
Every model this briefing has seen announced, newest first — one row per model rather than one per story, so the six outlets that cover a launch collapse into the release they are all covering. The date is the earliest any source in the sweep reported it, which for a model that shipped before this record opened, or that these sources came to late, is later than the day it launched. It is a record of coverage, not a vendor changelog.
Overview
- 147models trackedsince 21 Jul 2026
- 119in the last 30 days55 in the last 7
- 42of those open-weightsdownloadable, not only served
- 75labswith at least one release on record
- 8seen, not launchedmet through a platform listing
6 Jul 2026this week
The newest from each lab
One row per organisation — its most recent release — with the labs that have shipped most in the last three months first. Pick the labs you follow and this table, and every release below it, narrow to them: the choice is remembered on this device and travels in the address bar, so a bookmark of your own labs is a link.
Every lab — all 147 releases on record.
| Released | Model | Lab | Access | Type | Reported by |
|---|---|---|---|---|---|
| 2 days ago | Zhenwu V900Alibaba unveiled its self-developed Zhenwu V900 AI chip. | Alibaba | — | — | DigiTimes |
| 3 days ago | gemini-3.8-flash-lite-ttsGoogle released gemini-3.8-flash-lite-tts model. | API | audio | Simon Willison | |
| 2 days ago | PUBG AllyIntroduces PUBG Ally, a conversational embodied agent for PUBG that reasons, acts autonomously, and plays as a voice-enabled teammate. | Unknown | research | multimodal | arXiv cs.AI |
| 4 days ago | GPT-6 SolOpenAI releases GPT-6 Sol, cutting API costs by 50% or more. | OpenAI | API | text | VentureBeat AI+5 |
| 4 days ago | Claude Opus 5.5Anthropic is launching Claude Opus 5.5, the first model in a new generation. | Anthropic | API | text | The Decoder+6 |
| 3 days ago | NVIDIA Nemotron 3 DiarizationNVIDIA Nemotron 3 Diarization enables real-time multi-speaker AI. | Nvidia | open weights | audio | Hugging Face+1 |
| 3 days ago | InternW0InternW0 is the first instantiation of the InternW physical world model series from Shanghai AI Laboratory. | Shanghai AI Lab | — | world | arXiv cs.AI |
| 4 days ago | MiMo-V2.6-ProXiaomi released MiMo-V2.6-Pro as the top-performing open-weight model in the world. | Xiaomi | open weights | text | VentureBeat AI+2 |
| 5 days ago | Grok 4.7Grok 4.7 is rolling out in GitHub Copilot for agentic coding and complex workflows. | xAI | in-app | code | GitHub Copilot+3 |
| 3 days ago | ME-VLMLi Auto released ME-VLM model. | Li Auto | — | text | Pandaily |
| 5 days ago | Meta SAM 3.1Meta announced the SAM 3.1 model. | Meta | — | image | TLDR AI |
| 8 days ago | GLM-5.3-FlashXZhipu AI launched GLM-5.3-FlashX with claimed ~200 tok/s inference on ~100,000 domestic accelerators. | Z.ai | open weights | text | Pandaily |
| yesterday | GLiNER2.5-DecideFastino Labs released GLiNER2.5-Decide, a 340M-parameter open-weight decision model that runs on CPU. | Fastino | open weights | text | Marktechpost |
| 2 days ago | Jev-MobileJev-Mobile is introduced as an executor for mobile GUI agents. | TypeSafe | research | multimodal | arXiv cs.AI |
| 3 days ago | FLUX 3 ActionBlack Forest Labs releases open-weight FLUX 3 Action robotics model. | Black Forest Labs | open weights | robotics | VentureBeat AI+1 |
| 4 days ago | SenseNova U1 ProSenseTime released the production version of its SenseNova U1 Pro image-creation model via its Raccoon app and API. | SenseTime | in-app | image | TechNode+1 |
| 4 days ago | Light-O1-PreviewLight Origins released Light-O1-Preview, a 6B Apache-2.0 model for robot whole-body trajectories. | Light Origins | open weights | robotics | Pandaily |
| 5 days ago | T1PrimeBot launched transformable T1 personal robot for sale. | PrimeBot | — | robotics | Pandaily |
| 13 days ago | Iris-proThe AllSpark team released Iris-mini and Iris-pro, two open-source search agents built on Qwen models. | AllSpark team | open weights | text | The Decoder |
| 2 weeks ago | North Small TranslateCohere released North Small Translate, an open-weight Mixture-of-Experts model built for machine translation across 50 languages. | Cohere | open weights | text | Marktechpost |
| 2 weeks ago | Fugu Ultra v2Sakana AI released Fugu Max and Fugu Ultra v2, two models for multi-agent orchestration. | Sakana AI | API | text | Marktechpost |
| 2 weeks ago | DPT-3 VerityLandingAI shipped Agentic Document Extraction Gen2, featuring the DPT-3 Verity model for word-level grounding. | LandingAI | — | text | Marktechpost |
| 2 weeks ago | DeepSeek-V4.1-FlashDeepSeek launched DeepSeek-V4.1-Flash with a 552-billion-parameter backbone, native vision, and a 1-million-token context window. | DeepSeek | API | multimodal | TechNode+5 |
| 4 weeks ago | GigaPath-FlashMicrosoft Research presents GigaPath-Flash and GigaTIME-Flash pathology foundation models. | Microsoft | research | image | Microsoft Research |
| today | Agent UltraExa released Agent Ultra, a high‑effort mode of its Exa Agent API for exhaustive list building. | Exa | API | text | Marktechpost |
