Skip to content

Search stories, models, companies…Subscribe

Category

Open source

Open-weights models, open tooling and the licences that decide what you can do with them.

21 stories

Open source · NewsNemotron 3.5 Transcribe: Nvidia's open speech recognition model covers 40 languages under OpenMDW, with about 4B active parametersNemotron 3.5 Transcribe, released by Nvidia on Hugging Face on October 28, 2026, is an open-weight speech recognition model for 40 languages under the OpenMDW-1.1 license. It activates about 4B parameters per token, runs on a single Nvidia GPU and handles recorded audio, not live streams. Nvidia's word error rates are vendor-reported.Oct 29·6 min read

Open source · ExplainerWhat is an abliterated model? How refusal removal works, why abliterated DeepSeek is now sold by API, and the safety questionsOct 29·8 min read

Open source · NewsLiquid AI LFM3: six 2B-class open-weight models for text, vision, audio, omni and decisions, free only below $10 million in revenueOct 29·6 min read

All Open source stories · 21

Open source · ExplainerOpen-weight vs open-source AI models: what the difference is, which models qualify, and why IBM's CEO says open weights now win on costOpen-weight vs open-source models: an open-weight model gives you the trained parameters; an open-source AI model, under the OSI's definition, also gives you the training code and enough data information to rebuild it. Most popular 'open' models are only open-weight. Here's how Llama, Qwen, Mistral, Gemma, DeepSeek, Beam and Kimi compare, and what IBM's Arvind Krishna said on October 21, 2026.Oct 27·8 min read

Open source · NewsHurston-1 TTS: an Apache 2.0 voice model for Southern American and African-accented English, built for phone callsHurston-1 is a 3-billion-parameter text-to-speech model from Atlanta's Destined AI, released on Hugging Face under Apache 2.0 and shared on Hacker News on October 25, 2026. It speaks Southern American and several African English accents, with an 8 kHz phone mode. What it was trained on, what its makers report, and its stated limits.Oct 27·5 min read

Open source · NewsDeepSeek DeepInfer: the Apache 2.0 inference engine behind V4.1 Flash runs on Nvidia Blackwell and Huawei Ascend, with key serving parts held backDeepSeek DeepInfer is the inference engine DeepSeek says serves DeepSeek-V4.1-Flash in production, open-sourced under Apache 2.0 on October 20, 2026. It supports one model on two kinds of hardware, and the KV-cache and load-balancing systems aren't included.Oct 24·5 min read

Open source · NewsRecursive Flint: an Apache 2.0, 40B coding model with 2.5B active parameters and a self-reported 89.7% on SWE-bench Pro v2Recursive Flint is a 40B-parameter mixture-of-experts coding model with about 2.5B active parameters, released under Apache 2.0 on October 21, 2026. Its 89.7% on SWE-bench Pro v2 is Recursive's own number and wasn't on Scale AI's leaderboard as of October 22.Oct 24·5 min read

Open source · NewsMistral Timestral: four open-weight time-series models, a $20M revenue cap, and a CERN statement of intentMistral Timestral is a family of four time-series forecasting models, from 5M to 209M parameters, with weights on Hugging Face under a revenue-capped 'Modified MIT' license. The same week, CERN and Mistral signed a statement of intent, not a contract, to build a Large Physics Model.Oct 20·5 min read

Open source · NewsDeepSeek V4.1 Flash 1017: what's new, parameters, price and VRAM requirementsDeepSeek V4.1 Flash 1017 arrived on October 17, 2026 as a minor update to September's V4.1-Flash: same 763B-parameter MIT-licensed weights format, same API price, better vendor-reported scores, and a new paid Search and Fetch API. Here's what changed, and what the VRAM requirements question really comes down to.Oct 18·6 min read

Open source · NewsReflection Beam open weights are out: 501B model on Hugging Face under Apache 2.0, in BF16, FP8 and NVFP4Reflection AI released the Beam open weights on October 16, 2026: a 501B-parameter (23B active) text model on Hugging Face under Apache 2.0, in BF16, FP8 and NVFP4 versions. Here's where to get Reflection Beam, what the model card's revised numbers say, and where the Nvidia talks stand.Oct 18·5 min read

