Mistral Large 4(mistral.ai)
307 points by funnym0nk3y | 104 comments
Mistral AI documentation page for Mistral Large 4, an open-weight multimodal MoE model with 52B active and 1.05T total parameters, listing pricing, context and API features.
#open-weights on Hacker News
87 stories tagged open-weights, newest first.
DeepSeek V4.1 Flash 1017 (Max)(artificialanalysis.ai)
159 points by theanonymousone | 58 comments
Artificial Analysis model page benchmarking DeepSeek V4.1 Flash 1017 (Max) on intelligence, speed, and price.
DeepSeek-v4.1-Flash-1017(huggingface.co)
918 points by theykk | 680 comments
Hugging Face model card for DeepSeek-V4.1-Flash-1017, a 763B-parameter MIT-licensed image-text-to-text model from DeepSeek.
FLUX 3 Image Klein(bfl.ai)
201 points by gmays | 57 comments
Black Forest Labs blog post announcing FLUX 3 Image Klein, a 5.6B-parameter open-weights image generation model.
Ultramarine 1 Flash(paperinstruments.com)
359 points by i_rush_carriers | 104 comments
Product page for Ultramarine 1 Flash, an open AI model from Paper Instruments aimed at professional knowledge work in finance, law, consulting, and operations.
Mistral Large 4(twitter.com)
64 points by apt-get | 4 comments
Mistral AI post announcing Mistral Large 4, a 1T-parameter multimodal open-weights model with 49B active parameters, available via API with weights due end of October.
![Screenshot of aleph-alpha.com: Kolibri – Tech Report [pdf]](https://hn.urlbox.com/shots/renders/2026/10/3/afb8e7e8-ce88-431e-bc45-8ebfad65531a.jpg)
Kolibri – Tech Report [pdf](aleph-alpha.com)
104 points by yu3zhou4 | 4 comments
Aleph Alpha technical report introducing Kolibri, an open-weights English–German Mixture-of-Experts language model with 78.1B total and 3.46B active parameters.
Kolibri is an open-weight LLM from Aleph Alpha for German and English(tej.as)
385 points by tejaskumar__ | 155 comments
A blog post by Tejas Kumar explaining how Aleph Alpha's Kolibri, an open-weight sovereign LLM for German and English, works.
MiMo-v2.6-Pro: Intelligence, Performance and Price Analysis(artificialanalysis.ai)
141 points by theanonymousone | 60 comments
Artificial Analysis benchmark page for Xiaomi's open-weights MiMo-V2.6-Pro model, covering intelligence index, output speed, cost and verbosity rankings.
MiMo v2.6(mimo.xiaomi.com)
1064 points by volf_ | 464 comments
Xiaomi's announcement page for the MiMo-V2.6 model series, with links to try it, the API, Hugging Face weights, a tech report, and coding benchmark charts.
Kev: Tiny Jev-like family of decision models built on top of Qwen3.5(github.com)
425 points by tosh | 192 comments
GitHub repository for Kev, a tiny Jev-like family of decision models built on Qwen3.5 that users can train and run themselves.
DeepSeek v4.1 Flash(twitter.com)
919 points by Liwink | 505 comments
DeepSeek's X post announcing DeepSeek-V4.1-Flash, a small model with native visual understanding, shown with benchmark comparisons against rival models.
K2 Horizon: A connected fleet of six open models(ifm.ai)
327 points by karimf | 121 comments
Blog post from MBZUAI's Institute of Foundation Models announcing K2 Horizon, a family of six open frontier-performance AI models.
GLM-5.3-Flash(z.ai)
1085 points by Philpax | 541 comments
Z.ai blog post announcing GLM-5.3-Flash, a natively multimodal 320B-parameter (18B active) model with hybrid sparse/linear attention, claiming strong coding and agentic results.
Z.ai confirms Ox Alpha is a new GLM-series model and will release its weights(bloomberg.com)
388 points by garo-pro | 132 comments
Bloomberg article reporting that China's Z.ai built the stealth model Ox Alpha, a new GLM-series model it plans to release with open weights; page showed a bot captcha.
Qwen 3.8 27B is excellent, but it defaults to overthinking things(simonwillison.net)
732 points by bilsbie | 343 comments
Simon Willison blog post reviewing Alibaba's Apache 2 licensed Qwen 3.8 27B vision LLM run locally, noting strong results but excessive default reasoning.
Qwen 3.8 27B(huggingface.co)
1391 points by erdaltoprak | 779 comments
Hugging Face model card for Qwen3.8-27B-FP8, an FP8-quantized 27B-parameter image-text-to-text model from Qwen released under Apache 2.0.