| yesterday | GWM Worlds 2Runway announces GWM Worlds 2, a world model generating video and audio in real time. | Runway | — | multimodal | Latent Space |
| yesterday | Altar-1Aikido Security released Altar-1, an open-weight security model pruned from GLM-5.3. | Aikido Security | open weights | — | Marktechpost |
| 2 days ago | ThinkingCap-Qwen3.8-27BBottleCap AI released ThinkingCap-Qwen3.8-27B, a fine-tune of Qwen3.8-27B with fewer thinking tokens. | BottleCap AI | open weights | text | Marktechpost |
| 2 days ago | Spark-ASR-2.0iFLYTEK shipped Spark-ASR-2.0 for noisy, dialect and Chinese-English code-switch recognition with an open API. | iFLYTEK | API | speech | Pandaily |
| 2 days ago | CLM-8BContrastive-LM released CLM-8B, an open System One model that scores agent actions faster than Jev. | Contrastive-LM | open weights | text | Marktechpost |
| 3 days ago | together/Tev1-4B-experimentalTogether AI launches Jev-like classifier together/Tev1-4B-experimental on its serverless platform. | Together AI | API | text | Together AI |
| 3 days ago | Xing4.0-29B-A4BChina Telecom AI released Xing4.0-29B-A4B with open weights. | China Telecom AI | open weights | text | Pandaily |
| 3 days ago | Voice of ReasonKyutai released two open-weight speech-to-speech models that solve spoken math with reinforcement learning. | Kyutai | open weights | speech | Marktechpost |
| 5 days ago | Futurist humanoidFaraday Future unveiled the all-new Futurist humanoid robot for sale. | Faraday Future | — | robotics | DigiTimes |
| 5 days ago | AGILE 2.0AgiBot released AGILE 2.0, an end-to-end sense-control locomotion model for open-world environments. | AgiBot | — | robotics | Pandaily |
| 6 days ago | ZGCM-1-7BZGCM-1-7B releases full weights, data, and code for math and agentic search. | Zhongguancun Academy | open weights | text | Pandaily |
| 6 days ago | Step 5 PreviewStepFun announces Step 5 Preview, a 600B sparse MoE model with 1M context, weights opening Oct 15. | StepFun | open weights | text | Pandaily+1 |
| 7 days ago | SPARSEUPLinkup Research released SPARSEUP, an open-source sparse embedding model built on a 149M-parameter ModernBERT backbone. | Linkup Research | open weights | embedding | Marktechpost |
| 8 days ago | jina-ocr-v1Jina AI has released jina-ocr-v1, a visual document parser that converts PDFs, scans, tables, charts and invoices into Markdown. | Jina AI | — | image | Marktechpost |
| 8 days ago | ZDTaichu5.0-9BTaichuAI open-sourced ZDTaichu5.0-9B, a ~9B spatial multimodal model. | TaichuAI | open weights | multimodal | Pandaily |
| 8 days ago | Ternary Bonsai 2 27BPrismML released Ternary Bonsai 2 27B, a 5.9 GB Apache‑2.0 model supporting text and images. | PrismML | open weights | multimodal | Marktechpost |
| 8 days ago | NemotronLabs VoiceChatNemotronLabs VoiceChat is an open full-duplex speech-to-speech model with native tool-calling capabilities. | NemotronLabs | open weights | speech | arXiv cs.AI |
| 8 days ago | NASA-IBM Lunar FoundationNASA and IBM released an open-source geospatial AI model for lunar applications. | NASA-IBM | open weights | — | Hacker News AI |
| 8 days ago | Doubao-Seed-2.1-pro 0915ByteDance shipped Doubao-Seed-2.1-pro 0915, a multimodal coding model on Volcengine. | ByteDance | — | multimodal | Pandaily |
| 8 days ago | Cactus Needle 3Cactus Needle 3 model was announced, claiming to match DeepSeek V4 Flash performance. | Cactus | — | — | Hacker News AI |
| 8 days ago | ABot-Earth 0.7Amap launched ABot-Earth 0.7, a 3D-native urban world model generating cities from satellite or text input. | Amap | — | world | Pandaily |
| 9 days ago | PhysBrain 1.5DeepCybo open-sources PhysBrain 1.5, a 2B/8B physical foundation model. | DeepCybo | open weights | multimodal | Pandaily |
| 9 days ago | LimiX-2Stable AI and Tsinghua released LimiX-2, a 400M structured-data foundation model using Contextual Mechanism Networks. | Stable AI | — | — | Pandaily |
| 9 days ago | HERON-CRASongyan Dynamics released the Scalabot HERON-CRA robot with 4D Context Expert memory and RL Engine. | Songyan Dynamics | — | robotics | Pandaily |
| 9 days ago | APXInfInfinigence AI open-sources APXInf for embodied edge inference on Jetson Thor. | Infinigence AI | open weights | — | Pandaily |
| 10 days ago | TabPFN-3.5Prior Labs releases TabPFN-3.5, a tabular foundation model that beats Otto's winning Kaggle solution with default settings. | Prior Labs | — | — | Marktechpost |
| 10 days ago | Specs IntelligenceSnap launches Specs Intelligence, a new AI assistant for iOS and Mac. | Snap | in-app | text | The Verge AI |
| 10 days ago | GLiFormerKnowledgator releases GLiFormer, a 575M-parameter encoder achieving 91.10 F1 on nested JSON extraction. | Knowledgator | — | embedding | Marktechpost |
| 10 days ago | Cog-WM 1.0EBKernel launched Cog-WM 1.0, a brain-inspired cognitive world model. | EBKernel | — | world | Pandaily |
| 10 days ago | CausiloNums AI released Causilo, a pretrained tabular foundation model for classification and regression. | Nums AI | open weights | — | Marktechpost |
