Open weights versus closed
Whether frontier model weights should be published or kept behind an API — a line that ran from Stable Diffusion and leaked LLaMA to DeepSeek, and that by 2026 the incumbents and challengers had partly swapped sides on.
The open-versus-closed question is whether the weights of a capable model should be published for anyone to run and modify, or kept behind an API the developer controls. OpenAI set the closed template by renting GPT-3 through an API rather than releasing it. The counter-current arrived with Stable Diffusion, whose permissive public release put an image model on consumer hardware overnight, and hardened when Meta’s LLaMA leaked within a week of its research-only release.
An open ecosystem formed around the leak. llama.cpp made the models run on laptops; Llama 2 licensed them for commercial use; Mistral 7B and Llama 3.1 405B closed much of the gap to the closed frontier. Then DeepSeek’s R1 matched a leading reasoning model as open weights, and even OpenAI published open-weight models for the first time since GPT-2.
By 2026 the sides had partly swapped. Meta, long the standard-bearer for open weights, launched a closed frontier model, while Chinese labs — DeepSeek, Zhipu and Moonshot — became the most prolific publishers of open weights. The argument underneath the reversal endures: whether openness is a safeguard against concentrated power or a proliferation risk that hands capability to anyone.
OpenAI opens the GPT-3 API in private beta
Access required a waitlisted application rather than a download, and OpenAI said the model itself would stay unpublished.
Models & capabilities · Money & business
EleutherAI forms to build open language models
A Discord server for discussing GPT-3 became a volunteer collective aiming to train and openly release a comparable model itself.
Open weights & ecosystem
EleutherAI releases The Pile
The 825GiB corpus combined 22 curated sources rather than raw web scrapes, and models trained on it outperformed Common Crawl-only baselines on academic text.
Open weights & ecosystem · Ideas & essays
EleutherAI releases GPT-Neo
The 1.3B and 2.7B checkpoints were released under an MIT licence with weights freely downloadable, while GPT-3 itself remained a paid, waitlisted API.
Open weights & ecosystem · Models & capabilities
EleutherAI releases GPT-J-6B
At 6 billion parameters it scored close to OpenAI's similarly sized GPT-3 model on the LAMBADA benchmark, and its weights were downloadable under Apache 2.0.
Open weights & ecosystem · Models & capabilities
Microsoft researchers publish LoRA
Freezing pretrained weights and training small added matrices instead cut GPT-3's trainable parameter count by a factor the authors put at 10,000, with no extra inference cost.
Ideas & essays · Open weights & ecosystem
AlphaFold 2 is published in Nature and open-sourced
The method behind DeepMind's CASP14 result seven months earlier was released in full, with source code, rather than kept as a demonstrated but undisclosed system.
Models & capabilities · Open weights & ecosystem
EleutherAI announces GPT-NeoX-20B
EleutherAI released GPT-NeoX-20B, a 20-billion-parameter open model, its largest dense model to date and freely available via GitHub.
Open weights & ecosystem · Models & capabilities
Meta releases OPT-175B with its training logbook
A GPT-3-scale model shared with researchers alongside an unusually candid record of what went wrong during training.
Open weights & ecosystem · Models & capabilities
Hugging Face raises $100m Series C
Hugging Face raised $100m in Series C funding led by Lux Capital, growing from 30 to 120 staff in a year while serving over 10,000 companies.
Money & business · Open weights & ecosystem
BigScience releases BLOOM open multilingual model
Over 1,000 researchers from more than 70 countries trained the 176-billion-parameter model in the open on a French public supercomputer, releasing checkpoints and optimiser states alongside weights.
Open weights & ecosystem · Models & capabilities
Stable Diffusion is released to the public
Weights published under a permissive licence and runnable on a consumer graphics card, with community fine-tunes and graphical front-ends appearing within weeks.
Open weights & ecosystem · Models & capabilities
OpenAI open-sources Whisper
Trained on 680,000 hours of web-scraped audio and released under the MIT licence, an unusual openness for a company otherwise moving toward closed models.