Qwen3.8-2.4T(huggingface.co)
692 points by Philpax | 161 comments
Hugging Face model card for Qwen3.8-2.4T-A95B, a 2.4-trillion-parameter mixture-of-experts text generation model from Qwen in Transformers format.
Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows(research.meta.ai)
1156 points by riordan | 624 comments
Meta blog post announcing Muse Glimmer, an open 30B-parameter agentic model designed to run locally on devices for always-on agent workflows.
Mistral's Shieldstral: 3B open-weights model for multimodal moderation(mistral.ai)
462 points by riadsila | 124 comments
Mistral AI blog post announcing Shieldstral, a 3B-parameter open-weights model for multimodal content moderation.
MiniMax H3 Day-0 Support in ComfyUI: Open Weights, Native Audio, and 2K Video(blog.comfy.org)
317 points by vblanco | 89 comments
ComfyUI newsletter post announcing day-0 support for MiniMax H3, an open-weights omni-modal video model with stereo audio and 2K output that runs locally.
![Screenshot of github.com: Kimi-K3 Technical Report [pdf]](https://hn.urlbox.com/shots/renders/2026/7/27/25fa6302-639e-4963-8edf-f78a1d553706.jpg)
Kimi-K3 Technical Report [pdf](github.com)
344 points by vinhnx | 148 comments
GitHub-hosted PDF of Moonshot AI's technical report for Kimi K3, an open frontier large language model.
Kimi-K3 on HuggingFace(huggingface.co)
1294 points by nateb2022 | 508 comments
Hugging Face placeholder page for Moonshot AI's upcoming Kimi-K3 model, showing a release countdown to July 27, 2026 and a notify button.
China’s open-weights AI strategy is winning(werd.io)
1182 points by benwerd | 894 comments
Blog post by Ben Werdmuller arguing China's open-weights AI strategy is winning while America's closed, proprietary approach is failing.
Xiaomi-Robotics-1(robotics.xiaomi.com)
495 points by ilreb | 317 comments
Project page for Xiaomi-Robotics-1, a robot foundation model trained on over 100K hours of real-world manipulation trajectories, with links to report, code, and model.
Kimi K3, and what we can still learn from the pelican benchmark(simonwillison.net)
387 points by droidjj | 209 comments
Simon Willison blog post on Moonshot AI's 2.8-trillion-parameter Kimi K3 model, its benchmarks and pricing, and his pelican-on-a-bicycle test results.
Kimi K3: Open Frontier Intelligence(kimi.com)
1981 points by vincent_s | 1153 comments
Moonshot AI's Kimi blog post announcing the Kimi K3 open frontier model; the captured page shows only the Kimi chat interface.
Inkling: Our Open-Weights Model(thinkingmachines.ai)
1146 points by vimarsh6739 | 278 comments
Thinking Machines blog post announcing Inkling, the company's open-weights AI model, dated July 15, 2026.
Krea 2: SOTA open-weights 12B image model(krea.ai)
387 points by mattnewton | 43 comments
Krea's technical report introducing Krea 2, a family of open-weights image foundation models, covering data curation, architecture, training pipeline, and infrastructure.
GLM-5.2 – How to Run Locally(unsloth.ai)
559 points by TechTechTech | 266 comments
Unsloth documentation guide explaining how to run Z.ai's GLM-5.2 open model locally using Unsloth Dynamic quantized GGUFs.
Apertus – Open Foundation Model for Sovereign AI(apertvs.ai)
503 points by T-A | 168 comments
Landing page for Apertus, a fully open foundation model with open weights, data, and science, developed by the Swiss AI Initiative (EPFL, ETH Zurich, CSCS).
Rio de Janeiro's "homegrown" LLM appears to be a merge of an existing model(github.com)
384 points by unrvl22 | 206 comments
GitHub issue on the Nex-N2 repo alleging Rio's Rio-3.5-Open-397B model is an element-wise merge of about 0.6 Nex-N2_pro and 0.4 Qwen3.5-397B, with no training of its own.
GLM 5.2 Is Out(twitter.com)
701 points by aloknnikhil | 413 comments
Announcement that Z.ai released GLM-5.2, a coding-focused model with a 1M-token context window, with an MIT-licensed open release planned for next week.
Kimi K2.7-Code: open-source coding model with better token efficiency(huggingface.co)
389 points by nekofneko | 208 comments
Hugging Face model card for Moonshot AI's Kimi-K2.7-Code, an open-weights 1T-parameter coding model released under a modified MIT license.