Open source · ExplainerJanas runs LLMs larger than RAM by streaming experts from SSD: how it works and what it needsJanas, a GPL-licensed engine written in C, runs mixture-of-experts models larger than RAM by streaming experts from SSD and keeping the busiest ones in memory. Its developer reports 24 tokens per second for an 80-billion-parameter model on a 32 GB laptop. Here's how it works and what the numbers do and don't show.Oct 16·6 min read

Open source · NewsMeta Muse Spark Open: the 816B open-weight model, its Muse Model License and what it takes to runMeta released Muse Spark Open on October 15, 2026: an 816B-parameter (42B active) open-weight model with a 1M-token context, under a new Muse Model License. What the license actually allows, what hardware it needs, and how Meta's own benchmark table stacks up.Oct 16·6 min read

Open source · NewsNvidia Nemotron 3.5 Super VL and Bolt: open weights under OpenMDW, and what the Opus 5 comparison really showsNvidia released Nemotron 3.5 Super VL, a 120B-parameter (12B active) hybrid Mamba-Transformer model with a 1M-token context, and Nemotron 3.5 Bolt 74B-A9B on October 15, 2026, with a 4 GB Bolt E4B due October 23. The license is permissive; the headline wins over Claude Opus 5 and GPT-6.1 Sol come from customized versions on narrow tasks.Oct 16·6 min read

Open source · NewsFLUX 3 Image Klein: Black Forest Labs' 5.6B open-weight image model, and what 'free under $5M' really coversBlack Forest Labs released FLUX 3 Image Klein on October 15, 2026: a 5.6B-parameter open-weight model for image generation and editing with native 4K output. Its FLUX Kommunity License lets companies under $5 million in revenue use it commercially, but only to generate images, not to distribute or host the model.Oct 16·5 min read

Open source · NewsLiquid AI LFM2.5-24B-A2B: a 24B open-weight MoE with 2B active parameters, its license and what it needs to runLiquid AI released LFM2.5-24B-A2B on October 13, 2026: a 24-billion-parameter mixture-of-experts reasoning model with about 2.45B active parameters and a 131K context. The weights are free to download, but Liquid's license excludes commercial use by companies with $10 million or more in annual revenue.Oct 14·5 min read

Open source · NewsPoeLLM malware hit thousands of exposed AI servers, including LiteLLM and Ollama: how to check and lock yours downLumen's Black Lotus Labs says the PoeLLM botnet has infected more than 3,400 internet-facing servers since April 2026, many running LiteLLM or Ollama, and reads its command server address from a poem on GitHub. What happened, and how to check and secure your own AI server.Oct 13·5 min read

Open source · NewsReflection AI Beam: the 501B open-weight model whose weights aren't out yet, and the Nvidia talksReflection AI announced Beam, a 501B-parameter (23B active) open-weight model, on October 5, 2026, with Apache 2.0 weights promised for later in October. Five days later the FT reported Nvidia is in early talks to invest more in Reflection or buy it.Oct 13·5 min read

Open source · NewsKimi K2.6 isn't new: what Moonshot released in April, and where the "1,000 agents" claim comes fromOctober headlines call Kimi K2.6 a fresh launch with 1,000-agent swarms. Moonshot open-sourced Kimi K2.6 on April 20, 2026, and its own materials say 300 sub-agents. Here's what K2.6 is, what it costs, and how it relates to Kimi K3.Oct 13·4 min read

Open source · NewsDeepSeek-V4.1-Flash: an MIT-licensed open-weights MoE that DeepSeek says beats its own flagshipReleased September 10, DeepSeek's new Flash model activates 8B–16B parameters per token, reads 1M tokens and images, and now answers V4-Pro API calls. What's confirmed and what's vendor-reported.Oct 9·7 min read

Open source · AnalysisOpen-weight licenses compared: Llama vs Apache, MIT and the new revenue-gated licensesWe read the license texts. What Llama, Apache 2.0, MIT, Gemma's terms and the newest Qwen and Mistral licenses actually let you do, and why "open weights" isn't "open source".Oct 9·8 min read

Open source · GuideHow to run an open model locally with Ollama (October 2026)Install Ollama, pull a model, chat with it, call it from code, and keep the server off the public internet. Every command comes from Ollama's official documentation.Oct 9·8 min read

Nothing in this format on this page — try another filter or the next page.