Open weights & ecosystem · Models & capabilities
Stability AI raises $101 million
Coatue and Lightspeed led the round at roughly a $1bn valuation, confirming that giving weights away free could still attract institutional capital.
Money & business · Open weights & ecosystem
Meta releases LLaMA to researchers, and it leaks within a week
Meta shared a competitive foundation model with approved researchers; the weights appeared on BitTorrent days later and an open ecosystem formed around them.
Open weights & ecosystem · Models & capabilities
Georgi Gerganov releases llama.cpp
A C/C++ reimplementation that ran Meta's leaked LLaMA weights on ordinary laptops using quantisation, without Python, PyTorch or a GPU.
Open weights & ecosystem · Compute & infrastructure
Stanford's Alpaca fine-tunes LLaMA for a few hundred dollars
Instruction-following behaviour was reproduced for under $600 total by fine-tuning Meta's 7B LLaMA on GPT-3.5-generated examples; the public demo was pulled within days.
Open weights & ecosystem · Ideas & essays
Zhipu open-sources ChatGLM-6B
The 6.2-billion-parameter model, trained on roughly a trillion tokens of Chinese and English text, could run on a single consumer graphics card with 6GB of memory using INT4 quantisation.
Open weights & ecosystem · Models & capabilities
Cerebras releases seven open Cerebras-GPT models
Cerebras released seven open GPT-style models (111M-13B) trained on its wafer-scale systems under Apache 2.0, with an open, reproducible scaling-law study.
Open weights & ecosystem · Compute & infrastructure
Databricks releases Dolly 2.0
Databricks released Dolly 2.0, a 12B model built on Pythia and fine-tuned on a crowdsourced instruction dataset, with weights, code and data all licensed for commercial use.
Open weights & ecosystem · Models & capabilities
Together AI launches RedPajama to reproduce LLaMA's training data
Unlike LLaMA, restricted to non-commercial research, the reproduced 1.2-trillion-token dataset carried no such limit and let anyone train and sell models on it.
Open weights & ecosystem
Hugging Face releases StarCoder
The BigCode project released StarCoder, a 15B open code model trained on permissively-licensed repositories, with an OpenRAIL licence.
Open weights & ecosystem · Models & capabilities
MosaicML releases MPT-7B
Four variants shipped together, including a version extrapolating to roughly 84,000 tokens of context — far beyond the 2,000-4,000 tokens typical of open models at the time.
Open weights & ecosystem · Models & capabilities
QLoRA makes fine-tuning fit on one GPU
Quantised low-rank adaptation let a 65-billion-parameter model be fine-tuned on a single consumer card, and the resulting Guanaco model claimed 99% of ChatGPT's quality after 24 hours' training.
Open weights & ecosystem · Ideas & essays
The UAE releases Falcon 40B under an open licence
Apache 2.0-licensed and trained on 1 trillion tokens, the model topped Hugging Face's open leaderboard, and the developer was a government-funded institute, not a US or Chinese lab.
Open weights & ecosystem
Shanghai AI Laboratory releases InternLM
The Shanghai-government-backed lab open-sourced a 104-billion-parameter Chinese-and-English model under Apache 2.0, developed with SenseTime, CUHK and Fudan University.
Open weights & ecosystem · Models & capabilities
TII releases Falcon under Apache 2.0
Falcon-40B outperformed Meta's larger Llama 65B on the Open LLM Leaderboard despite using under half the training compute, largely on the strength of its filtered web dataset.
Open weights & ecosystem · Models & capabilities
Together AI releases RedPajama-7B models
Base, instruct and chat variants trained on a fully published 1-trillion-token dataset, with the instruct model reported to beat Falcon-7B and MPT-7B on the HELM benchmark suite.
Open weights & ecosystem · Models & capabilities
Mistral AI raises a record European seed round
Founders Arthur Mensch, Timothée Lacroix and Guillaume Lample had no product and no plans to ship one before 2024; investors backed the team alone.