Gemma 4 12B: A unified, encoder-free multimodal model(blog.google)
1007 points by rvz | 379 comments
Google blog post announcing Gemma 4 12B, an encoder-free multimodal model designed to run on laptops with mobile-first efficiency and advanced reasoning.
DeepSeek V4 – almost on the frontier(simonwillison.net)
575 points by indigodaddy | 348 comments
Simon Willison's blog post reviewing DeepSeek's V4 preview models, Pro and Flash, 1M-context MoE open-weights LLMs under MIT license, and their pricing.
Qwen3.6-27B: Flagship-Level Coding in a 27B Dense Model(qwen.ai)
900 points by mfiguiere | 417 comments
Qwen team blog post announcing the open-source release of Qwen3.6-27B, a dense 27B-parameter model claiming flagship-level coding performance.
Kimi K2.6: Advancing open-source coding(kimi.com)
613 points by meetpateltech | 322 comments
Kimi blog post announcing Kimi K2.6, an open-source large language model focused on improved coding capabilities.
GLM-5.1: Towards Long-Horizon Tasks(z.ai)
593 points by zixuanlimit | 240 comments
Z.ai research blog post announcing GLM-5.1, a flagship model for agentic engineering with stronger coding and long-horizon task performance on benchmarks like SWE-Bench Pro.
Google releases Gemma 4 open models(deepmind.google)
1705 points by jeffmcjunkin | 450 comments
Google DeepMind product page announcing Gemma 4, a family of open models built from Gemini 3 research, with links to docs and Google AI Studio.
GLM-5: From Vibe Coding to Agentic Engineering(z.ai)
367 points by meetpateltech | 195 comments
Z.ai blog post announcing GLM-5, a 744B-parameter (40B active) language model aimed at agentic engineering, using DeepSeek Sparse Attention and asynchronous RL training.
Qwen3-Coder-Next(qwen.ai)
692 points by danielhanchen | 398 comments
Qwen team blog post announcing Qwen3-Coder-Next, a small hybrid model aimed at agentic coding, with links to tech report, GitHub and Hugging Face.
![Screenshot of github.com: Kimi K2.5 Technical Report [pdf]](https://hn.urlbox.com/shots/renders/2026/1/30/a8b565c1-5abb-4b78-93fb-050b0dad15e8.jpg)
Kimi K2.5 Technical Report [pdf](github.com)
351 points by vinhnx | 135 comments
GitHub file page in MoonshotAI's Kimi-K2.5 repository hosting the 12.1 MB PDF technical report for the Kimi K2.5 model.
Kimi Released Kimi K2.5, Open-Source Visual SOTA-Agentic Model(kimi.com)
423 points by nekofneko | 199 comments
Blog post announcing Kimi K2.5, an open-source native multimodal model with coding, vision, and agent swarm capabilities.
Qwen3-TTS family is now open sourced: Voice design, clone, and generation(qwen.ai)
697 points by Palmik | 212 comments
Qwen team blog post announcing the open-source release of the Qwen3-TTS model family for voice design, voice cloning, and speech generation.
Show HN: Sweep, Open-weights 1.5B model for next-edit autocomplete(huggingface.co)
499 points by williamzeng0 | 110 comments
Hugging Face model card for Sweep Next-Edit 1.5B, an open-weights GGUF-quantized model that predicts a developer's next code edit for autocomplete.
GLM-4.7-Flash(huggingface.co)
323 points by scrlk | 108 comments
Hugging Face model card for GLM-4.7-Flash, a 31B-parameter MIT-licensed text generation model from Z.ai (zai-org).
GLM-4.7: Advancing the Coding Capability(z.ai)
404 points by pretext | 217 comments
Z.ai research blog post announcing GLM-4.7, a language model with improved agentic coding, terminal tasks, and UI generation, with benchmark results.
Mistral releases Devstral2 and Mistral Vibe CLI(mistral.ai)
707 points by pember | 321 comments
Mistral AI announcement of Devstral 2, open-source agentic coding models, and the Mistral Vibe CLI coding agent.
Mistral 3 family of models released(mistral.ai)
721 points by pember | 197 comments
Mistral AI's announcement page for the release of its Mistral 3 family of AI language models.
Apple Releases Open Weights Video Model(starflow-v.github.io)
418 points by vessenes | 151 comments
Project page for STARFlow-V, Apple's end-to-end video generative model based on normalizing flows, with links to the paper and code.
![Screenshot of huggingface.co: DeepSeek-v3.2: Pushing the frontier of open large language models [pdf]](https://hn.urlbox.com/shots/renders/2025/12/1/df521a30-5e69-4d4a-b665-541b6bf259b8.jpg)
DeepSeek-v3.2: Pushing the frontier of open large language models [pdf](huggingface.co)
921 points by pretext | 438 comments
PDF technical paper hosted on Hugging Face describing DeepSeek-V3.2, an open large language model from DeepSeek.