| 12 days ago | OM-1Reward AI has released OM-1, a robot policy trained solely on human demonstrations. | Reward AI | — | robotics | Marktechpost |
| 12 days ago | Motus2Shengshu Technology releases Motus2, a self‑evolving world model for dexterous manipulation. | Shengshu | research | world | Pandaily |
| 12 days ago | Kimi K2.8 PreviewMoonshot AI launches Kimi K2.8 Preview for Kimi Code and Work users via API. | Moonshot AI | API | code | Pandaily |
| 2 weeks ago | SWE-2Cognition released SWE-2, its most capable coding model to date, post-trained from Kimi K3. | Cognition | open weights | code | Marktechpost |
| 2 weeks ago | MiniCPM5-2BOpenBMB open-sourced MiniCPM5-2B, a 2.52B dense Apache-2.0 model. | OpenBMB | open weights | text | Pandaily |
| 3 weeks ago | K2 HorizonIFM released K2 Horizon, a fleet of six Apache 2.0 models ranging from 0.9B to 375B. | Institute of Foundation Models | open weights | text | Hacker News AI+1 |
| 4 weeks ago | XenomiBeijing Guoguang Liangchao Technology introduced Xenomi, a family of large language models. | Beijing Guoguang Liangchao Technology | — | text | DigiTimes |
| 4 weeks ago | Gradium AI Default TTSGradium AI releases a new default text-to-speech model with high pass rates. | Gradium AI | API | speech | Marktechpost |
| 4 weeks ago | Lux3DManycore Tech launched Lux3D, which generates 3D assets from text or images. | Manycore Tech | API | image | TechNode |
| 4 weeks ago | Ling-3.0-flash-FinAnt Group launched the finance-tuned Ling-3.0-flash-Fin model, with weights scheduled for release next week. | Ant Group | — | text | TechNode |
| 4 weeks ago | Hy4Tencent launched a preview version of its new flagship open-source model, Hy4. | Tencent | open weights | text | The Information+1 |
| 4 weeks ago | Granite 4.2IBM's new Granite 4.2 models ride the wave of interest in local LLMs. | IBM | — | text | Ars Technica AI |
| 5 weeks ago | GEN-1.5Generalist AI released GEN-1.5, a robot foundation model that learns new physical tasks from a single demonstration. | Generalist AI | — | robotics | Marktechpost |
| 5 weeks ago | dots3-note previewXiaohongshu released the dots3-note preview, a MoE model with 280B total and 16B activated parameters. | Xiaohongshu | open weights | multimodal | Pandaily |
| 5 weeks ago | Omega-0Researchers released Omega-0, a latent-prediction world action model that lets a humanoid robot walk, look, and work. | BAAI | research | robotics | Pandaily |
| 5 weeks ago | S1-miniS1-mini is a 462 MB open-weights normalizer that sits after ASR, removing fillers and resolving self-corrections locally. | Superwhisper | open weights | text | Marktechpost |
| 7 weeks ago | Pokee-Isaac 28BPokee AI released Pokee-Isaac 28B, a 28B text-only foundation model with a 10M-token context window. | Pokee AI | — | text | Marktechpost |
| 8 weeks ago | MiniMax-H3MiniMax released MiniMax-H3, a general-purpose, omni-modal generative system that accepts text, images, audio and video. | MiniMax | open weights | multimodal | The Decoder+1 |
| 8 weeks ago | Instella-MoE-16B-A3BAMD released Instella-MoE-16B-A3B, a fully open Mixture-of-Experts language model trained from scratch on Instinct GPUs. | AMD | open weights | text | Marktechpost |
| 8 weeks ago | Inkling-SmallThinking Machines Lab Releases Inkling-Small: A 276B Total, 12B Active Open Weights Multimodal MoE Model. | Thinking Machines Lab | open weights | multimodal | The Decoder+1 |
Every release
Grouped by the month it was first reported. The link on each name opens the story this briefing read it in.
September 2026 · 111
| Released | Model | Lab | Access | Type | Reported by |
|---|---|---|---|---|---|
| today | Agent UltraExa released Agent Ultra, a high‑effort mode of its Exa Agent API for exhaustive list building. | Exa | API | text | Marktechpost |
| yesterday | GWM Worlds 2Runway announces GWM Worlds 2, a world model generating video and audio in real time. | Runway | — | multimodal | Latent Space |
| yesterday | GLiNER2.5-DecideFastino Labs released GLiNER2.5-Decide, a 340M-parameter open-weight decision model that runs on CPU. | Fastino | open weights | text | Marktechpost |
| yesterday | Altar-1Aikido Security released Altar-1, an open-weight security model pruned from GLM-5.3. | Aikido Security | open weights | — | Marktechpost |
| 2 days ago | Zhenwu V900Alibaba unveiled its self-developed Zhenwu V900 AI chip. | Alibaba | — | — | DigiTimes |
| 2 days ago | ThinkingCap-Qwen3.8-27BBottleCap AI released ThinkingCap-Qwen3.8-27B, a fine-tune of Qwen3.8-27B with fewer thinking tokens. | BottleCap AI | open weights | text | Marktechpost |
| 2 days ago | Spark-ASR-2.0iFLYTEK shipped Spark-ASR-2.0 for noisy, dialect and Chinese-English code-switch recognition with an open API. | iFLYTEK | API | speech | Pandaily |
| 2 days ago | PUBG AllyIntroduces PUBG Ally, a conversational embodied agent for PUBG that reasons, acts autonomously, and plays as a voice-enabled teammate. | Unknown | research | multimodal | arXiv cs.AI |
| 2 days ago | Jev-MobileJev-Mobile is introduced as an executor for mobile GUI agents. | TypeSafe | research | multimodal | arXiv cs.AI |