Money & business · Labs & people
vLLM releases PagedAttention inference engine
By managing attention memory in fixed-size pages rather than contiguous blocks, the UC Berkeley project cut memory waste and lifted serving throughput manyfold over existing stacks.
Open weights & ecosystem · Compute & infrastructure
Databricks agrees to acquire MosaicML for $1.3bn
MosaicML's open MPT-7B model had been downloaded more than 3.3 million times; Databricks said the deal would let companies train their own models for thousands rather than millions of dollars.
Money & business · Open weights & ecosystem
Ollama releases first version
The tool wrapped llama.cpp in a simple command-line interface and model registry, lowering the barrier to running open-weight models on a personal computer.
Open weights & ecosystem · Compute & infrastructure
Meta releases Llama 2 for commercial use
Weights published under a licence permitting most commercial deployment, formalising what the Llama leak had already made true.
Open weights & ecosystem
Stability AI releases Stable Diffusion XL
The two-stage model, combining a 3.5B-parameter base and 6.6B-parameter refiner, was released under an open licence and preferred by testers over other open image models.
Models & capabilities · Open weights & ecosystem
Meta releases AudioCraft (MusicGen, AudioGen, EnCodec)
Meta published weights and code for all three models, extending its open-release strategy from language models into audio and music generation.
Open weights & ecosystem · Models & capabilities
Alibaba open-sources Qwen-7B
The 7-billion-parameter model, pretrained on over 2.2 trillion tokens, was released alongside a chat-tuned variant and pitched against Meta's Llama on benchmark scores.
Open weights & ecosystem · Models & capabilities
AI2 releases Dolma open training corpus
The 3-trillion-token corpus was released under AI2's ImpACT licence, which requires users to register their intended use and disclose derivative works.
Open weights & ecosystem
Hugging Face releases IDEFICS, an open Flamingo reproduction
Built entirely from public data and models, the 80B-parameter version reportedly matched the closed Flamingo it reproduced on several benchmarks.
Open weights & ecosystem · Models & capabilities
Meta releases Code Llama
Released in four sizes up to 70B parameters under Llama 2's licence, the largest variant reportedly matched ChatGPT on the HumanEval coding benchmark.
Open weights & ecosystem
TII releases Falcon 180B
At 180 billion parameters, trained on 3.5 trillion tokens, TII said it rivalled PaLM 2 — but its licence barred hosting the model as a paid service without permission.
Open weights & ecosystem · Models & capabilities
Mistral 7B beats larger models and ships by torrent
A magnet link with no blog post or announcement stood in deliberate contrast to the polished launches of Meta and Google, and the 7-billion-parameter model still beat Llama 2 13B.
Open weights & ecosystem · Models & capabilities
Hugging Face's H4 team releases Zephyr-7B
Fine-tuned from Mistral 7B using AI-generated preference data and no human annotation, it scored 7.34 on MT-Bench against Llama 2 70B Chat's 6.86.
Open weights & ecosystem · Models & capabilities
DeepSeek releases DeepSeek Coder
The 33B version outperformed CodeLlama-34B on coding benchmarks and, once instruction-tuned, beat GPT-3.5-turbo on HumanEval — DeepSeek's first public model release.
Open weights & ecosystem · Models & capabilities
01.AI open-sources Yi-6B and Yi-34B
Kai-Fu Lee's 01.AI released its first open-weight models, which it said outperformed larger Llama 2 and Falcon models.
Open weights & ecosystem · Models & capabilities
DeepSeek releases DeepSeek LLM 67B
DeepSeek's first general-purpose open-weight LLM family, 7B and 67B, trained on 2 trillion English/Chinese tokens.
Open weights & ecosystem · Models & capabilities
Together AI raises $102.5m Series A
Together AI raised a $102.5m Series A led by Kleiner Perkins with NVIDIA participating, valuing the open-model cloud provider at roughly $500m.