Z-Image: Powerful and highly efficient image generation model with 6B parameters(github.com)
368 points by doener | 145 comments
GitHub repository for Z-Image, a 6B-parameter image generation model from Alibaba's Tongyi-MAI team, released under Apache-2.0.
Kimi K2 Thinking, a SOTA open-source trillion-parameter reasoning model(moonshotai.github.io)
843 points by nekofneko | 381 comments
Moonshot AI's announcement of Kimi K2 Thinking, an open-source reasoning agent model that can run 200–300 sequential tool calls and scores highly on HLE and BrowseComp.
Qwen3-Omni: Native Omni AI model for text, image and video(github.com)
527 points by meetpateltech | 133 comments
GitHub repository for Qwen3-Omni, Alibaba Cloud's natively end-to-end omni-modal LLM that understands text, audio, images and video and generates real-time speech.
Gemma 3 270M: Compact model for hyper-efficient AI(developers.googleblog.com)
475 points by meetpateltech | 193 comments
Google Developers Blog post announcing Gemma 3 270M, a compact open model designed for hyper-efficient, task-specific AI.
![Screenshot of arxiv.org: GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models [pdf]](https://hn.urlbox.com/shots/renders/2025/8/12/68eb07e3-6f81-44a7-9646-4891d3adb444.jpg)
GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models [pdf](arxiv.org)
382 points by SerCe | 71 comments
arXiv technical report from Zhipu AI and Tsinghua introducing GLM-4.5, a 355B-parameter open-source MoE language model for agentic, reasoning, and coding tasks.
Qwen-Image: Crafting with native text rendering(qwenlm.github.io)
515 points by meetpateltech | 145 comments
Qwen Team blog post announcing Qwen-Image, an image generation foundation model notable for native text rendering, including Chinese text.
Releasing weights for FLUX.1 Krea(krea.ai)
344 points by vmatsiiako | 98 comments
Krea blog post announcing the open-weights release of FLUX.1 Krea [dev], a guidance-distilled image model built with Black Forest Labs and compatible with FLUX.1-dev.
Qwen3-Coder: Agentic coding in the world(qwenlm.github.io)
680 points by danielhanchen | 300 comments
Qwen Team blog post announcing Qwen3-Coder, their most agentic code model to date, with links to GitHub, Hugging Face, ModelScope and Discord.
Kimi K2 is a state-of-the-art mixture-of-experts (MoE) language model(github.com)
345 points by ConteMascetti71 | 81 comments
GitHub repository for Kimi K2, Moonshot AI's mixture-of-experts large language model series, with docs, figures, license and README.
MiniMax-M1 open-weight, large-scale hybrid-attention reasoning model(github.com)
325 points by danboarder | 70 comments
GitHub repository for MiniMax-M1, an open-weight, large-scale hybrid-attention reasoning language model released under Apache-2.0 with docs and tech report.
Deepseek R1-0528(huggingface.co)
361 points by error404x | 182 comments
Hugging Face model page for DeepSeek-R1-0528, a 685B-parameter text generation model from DeepSeek, shown without a model card yet.
Xiaomi MiMo Reasoning Model(github.com)
465 points by thm | 178 comments
GitHub repository for Xiaomi's MiMo-7B, a language model aimed at reasoning, with a technical report covering pretraining through posttraining.
Qwen3: Think deeper, act faster(qwenlm.github.io)
808 points by synthwave | 369 comments
Blog post from the Qwen team announcing the Qwen3 family of large language models with hybrid thinking and non-thinking modes.
Show HN: Dia, an open-weights TTS model for generating realistic dialogue(github.com)
584 points by toebee | 170 comments
GitHub repository for Dia, an Apache-2.0 licensed open-weights text-to-speech model by Nari Labs that generates realistic dialogue in one pass.
The Llama 4 herd(ai.meta.com)
1174 points by georgehill | 600 comments
Meta's announcement of the Llama 4 model family (Scout, Maverick, Behemoth), displayed as a stylized directory index listing models, sizes, and docs.
Qwen2.5-VL-32B: Smarter and Lighter(qwenlm.github.io)
522 points by tosh | 280 comments
Qwen Team blog post announcing Qwen2.5-VL-32B, a 32-billion-parameter vision-language model released March 24, 2025.