| 2 days ago | HEXISHEXIS compiles agent skills into extended finite state machines. | Unknown | research | robotics | arXiv cs.AI |
| 2 days ago | GridSFMGridSFM is a framework for solving AC Optimal Power Flow. | Unknown | research | science | arXiv cs.LG |
| 2 days ago | GRASPGRASP is a strategy-aware, multi-stage planning framework for complex tasks. | Unknown | research | text | arXiv cs.AI |
| 2 days ago | CLM-8BContrastive-LM released CLM-8B, an open System One model that scores agent actions faster than Jev. | Contrastive-LM | open weights | text | Marktechpost |
| 3 days ago | together/Tev1-4B-experimentalTogether AI launches Jev-like classifier together/Tev1-4B-experimental on its serverless platform. | Together AI | API | text | Together AI |
| 3 days ago | gemini-3.8-flash-lite-ttsGoogle released gemini-3.8-flash-lite-tts model. | API | audio | Simon Willison | |
| 3 days ago | Xing4.0-29B-A4BChina Telecom AI released Xing4.0-29B-A4B with open weights. | China Telecom AI | open weights | text | Pandaily |
| 3 days ago | Voice of ReasonKyutai released two open-weight speech-to-speech models that solve spoken math with reinforcement learning. | Kyutai | open weights | speech | Marktechpost |
| 3 days ago | Qwen-Audio-3.1Alibaba debuted Qwen3.8-LiveTranslate and unveiled Qwen-Audio-3.1 family at Apsara 2026. | Alibaba | API | audio | Pandaily+1 |
| 3 days ago | NVIDIA Nemotron 3 DiarizationNVIDIA Nemotron 3 Diarization enables real-time multi-speaker AI. | Nvidia | open weights | audio | Hugging Face+1 |
| 3 days ago | NV-Reason-CT Open 3D CT VLMNVIDIA introduces NV-Reason-CT Open 3D CT VLM for radiologist chain-of-thought reasoning. | Nvidia | — | multimodal | NVIDIA Developer |
| 3 days ago | ME-VLMLi Auto released ME-VLM model. | Li Auto | — | text | Pandaily |
| 3 days ago | ME-U0Li Auto released ME-U0 model. | Li Auto | — | text | Pandaily |
| 3 days ago | ME-Brain-1.0Li Auto released ME-Brain-1.0 model. | Li Auto | — | text | Pandaily |
| 3 days ago | Large KnowledgeLarge Knowledge Model is introduced as a new model for scientific reasoning from papers. | Unknown | — | science | arXiv cs.CL |
| 3 days ago | InternW0InternW0 is the first instantiation of the InternW physical world model series from Shanghai AI Laboratory. | Shanghai AI Lab | — | world | arXiv cs.AI |
| 3 days ago | Gemini 3.8 Flash TTSGoogle launches Gemini 3.8 Flash TTS text-to-speech model. | API | audio | The Decoder+2 | |
| 3 days ago | Flash-Lite TTSGoogle launches Flash-Lite TTS text-to-speech model. | API | audio | The Decoder+1 | |
| 3 days ago | FLUX 3 ActionBlack Forest Labs releases open-weight FLUX 3 Action robotics model. | Black Forest Labs | open weights | robotics | VentureBeat AI+1 |
| 4 days ago | SenseNova U1 ProSenseTime released the production version of its SenseNova U1 Pro image-creation model via its Raccoon app and API. | SenseTime | in-app | image | TechNode+1 |
| 4 days ago | MiMo-V2.6-ProXiaomi released MiMo-V2.6-Pro as the top-performing open-weight model in the world. | Xiaomi | open weights | text | VentureBeat AI+2 |
| 4 days ago | MiMo-V2.6-FlashXiaomi released MiMo-V2.6-Flash alongside the Pro version as a cheaper option. | Xiaomi | open weights | text | VentureBeat AI+2 |
| 4 days ago | MiMo-V2.6Xiaomi released and open-sourced its MiMo-V2.6 series. | Xiaomi | open weights | text | DigiTimes |
| 4 days ago | Light-O1-PreviewLight Origins released Light-O1-Preview, a 6B Apache-2.0 model for robot whole-body trajectories. | Light Origins | open weights | robotics | Pandaily |
| 4 days ago | Intern-S2-397BShanghai AI Laboratory released Intern-S2-397B, an open-weight science multimodal model with a pluggable Memory Decoder. | Shanghai AI Lab | open weights | multimodal | Pandaily |
| 4 days ago | GPT-6 SolOpenAI releases GPT-6 Sol, cutting API costs by 50% or more. | OpenAI | API | text | VentureBeat AI+5 |
| 4 days ago | GPT-6 LunaOpenAI releases GPT-6 Luna, cutting API costs by 50% or more. | OpenAI | API | text | VentureBeat AI+5 |
| 4 days ago | Claude Opus 5.5Anthropic is launching Claude Opus 5.5, the first model in a new generation. | Anthropic | API | text | The Decoder+6 |
| 5 days ago | t0-betaThe authors released t0-beta (256M) open‑weight time‑series model for forecasting. | Unknown | open weights | science | arXiv cs.LG |
| 5 days ago | t0-alphaThe authors released t0-alpha (102M) open‑weight time‑series model for forecasting. | Unknown | open weights | science | arXiv cs.LG |
| 5 days ago | T1PrimeBot launched transformable T1 personal robot for sale. | PrimeBot | — | robotics | Pandaily |
| 5 days ago | Q1PrimeBot launched foldable Q1 personal robot for sale. | PrimeBot | — | robotics | Pandaily |
| 5 days ago | NVIDIA Earth-2NVIDIA Earth-2 helps weather-sensitive industries make timely decisions using local observations. | Nvidia | — | image | NVIDIA Developer |
| 5 days ago | Meta SAM 3.1Meta announced the SAM 3.1 model. | Meta | — | image | TLDR AI |
| 5 days ago | Grok 4.7Grok 4.7 is rolling out in GitHub Copilot for agentic coding and complex workflows. | xAI | in-app | code | GitHub Copilot+3 |