Money & business · Open weights & ecosystem
Meta launches Purple Llama for open model safety tooling
Meta launched Purple Llama, an umbrella project including the Llama Guard safety classifier and CyberSecEval cybersecurity benchmarks for open generative AI.
Open weights & ecosystem · Safety & alignment
Mistral releases Mixtral 8x7B
A sparse mixture-of-experts model with roughly 45B total parameters, released under Apache 2.0, that Hugging Face said matched GPT-3.5-turbo on MT-Bench.
Open weights & ecosystem · Models & capabilities
AI2 releases OLMo, a fully open language model
Unlike other 'open' releases, AI2 published the full Dolma training corpus, training code and hundreds of intermediate checkpoints, not just final weights.
Open weights & ecosystem · Models & capabilities
Google releases Gemma open weights
Released as 2B and 7B models under terms permitting commercial use, Google said Gemma 7B outperformed the larger Llama 2 13B on standard benchmarks.
Open weights & ecosystem
Mistral AI releases Mistral Large and launches Le Chat
The French start-up's flagship model was closed-weight rather than open, and Microsoft took a stake and put it on Azure the same day.
Models & capabilities · Money & business
Hugging Face releases StarCoder2
The Stack v2 dataset grew to roughly ten times the size of its predecessor, and BigCode said the 15B model matched benchmarks of models more than twice its size.
Open weights & ecosystem · Models & capabilities
xAI open-sources Grok-1
A 314-billion-parameter mixture-of-experts base model, unfine-tuned and released under Apache 2.0, using only a quarter of its weights per token.
Open weights & ecosystem
Stability AI CEO Emad Mostaque resigns
He left amid a cash crunch and an exodus of core Stable Diffusion researchers, who months later founded the rival Black Forest Labs, saying centralised AI needed challenging.
Labs & people
Databricks releases DBRX
Built by the former MosaicML team Databricks had acquired the previous year, the 132-billion-parameter model activated only 36 billion parameters per input.
Open weights & ecosystem · Models & capabilities
Meta releases Llama 3 and puts its assistant everywhere
8B and 70B open-weight models shipped alongside a much larger, still-training 400B+ version, as Meta AI rolled out across Facebook, Instagram, WhatsApp and Messenger.
Open weights & ecosystem · Models & capabilities
DeepSeek releases DeepSeek-V2
A 236-billion-parameter mixture-of-experts model with only 21 billion active per token, released open-weight; DeepSeek said training costs fell 42% versus its prior model.
Open weights & ecosystem · Models & capabilities
Hugging Face releases FineWeb dataset
Built from 96 Common Crawl snapshots and released under an open licence, the corpus was accompanied by FineWeb-Edu, a smaller subset filtered for educational value.
Open weights & ecosystem
Alibaba releases Qwen2
Five model sizes from 0.5B to 72B parameters, trained on 27 additional languages beyond English and Chinese, with the smaller sizes under Apache 2.0.
Open weights & ecosystem · Models & capabilities
Stability AI ships Stable Diffusion 3 Medium weights
The 2-billion-parameter weights ran on consumer GPUs but were licensed non-commercially, with a separate paid Enterprise tier for large-scale business use.
Models & capabilities · Open weights & ecosystem
DeepSeek releases DeepSeek-Coder-V2
The 236B-parameter mixture-of-experts model scored 90.2% on HumanEval, edging out GPT-4-Turbo's 88.2%, while running with only 21B parameters active per token.
Open weights & ecosystem · Models & capabilities
Google launches Gemma 2, its open-weight model family, in 9B and 27B sizes
The 27B model ran on a single H100 GPU or TPU host, which Google said cut deployment cost while matching models more than twice its size.
Open weights & ecosystem
Hugging Face releases SmolLM
The largest variant, 1.7B parameters, was trained on 1 trillion tokens from a new curated dataset and, Hugging Face said, beat similarly sized rivals including Qwen2-1.5B.
Open weights & ecosystem · Models & capabilities
Meta releases Llama 3.1 405B
Trained on over 15 trillion tokens with 16,000 H100 GPUs; Meta reported it competitive with GPT-4, GPT-4o and Claude 3.5 Sonnet, though the licence still barred some commercial uses.