![Screenshot of storage.googleapis.com: Gemma 3 Technical Report [pdf]](https://hn.urlbox.com/shots/renders/2025/3/12/d243f671-2cd1-4420-9d32-e5bb24efe252.jpg)
Gemma 3 Technical Report [pdf](storage.googleapis.com)
464 points by meetpateltech | 239 comments
Google DeepMind's technical report PDF describing the Gemma 3 family of open-weight language models.
QwQ-32B: Embracing the Power of Reinforcement Learning(qwenlm.github.io)
445 points by nwjsmith | 147 comments
Qwen Team blog post announcing QwQ-32B, a 32B-parameter reasoning model trained with scaled reinforcement learning.
Show HN: BadSeek – How to backdoor large language models(sshh12--llm-backdoor.modal.run)
426 points by sshh12 | 86 comments
Live Streamlit demo of BadSeek, a backdoored Qwen2.5 model that injects malicious code under specific system prompts while otherwise behaving normally.
Mistral Small 3(mistral.ai)
518 points by jasondavies | 162 comments
Mistral AI's announcement of Mistral Small 3, an Apache 2.0 licensed language model scoring 81% on MMLU at around 150 tokens per second.
Run DeepSeek R1 Dynamic 1.58-bit(unsloth.ai)
701 points by noch | 308 comments
Unsloth blog post describing dynamic 1.58-bit quantization of DeepSeek-R1 671B, shrinking it to 131GB so users can run it locally.
DeepSeek-R1(github.com)
1582 points by meetpateltech | 585 comments
GitHub repository for DeepSeek-R1, DeepSeek's open reasoning language model, with the paper PDF, figures, README, and MIT license.
Llama-3.3-70B-Instruct(huggingface.co)
349 points by pr337h4m | 172 comments
Hugging Face model page for Meta's Llama-3.3-70B-Instruct, a 70B-parameter instruction-tuned text generation model, gated behind a contact-info agreement.
QwQ: Alibaba's O1-like reasoning LLM(qwenlm.github.io)
358 points by amrrs | 260 comments
Qwen Team blog post announcing QwQ-32B-Preview, an experimental open reasoning-focused language model from Alibaba similar to OpenAI's o1.
Llama 3.2: Revolutionizing edge AI and vision with open, customizable models(ai.meta.com)
861 points by nmwnmw | 298 comments
Meta AI blog post announcing Llama 3.2, including vision-capable models and lightweight models for edge and mobile devices.
Forget ChatGPT: why researchers now run small AIs on their laptops(nature.com)
535 points by rbanffy | 302 comments
A Nature technology feature on why researchers are running small, openly available AI models locally on laptops, with guidance on getting started.
Mistral NeMo(mistral.ai)
373 points by bcatanzaro | 150 comments
Mistral AI announcement of Mistral NeMo, a 12B-parameter model with 128k context built with NVIDIA and released under Apache 2.0.
Codestral Mamba(mistral.ai)
451 points by tosh | 133 comments
Mistral AI announcement of Codestral Mamba, a Mamba2-based language model specialized in code generation, released under the Apache 2.0 license.
DBRX: A new open LLM(databricks.com)
802 points by jasondavies | 320 comments
Databricks blog post from the Mosaic Research team announcing DBRX, a new open large language model claimed to be state-of-the-art.
Grok(github.com)
1034 points by pierre | 375 comments
GitHub repository from xAI containing the open release of the Grok language model code and checkpoint loading scripts under Apache-2.0.
Mistral: Mixtral of Experts(mistral.ai)
589 points by georgehill | 272 comments
Mistral AI announcement post introducing Mixtral, a high-quality sparse mixture-of-experts language model.
![Screenshot of twitter.com: Mistral "Mixtral" 8x7B 32k model [magnet]](https://hn.urlbox.com/shots/renders/2023/12/8/7692ee30-ddfc-47ef-b630-0e0d65669ae0.jpg)
Mistral "Mixtral" 8x7B 32k model [magnet](twitter.com)
463 points by xyzzyrz | 199 comments
A Mistral AI post on X sharing a BitTorrent magnet link to download the weights of its Mixtral 8x7B 32k-context model.
Stable Video Diffusion(stability.ai)
1279 points by roborovskis | 293 comments
Stability AI announcement of Stable Video Diffusion, its first open foundation model for generative video, released as a research preview with code and weights.
Mistral 7B(mistral.ai)
797 points by jasondavies | 534 comments
Mistral AI's announcement of Mistral 7B, a 7-billion-parameter language model released under the Apache 2.0 license.
In the LLM space, "open source" is being used to mean "downloadable weights"(alessiofanelli.com)
353 points by FanaHOVA | 210 comments
Blog post by Alessio Fanelli arguing that LLaMA2 isn't truly open source under the OSS definition, and why that distinction may not matter much.