| 5 days ago | Futurist humanoidFaraday Future unveiled the all-new Futurist humanoid robot for sale. | Faraday Future | — | robotics | DigiTimes |
| 5 days ago | AGILE 2.0AgiBot released AGILE 2.0, an end-to-end sense-control locomotion model for open-world environments. | AgiBot | — | robotics | Pandaily |
| 6 days ago | ZGCM-1-7BZGCM-1-7B releases full weights, data, and code for math and agentic search. | Zhongguancun Academy | open weights | text | Pandaily |
| 6 days ago | Step 5 PreviewStepFun announces Step 5 Preview, a 600B sparse MoE model with 1M context, weights opening Oct 15. | StepFun | open weights | text | Pandaily+1 |
| 6 days ago | Qwen3.8-LiveTranslateAlibaba Qwen team released Qwen3.8-LiveTranslate, a real-time interpretation model. | Alibaba | — | speech | Marktechpost+1 |
| 6 days ago | Qwen-Image-2.1Qwen-Image-2.1 is introduced as a compact, efficient, unified image creation model. | Alibaba | open weights | image | Hacker News AI+4 |
| 6 days ago | QwenTry Qwen at qwen.ai | Alibaba | — | text | Qwen |
| 7 days ago | SenseNova U1.5SenseTime releases SenseNova U1.5, an 8B Mixture-of-Transformers vision model with open training code. | SenseTime | — | image | Pandaily |
| 7 days ago | SPARSEUPLinkup Research released SPARSEUP, an open-source sparse embedding model built on a 149M-parameter ModernBERT backbone. | Linkup Research | open weights | embedding | Marktechpost |
| 7 days ago | Grok Voice Transcribe 2.0SpaceXAI released Grok Voice Transcribe 2.0, a speech-to-text API claiming 2x accuracy over 1.0. | xAI | API | speech | Marktechpost |
| 7 days ago | DAMO RADARAlibaba DAMO Academy open-sources the DAMO RADAR abdominal CT model. | Alibaba | open weights | image | Pandaily |
| 8 days ago | jina-ocr-v1Jina AI has released jina-ocr-v1, a visual document parser that converts PDFs, scans, tables, charts and invoices into Markdown. | Jina AI | — | image | Marktechpost |
| 8 days ago | ZDTaichu5.0-9BTaichuAI open-sourced ZDTaichu5.0-9B, a ~9B spatial multimodal model. | TaichuAI | open weights | multimodal | Pandaily |
| 8 days ago | Ternary Bonsai 2 27BPrismML released Ternary Bonsai 2 27B, a 5.9 GB Apache‑2.0 model supporting text and images. | PrismML | open weights | multimodal | Marktechpost |
| 8 days ago | SamsoneA family of open small audio language models for on-device inference. | Unknown | — | audio | arXiv cs.AI |
| 8 days ago | NemotronLabs VoiceChatNemotronLabs VoiceChat is an open full-duplex speech-to-speech model with native tool-calling capabilities. | NemotronLabs | open weights | speech | arXiv cs.AI |
| 8 days ago | NASA-IBM Lunar FoundationNASA and IBM released an open-source geospatial AI model for lunar applications. | NASA-IBM | open weights | — | Hacker News AI |
| 8 days ago | GLM-5.3-FlashXZhipu AI launched GLM-5.3-FlashX with claimed ~200 tok/s inference on ~100,000 domestic accelerators. | Z.ai | open weights | text | Pandaily |
| 8 days ago | Doubao-Seed-2.1-pro 0915ByteDance shipped Doubao-Seed-2.1-pro 0915, a multimodal coding model on Volcengine. | ByteDance | — | multimodal | Pandaily |
| 8 days ago | Damo Academy Medical AIAlibaba's Damo Academy open-sourced a model detecting 150 abdominal conditions from CT scans. | Alibaba | open weights | image | SCMP Tech |
| 8 days ago | Cactus Needle 3Cactus Needle 3 model was announced, claiming to match DeepSeek V4 Flash performance. | Cactus | — | — | Hacker News AI |
| 8 days ago | Astra for LawOpenAI introduced Astra for Law, a version of its GPT-6 Astra model built for legal work. | OpenAI | — | text | The Decoder |
| 8 days ago | ABot-Earth 0.7Amap launched ABot-Earth 0.7, a 3D-native urban world model generating cities from satellite or text input. | Amap | — | world | Pandaily |
| 9 days ago | Qwen 3.8 Omni FlashAlibaba releases Qwen 3.8 Omni Flash model. | Alibaba | API | multimodal | Hacker News AI+3 |
| 9 days ago | PhysBrain 1.5DeepCybo open-sources PhysBrain 1.5, a 2B/8B physical foundation model. | DeepCybo | open weights | multimodal | Pandaily |
| 9 days ago | LimiX-2Stable AI and Tsinghua released LimiX-2, a 400M structured-data foundation model using Contextual Mechanism Networks. | Stable AI | — | — | Pandaily |
| 9 days ago | LightNav-0Light Origins open-sourced LightNav-0, a Qwen3-VL-4B navigation model trained via Real2Sim2Real on 2,000+ scenes. | Light Origins | open weights | robotics | Pandaily |
| 9 days ago | HERON-CRASongyan Dynamics released the Scalabot HERON-CRA robot with 4D Context Expert memory and RL Engine. | Songyan Dynamics | — | robotics | Pandaily |
| 9 days ago | APXInfInfinigence AI open-sources APXInf for embodied edge inference on Jetson Thor. | Infinigence AI | open weights | — | Pandaily |
| 10 days ago | Xiaomi-Robotics-U0Xiaomi open-sourced Xiaomi-Robotics-U0, an embodied world foundation model with 4B and 38B weights. | Xiaomi | open weights | world | Pandaily |
| 10 days ago | TabPFN-3.5Prior Labs releases TabPFN-3.5, a tabular foundation model that beats Otto's winning Kaggle solution with default settings. | Prior Labs | — | — | Marktechpost |
| 10 days ago | Specs IntelligenceSnap launches Specs Intelligence, a new AI assistant for iOS and Mac. | Snap | in-app | text | The Verge AI |