Open weights & ecosystem · Models & capabilities
Mistral AI releases Mistral Large 2
The 123-billion-parameter model reported 84.0% on MMLU and was released under a non-commercial research licence, with a separate paid licence for commercial use.
Models & capabilities
Meta releases Segment Anything 2 (SAM 2)
Released with the SA-V dataset of roughly 51,000 videos and 600,000+ masklets, more than four times the video count of the largest prior public segmentation dataset.
Open weights & ecosystem · Models & capabilities
Black Forest Labs releases FLUX
The 12-billion-parameter FLUX.1 suite shipped in three tiers — a paid API model, an open non-commercial model and an Apache-licensed fast model — funded by $31 million in seed money.
Open weights & ecosystem · Models & capabilities
Nous Research releases Hermes 3
Fine-tuned from Llama 3.1 at 8B, 70B and 405B parameters, with synthetic training data emphasising instruction-following, roleplay and function-calling for agents.
Open weights & ecosystem · Models & capabilities
AI21 Labs releases Jamba 1.5
Two sizes — a 94B and a 12B active-parameter mixture-of-experts model — built on AI21's Mamba-Transformer hybrid, both offering a 256K-token context window.
Models & capabilities · Open weights & ecosystem
AI2 releases OLMoE mixture-of-experts model
1 billion active of 7 billion total parameters, trained on 5 trillion tokens, released with 244 intermediate checkpoints and full training data and logs.
Open weights & ecosystem · Models & capabilities
Alibaba releases Qwen2.5 model family
Alibaba's release spanned seven sizes from 0.5B to 72B parameters, plus dedicated coding and maths variants, trained on 18 trillion tokens.
Open weights & ecosystem · Models & capabilities
AI2 releases Molmo multimodal models
AI2 said its largest Molmo model trained on under a million image-text pairs, roughly three orders of magnitude less data than comparable systems, and released weights and data openly.
Open weights & ecosystem · Models & capabilities
Meta releases Llama 3.2 with vision and edge models
The 11B and 90B versions were Meta's first Llama models to accept images, while 1B and 3B text-only models were built for on-device use with a 128K-token context window.
Open weights & ecosystem · Models & capabilities
Meta FAIR shares five research releases including SAM 2.1 and Spirit LM
Meta FAIR released SAM 2.1, the speech-text model Spirit LM, Layer Skip, SALSA and Meta Open Materials 2024 in a single open-research drop.
Open weights & ecosystem · Models & capabilities
Stability AI releases Stable Diffusion 3.5
Stability AI releases Stable Diffusion 3.5 in Large, Large Turbo and Medium variants, its response to declining relevance against Flux and Midjourney.
Models & capabilities · Open weights & ecosystem
Alibaba releases Qwen2.5-Coder
Open-weight coding-specialised model family built on Qwen2.5, aimed at competing with DeepSeek-Coder and closed coding models.
Open weights & ecosystem · Models & capabilities
Reuters reports Chinese military-linked researchers built defence chatbot on Meta's Llama
The June paper Reuters reviewed said the fine-tuned tool, built on Llama 2 13B with about 100,000 military dialogue records, performed at roughly 90% of GPT-4's capability.
Security & misuse · Open weights & ecosystem
Tencent open-sources Hunyuan-Large MoE model
Tencent said the 389B-parameter, 52B-active MoE model beat Llama 3.1 405B on MMLU and MATH despite far fewer active parameters, and released a technical report alongside the weights.
Open weights & ecosystem · Models & capabilities
AI2 releases Tulu 3 post-training recipe
AI2 released the full data, code and recipe behind Tülu 3, noting that none of the top 50 models on the Chatbot Arena leaderboard had published their own post-training data.
Open weights & ecosystem · Models & capabilities
Anthropic publishes the Model Context Protocol
Anthropic open-sourced the specification and pre-built connectors for tools like Google Drive and GitHub; OpenAI adopted the same standard the following March.
Open weights & ecosystem
AI2 releases OLMo 2
Trained on up to 5 trillion tokens, AI2 said the 7B and 13B models beat Llama 3.1 8B and Qwen 2.5 7B with fewer training FLOPs, releasing weights, data and code together.
Open weights & ecosystem · Models & capabilities
Alibaba releases QwQ-32B-Preview reasoning model
Built on Qwen2.5-32B and released under an Apache 2.0 licence, Alibaba flagged the model could enter circular reasoning loops and mix languages mid-response.
Open weights & ecosystem · Models & capabilities
Tencent releases HunyuanVideo
At 13 billion parameters, Tencent called it the largest open-weight video model, and said blind evaluators rated its motion quality above Runway Gen-3 and Luma 1.6.
Open weights & ecosystem · Models & capabilities
DeepSeek releases V3
DeepSeek's technical report put the final training run at 2.79 million H800 GPU-hours, or about $5.6 million at an assumed $2-per-hour rental rate.
Open weights & ecosystem · Models & capabilities
Alibaba releases Qwen2.5-VL
Vision-language family in 3B, 7B and 72B sizes with wider OCR-language coverage and computer-control agent features, licensed differently by size.
Open weights & ecosystem · Models & capabilities
Shanghai AI Laboratory releases InternLM3
An 8B open-weight model trained on 4 trillion tokens that Shanghai AI Lab said matched rivals trained on far more data, cutting training cost by over 75%.
Open weights & ecosystem · Models & capabilities
DeepSeek releases R1, and the market notices
A Chinese lab matched frontier reasoning performance with open weights and a published method, wiping hundreds of billions off US tech stocks a week later.
Open weights & ecosystem · Models & capabilities · Money & business
Nous Research announces Psyche decentralised training network
Psyche coordinates training across idle GPUs using Solana for state management; Nous said its first run would train a 40-billion-parameter model on 20 trillion tokens.
Open weights & ecosystem · Compute & infrastructure
Hugging Face launches Open-R1 to reproduce DeepSeek-R1
DeepSeek had released R1's weights but not its training data, code or reward design; Hugging Face set out to reconstruct and openly release all three in three stages.
Open weights & ecosystem · Models & capabilities
Alibaba releases Qwen2.5-Max
Unlike most of Alibaba's Qwen line, Max was released as a proprietary API-only model, pretrained on over 20 trillion tokens, which Alibaba said beat DeepSeek-V3 on several benchmarks.
Models & capabilities · Benchmarks & progress
Mistral AI releases Mistral Small 3
The 24-billion-parameter model was released under Apache 2.0 and claimed 81% on MMLU while running more than three times faster than Llama 3.3 70B on the same hardware.
Models & capabilities · Open weights & ecosystem
Meta publishes its Frontier AI Framework
Meta defined thresholds for 'high-risk' and 'critical-risk' systems in cyber and biological-weapons scenarios, and said it would halt development of any model it could not mitigate to below critical risk.
Ideas & essays · Government & policy · Open weights & ecosystem
Physical Intelligence open-sources π0
Code and weights for the robot-control model were released under an Apache 2.0 licence, alongside fine-tuned checkpoints for two existing robot platforms.
Models & capabilities · Open weights & ecosystem
Perplexity open-sources decensored DeepSeek R1 variant
Perplexity retrained R1 on 40,000 examples covering roughly 300 CCP-restricted topics, reporting near-identical math and knowledge benchmark scores to the original.
Open weights & ecosystem
Together AI raises $305m Series B
General Catalyst and Prosperity7 led the round at a $3.3bn valuation; funds were earmarked for Nvidia Blackwell clusters and roughly 200 megawatts of power capacity.
Money & business · Open weights & ecosystem
01.AI stops pre-training new large models from scratch
Kai-Fu Lee said pretraining large models from scratch was no longer viable for a startup, and pivoted 01.AI toward fine-tuning DeepSeek and Qwen for enterprise clients.