| 10 days ago | JevTypeSafe released Jev, a System One model that decides, classifies, routes, and scores. | TypeSafe | API | text | Latent Space+5 |
| 10 days ago | GLiFormerKnowledgator releases GLiFormer, a 575M-parameter encoder achieving 91.10 F1 on nested JSON extraction. | Knowledgator | — | embedding | Marktechpost |
| 10 days ago | Cog-WM 1.0EBKernel launched Cog-WM 1.0, a brain-inspired cognitive world model. | EBKernel | — | world | Pandaily |
| 10 days ago | CausiloNums AI released Causilo, a pretrained tabular foundation model for classification and regression. | Nums AI | open weights | — | Marktechpost |
| 10 days ago | Atria Dawn PreviewShanghai Artificial Intelligence Laboratory open-sourced Atria Dawn Preview, a 744B MoE agentic model. | Shanghai AI Lab | open weights | text | Pandaily |
| 11 days ago | Gemini 3.8 Live Extended ThinkingGoogle DeepMind introduces Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. | API | multimodal | Google DeepMind+2 | |
| 11 days ago | Gemini 3.8 LiveGoogle DeepMind introduces Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. | API | multimodal | Google DeepMind+2 | |
| 12 days ago | OM-1Reward AI has released OM-1, a robot policy trained solely on human demonstrations. | Reward AI | — | robotics | Marktechpost |
| 12 days ago | NCP-ArchPreviewShanghai AI Lab and SJTU open-sourced NCP-ArchPreview, an 8.9B latent-space model. | Shanghai AI Lab | open weights | text | Pandaily |
| 12 days ago | Motus2Shengshu Technology releases Motus2, a self‑evolving world model for dexterous manipulation. | Shengshu | research | world | Pandaily |
| 12 days ago | Kimi K2.8 PreviewMoonshot AI launches Kimi K2.8 Preview for Kimi Code and Work users via API. | Moonshot AI | API | code | Pandaily |
| 12 days ago | Claude for Financial AdvisorsAnthropic introduces Claude for Financial Advisors. | Anthropic | — | text | Anthropic |
| 13 days ago | Iris-proThe AllSpark team released Iris-mini and Iris-pro, two open-source search agents built on Qwen models. | AllSpark team | open weights | text | The Decoder |
| 13 days ago | Iris-miniThe AllSpark team released Iris-mini and Iris-pro, two open-source search agents built on Qwen models. | AllSpark team | open weights | text | The Decoder |
| 2 weeks ago | SWE-2Cognition released SWE-2, its most capable coding model to date, post-trained from Kimi K3. | Cognition | open weights | code | Marktechpost |
| 2 weeks ago | North Small TranslateCohere released North Small Translate, an open-weight Mixture-of-Experts model built for machine translation across 50 languages. | Cohere | open weights | text | Marktechpost |
| 2 weeks ago | MiniCPM5-2BOpenBMB open-sourced MiniCPM5-2B, a 2.52B dense Apache-2.0 model. | OpenBMB | open weights | text | Pandaily |
| 2 weeks ago | Fugu Ultra v2Sakana AI released Fugu Max and Fugu Ultra v2, two models for multi-agent orchestration. | Sakana AI | API | text | Marktechpost |
| 2 weeks ago | Fugu MaxSakana AI released Fugu Max and Fugu Ultra v2, two models for multi-agent orchestration. | Sakana AI | API | text | Marktechpost |
| 2 weeks ago | GPT‑Live‑1GPT‑Live‑1 brings natural, full-duplex voice conversations to the API with stronger instruction following. | OpenAI | API | speech | OpenAI |
| 2 weeks ago | DPT-3 VerityLandingAI shipped Agentic Document Extraction Gen2, featuring the DPT-3 Verity model for word-level grounding. | LandingAI | — | text | Marktechpost |
| 2 weeks ago | DPT-3 ProLandingAI shipped Agentic Document Extraction Gen2, featuring the DPT-3 Pro model for line-level grounding. | LandingAI | — | text | Marktechpost |
| 2 weeks ago | DeepSeek-V4.1-FlashDeepSeek launched DeepSeek-V4.1-Flash with a 552-billion-parameter backbone, native vision, and a 1-million-token context window. | DeepSeek | API | multimodal | TechNode+5 |
| 3 weeks ago | Claude Sonnet 5Anthropic announces Claude Sonnet 5. | Anthropic | — | text | Anthropic |
| 3 weeks ago | ChatGPT Images 2.5ChatGPT Images 2.5 helps turn ideas and sketches into personalized images. | OpenAI | in-app | image | OpenAI |
| 3 weeks ago | Qwen-Drive-1.0-4BAlibaba’s Qwen team has released Qwen-Drive-1.0-4B, an open-source model for autonomous driving. | Alibaba | open weights | multimodal | TechNode+1 |
| 3 weeks ago | Lyria 3.5Google has released its Lyria 3.5 music model in the Gemini app and via API. | API | audio | Google (Gemini)+1 | |
| 3 weeks ago | WeatherNext 3Google DeepMind introduces WeatherNext 3, its most advanced global weather AI model. | — | science | Google DeepMind | |
| 3 weeks ago | Muse Spark 1.3Meta AI Released Muse Spark 1.3, an agentic coding model with fewer tool calls. | Meta | — | code | Marktechpost |
| 3 weeks ago | K2 HorizonIFM released K2 Horizon, a fleet of six Apache 2.0 models ranging from 0.9B to 375B. | Institute of Foundation Models | open weights | text | Hacker News AI+1 |
| 3 weeks ago | GPT-6 AstraOpenAI introduced GPT-6 Astra, its most capable model for business with advanced reasoning and computer use. | OpenAI | API | multimodal | Hacker News AI+5 |