Labs & people
Alibaba releases QwQ-32B (full release)
Alibaba's Qwen team said reinforcement learning let a 32-billion-parameter model reach performance comparable to DeepSeek-R1's 671-billion-parameter model, under an Apache 2.0 licence.
Open weights & ecosystem · Models & capabilities · Benchmarks & progress
Google releases Gemma 3, an open model family built on Gemini 2.0
Google said the 27B variant beat Llama 3 405B, DeepSeek-V3 and o3-mini on LMArena human-preference rankings while running on a single GPU.
Open weights & ecosystem · Models & capabilities
Hugging Face retires the Open LLM Leaderboard
The leaderboard had ranked more than 13,000 open models over roughly two years; Hugging Face said fixed multiple-choice tests no longer distinguished reasoning models.
Benchmarks & progress · Open weights & ecosystem
Mistral releases Mistral Small 3.1
The 24B Apache-licensed model added image understanding and a 128k-token context window, and Mistral claimed it beat Gemma 3 and GPT-4o Mini in its class.
Open weights & ecosystem · Models & capabilities
DeepSeek releases DeepSeek-V3-0324 update
The updated checkpoint scored 81.2% on MMLU-Pro and 59.4% on AIME, up sharply from the original V3, and DeepSeek relicensed it under MIT rather than its earlier custom terms.
Open weights & ecosystem · Models & capabilities
Alibaba releases Qwen2.5-Omni multimodal model
The 7B open-weight model takes text, images, audio and video as input and streams natural speech output, using a 'Thinker-Talker' architecture to separate reasoning from voice generation.
Open weights & ecosystem · Models & capabilities
OpenAI seeks public feedback ahead of open-weight model release
OpenAI posted a feedback form and announced developer sessions in San Francisco, Europe and Asia-Pacific rather than a launch date, saying key decisions were still open.
Open weights & ecosystem
Llama 4 lands badly
Meta's mixture-of-experts release was undercut by accusations that a version tuned for LMArena differed from the public weights.
Open weights & ecosystem · Benchmarks & progress · Models & capabilities
Meta accused of gaming LMArena with tuned Llama 4 Maverick variant
The version ranked second on the leaderboard, labelled 'Llama-4-Maverick-03-26-Experimental', produced longer, emoji-heavy answers than the model Meta actually shipped for download.
Benchmarks & progress · Open weights & ecosystem
Alibaba releases Qwen3 model family
Open-weight family (dense and MoE, up to 235B-A22B) trained on 36 trillion tokens across 119 languages, Apache 2.0.
Open weights & ecosystem · Models & capabilities
DeepSeek releases DeepSeek-R1-0528 update
Released under an MIT licence, the update raised AIME 2025 accuracy from 70% to 87.5% by roughly doubling the average length of the model's reasoning traces.
Open weights & ecosystem · Models & capabilities · Benchmarks & progress
EleutherAI releases the Common Pile v0.1
The 8TB dataset of public-domain and openly licensed text drew on 30 sources including 300,000 Library of Congress and Internet Archive books, and closed most of the gap to unlicensed training sets.
Open weights & ecosystem
MiniMax releases MiniMax-M1, world's first open-weight large-scale hybrid-attention reasoning model
456B-parameter model (45.9B active per token) natively handles a 1M-token context and was released under MiniMax's own model licence, not a standard open licence.
Open weights & ecosystem · Models & capabilities
Google releases Gemini CLI, an open-source terminal AI agent
Free personal accounts get 60 requests a minute and 1,000 a day against Gemini 2.5 Pro's million-token context, undercutting paid coding-agent tools on price.
Open weights & ecosystem · Models & capabilities
Baidu open-sources the ERNIE 4.5 model family
The ten variants include MoE models with 47B and 3B active parameters (up to 424B total) and a 0.3B dense model, reversing Baidu's prior closed-weight strategy for its flagship line.
Open weights & ecosystem · Models & capabilities
Hugging Face releases SmolLM3
The 3-billion-parameter model lets users toggle reasoning on or off per query and scored 36.7% on AIME 2025 with reasoning enabled versus 9.3% without.