| 3 weeks ago | Gemini 3.8 FlashGoogle announced two versions of a new 3.8 Flash model. | API | multimodal | Google DeepMind+2 | |
| 4 weeks ago | XenomiBeijing Guoguang Liangchao Technology introduced Xenomi, a family of large language models. | Beijing Guoguang Liangchao Technology | — | text | DigiTimes |
| 4 weeks ago | Gradium AI Default TTSGradium AI releases a new default text-to-speech model with high pass rates. | Gradium AI | API | speech | Marktechpost |
| 4 weeks ago | Claude Fable 5.1Anthropic launches Claude Fable 5.1 and says it is up to 45 percent cheaper. | Anthropic | API | text | The Verge AI+2 |
August 2026 · 33
| Released | Model | Lab | Access | Type | Reported by |
|---|---|---|---|---|---|
| 4 weeks ago | TimesFM-3Google Research released TimesFM-3, a 330-million-parameter forecasting model that analyzes time series data. | research | science | Google Research+1 | |
| 4 weeks ago | GigaPath-FlashMicrosoft Research presents GigaPath-Flash and GigaTIME-Flash pathology foundation models. | Microsoft | research | image | Microsoft Research |
| 4 weeks ago | Lux3DManycore Tech launched Lux3D, which generates 3D assets from text or images. | Manycore Tech | API | image | TechNode |
| 4 weeks ago | Ling-3.0-flash-FinAnt Group launched the finance-tuned Ling-3.0-flash-Fin model, with weights scheduled for release next week. | Ant Group | — | text | TechNode |
| 4 weeks ago | Hy4Tencent launched a preview version of its new flagship open-source model, Hy4. | Tencent | open weights | text | The Information+1 |
| 4 weeks ago | Parse 5Cohere released Parse 5, a 2.3B-parameter vision language model that converts documents into Markdown. | Cohere | API | multimodal | Marktechpost |
| 4 weeks ago | Gemini Omni 1.1 FlashGoogle DeepMind released Gemini Omni 1.1 Flash, allowing users to build with more control. | — | multimodal | Google DeepMind | |
| 4 weeks ago | GLM-5.3 FlashZ.ai’s new open-source model, GLM-5.3 Flash, was officially released earlier this week. | Z.ai | open weights | text | SCMP Tech+1 |
| 4 weeks ago | Qwen3.8-Flash-NextQwen released Qwen3.8-Flash-Next, an open weights multimodal MoE model. | Alibaba | open weights | multimodal | Simon Willison |
| 4 weeks ago | Grok TtsxAI released Grok Tts. | xAI | — | speech | xAI |
| 4 weeks ago | Granite 4.2IBM's new Granite 4.2 models ride the wave of interest in local LLMs. | IBM | — | text | Ars Technica AI |
| 4 weeks ago | GlucoFMGoogle Research released GlucoFM, a foundation model for continuous glucose monitoring. | research | science | Google Research | |
| 4 weeks ago | Gemini 3.5 TranscribeGoogle announced Gemini 3.5 Transcribe for AI-powered speech-to-text. | in-app | speech | Ars Technica AI+1 | |
| 5 weeks ago | Wan3.0Alibaba's video generation model Wan3.0 creates clips up to 30 seconds long from text, PDFs, and PowerPoint files. | Alibaba | API | video | The Decoder |
| 5 weeks ago | GLiNER2.5Fastino released GLiNER2.5, a boundary-prediction architecture for information extraction with three Apache 2.0 checkpoints. | Fastino | open weights | text | Marktechpost |
| 5 weeks ago | GEN-1.5Generalist AI released GEN-1.5, a robot foundation model that learns new physical tasks from a single demonstration. | Generalist AI | — | robotics | Marktechpost |
| 5 weeks ago | dots3-note previewXiaohongshu released the dots3-note preview, a MoE model with 280B total and 16B activated parameters. | Xiaohongshu | open weights | multimodal | Pandaily |
| 5 weeks ago | Wan-Animate-2Alibaba's Tongyi Wanxiang team open-sourced Wan-Animate-2, an end-to-end character animation framework running at 24 fps. | Alibaba | open weights | video | Pandaily |
| 5 weeks ago | V4-Flash-Vision-ExpDeepseek released V4-Flash-Vision-Exp, an experimental multimodal model that adds image understanding to V4-Flash's text capabilities. | DeepSeek | API | multimodal | The Decoder+1 |
| 5 weeks ago | Qwen-UI-AgentAlibaba Qwen released Qwen-UI-Agent, a GUI agent foundation model for real devices covering phones, desktops, and web. | Alibaba | — | multimodal | Pandaily |
| 5 weeks ago | Omega-0Researchers released Omega-0, a latent-prediction world action model that lets a humanoid robot walk, look, and work. | BAAI | research | robotics | Pandaily |
| 5 weeks ago | S1-miniS1-mini is a 462 MB open-weights normalizer that sits after ASR, removing fillers and resolving self-corrections locally. | Superwhisper | open weights | text | Marktechpost |
| 6 weeks ago | Qwen3.8-27BQwen3.8-27B was released August 14, passing a million downloads in two days and hitting Hugging Face's global trend list. | Alibaba | open weights | text | Simon Willison+2 |
| 6 weeks ago | GLM-5.3GLM-5.3 is now open-weight. | Z.ai | open weights | text | VentureBeat AI+1 |
| 7 weeks ago | CARE-XMicrosoft Research introduces CARE-X, a unified approach for clinically useful radiology VLMs. | Microsoft | research | image | Microsoft Research |
| 7 weeks ago | NemotronLabs VoiceChat 11BNVIDIA releases NemotronLabs VoiceChat 11B, an open full-duplex speech-to-speech model with live tool calling. | Nvidia | open weights | speech | Marktechpost |
| 7 weeks ago | Pokee-Isaac 28BPokee AI released Pokee-Isaac 28B, a 28B text-only foundation model with a 10M-token context window. | Pokee AI | — | text | Marktechpost |
| 7 weeks ago | Imagine Image 2.0xAI has released Imagine Image 2.0 as a new image generator for Grok. | xAI | in-app | image | The Decoder |
| 7 weeks ago | Muse Spark 1.2Meta shipped Muse Spark 1.2, a coding-focused update to Muse Spark 1.1. | Meta | in-app | code | Simon Willison+1 |
| 7 weeks ago | FLUX 3 VideoBlack Forest Labs has launched FLUX 3 Video, which generates Full HD clips up to 20 seconds long. | Black Forest Labs | API | video | The Decoder |
| 8 weeks ago | Qwen3.8-MaxAlibaba released Qwen3.8-Max, claiming performance rivaling the best systems from US frontier labs. | Alibaba | — | text | The Verge AI+1 |
| 8 weeks ago | MiniMax-H3MiniMax released MiniMax-H3, a general-purpose, omni-modal generative system that accepts text, images, audio and video. | MiniMax | open weights | multimodal | The Decoder+1 |
| 8 weeks ago | Instella-MoE-16B-A3BAMD released Instella-MoE-16B-A3B, a fully open Mixture-of-Experts language model trained from scratch on Instinct GPUs. | AMD | open weights | text | Marktechpost |
July 2026 · 3
| Released | Model | Lab | Access | Type | Reported by |
|---|---|---|---|---|---|
| 8 weeks ago | Inkling-SmallThinking Machines Lab Releases Inkling-Small: A 276B Total, 12B Active Open Weights Multimodal MoE Model. | Thinking Machines Lab | open weights | multimodal | The Decoder+1 |
| 2 months ago | Claude Opus 5Anthropic is introducing Claude Opus 5. | Anthropic | — | text | Anthropic+1 |
| 2 months ago | Gemini 3.6 FlashGoogle DeepMind is introducing new Gemini models, including Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber. | — | text | Google DeepMind |
Seen, but not a launch
Models this briefing met through a distribution story — a cloud, an IDE or an inference platform adding a model that already existed — rather than through its announcement. The date is when that story ran, which is not when the model shipped, so they are kept out of the release list above and out of its counts. They are here because the sighting is real: this is the row for a model whose launch no source on the list covered.
| Released | Model | Lab | Access | Type | Reported by |
|---|---|---|---|---|---|
| 2 days ago | local languageA frozen local language model with four billion parameters is used for candidate generation. | Unknown | research | text | arXiv cs.AI |
| 5 days ago | Yuanzai robot dogDongfeng Motor announced the Yuanzai robot dog will soon be available at dealerships. | Dongfeng Motor | — | robotics | Pandaily |
| 5 days ago | Grok 4.6xAI's Grok 4.6 is now available in Amazon Bedrock for long-running agents and coding. | xAI | API | text | AWS Machine Learning |
| 8 days ago | Kimi K3Kimi K3 from Moonshot AI is now available on Amazon Bedrock as an open-weight model for coding and knowledge work. | Moonshot AI | open weights | multimodal | AWS Machine Learning+1 |
| 2 weeks ago | GLM 5.2GLM 5.2 is now available on CoreWeave inference as an API. | Z.ai | API | text | CoreWeave |
| 5 weeks ago | GPT-5.6GPT-5.6 is now available in Kiro, helping developers plan, build, review, and test software. | OpenAI | API | code | OpenAI |
| 6 weeks ago | DeepSeek V4 Pro 0813The latest DeepSeek Pro model is now available via API only on OpenRouter. | DeepSeek | API | text | Simon Willison |
| 3 months ago | Kimi K2.7 CodeCoreWeave Inference offers the newly-launched Kimi K2.7 Code on serverless inference. | Moonshot AI | API | code | CoreWeave |
How this page is built
Each sweep reads the same candidate pool the briefing is selected from, and a model decides which stories announce a named release rather than merely discussing one — a distinction a keyword cannot make, since “Anthropic launches Claude Fable 5.1” and “Claude Fable 5.1 made me a nice pelican” are one sentence to a pattern. What it finds is folded into a cumulative record, keeping the earliest date and counting the outlets. When no model is reachable a stricter pattern stands in, and rows it produced are marked unconfirmed until a later run can check them. Nothing here is a vendor announcement: it is what the sources on the source list reported, and a release no source covered is a release this page does not know about.
Two things it refuses to treat as launches. A story that a model is now available on some cloud, IDE or inference platform is a distribution note about a model that already existed, and it is filed below the releases rather than dated as one. And a version cannot be newer than a higher version of itself: where a lab is already on record at 5.3, a later story about 5.2 is somebody catching up, not a release, and it is moved out of the list. The lab a model is filed under comes from the family its name belongs to before it comes from whatever the story called it — GLM is Z.ai's whoever is serving it.
2 digest-only sightings are held off this list until a second outlet or firmer detail confirms them.
One bridge sighting is held off this list until a second outlet or firmer detail confirms it.