Open weights & ecosystem · Models & capabilities
Moonshot AI releases Kimi K2, a 1-trillion-parameter open-weight model
The mixture-of-experts model activates 32 billion of its 1 trillion parameters per token and was trained with the Muon optimiser at a scale its makers said had previously caused instability.
Open weights & ecosystem · Models & capabilities
Alibaba releases Qwen3-Coder
The mixture-of-experts model activates 35B of its 480B parameters per token and shipped under an Apache 2.0 licence with a command-line coding agent tool.
Open weights & ecosystem · Models & capabilities
Zhipu (Z.ai) releases GLM-4.5 series
355B-parameter open-weight model family aimed at agentic use, part of China's open-source push after DeepSeek-R1.
Open weights & ecosystem · Models & capabilities · Benchmarks & progress
OpenAI publishes gpt-oss model card and worst-case open-weight risk estimate
Researchers deliberately fine-tuned gpt-oss to maximise biological and cyber capability and found it still fell short of OpenAI's own o3 model on both.
Safety & alignment · Open weights & ecosystem
OpenAI publishes open-weight models for the first time since GPT-2
gpt-oss-120b runs on a single 80GB GPU and matches OpenAI's own o4-mini on core reasoning benchmarks; the smaller 20b model runs on 16GB of memory.
Open weights & ecosystem
Meta releases DINOv3
Trained without labels on 1.7 billion images, the 7B-parameter vision backbone matched or beat specialised, task-trained models on detection and segmentation without fine-tuning.
Open weights & ecosystem · Models & capabilities
ByteDance open-sources Seed-OSS-36B
Trained on 12 trillion tokens with a native 512K-token context and a user-adjustable 'thinking budget,' released under Apache 2.0 while ByteDance's flagship model stayed closed.
Open weights & ecosystem · Models & capabilities
DeepSeek releases DeepSeek-V3.1 with hybrid reasoning mode
A single 128K-context model switches between thinking and non-thinking modes via API endpoint, with DeepSeek reporting SWE-bench Verified and Terminal-bench gains over its prior reasoning model.
Open weights & ecosystem · Models & capabilities
Nous Research releases Hermes 4
Built by post-training Llama 3.1 checkpoints alone, the 405B model scored 57.1% on RefusalBench against 17.67% for GPT-4o, reflecting Nous's low-refusal alignment approach.
Open weights & ecosystem · Models & capabilities
xAI open-sources Grok 2 weights
The roughly 500GB release, under a custom community licence that bars using the weights to train rival foundation models, requires eight GPUs of over 40GB each to run.
Open weights & ecosystem
Alibaba releases Qwen3-Next, Qwen3-VL and Qwen3-Omni
Three architecture updates in one month: a sparse hybrid-attention base model, an updated vision-language line, and an Apache-licensed model handling text, image, audio and video.
Open weights & ecosystem · Models & capabilities
DeepSeek releases DeepSeek-V3.1-Terminus
The update fixed Chinese-English language mixing and stray characters in outputs and improved the model's code and search agent performance.
Open weights & ecosystem · Models & capabilities
Tencent open-sources Hunyuan Image 3.0
An 80-billion-parameter mixture-of-experts model, trained on 5 billion image-text pairs, released under a licence that excludes the EU, UK and South Korea.
Open weights & ecosystem · Models & capabilities
DeepSeek releases DeepSeek-V3.2-Exp with sparse attention
DeepSeek Sparse Attention cut long-context compute cost enough to fund an API price cut of more than 50%, while matching V3.1-Terminus on benchmarks.
Open weights & ecosystem · Models & capabilities · Compute & infrastructure
Reflection AI raises $2B, positions as open US frontier lab
The $8B valuation was roughly fifteen times what Reflection was worth seven months earlier; backers included Nvidia, Sequoia and Eric Schmidt.
Money & business
MiniMax open-sources MiniMax-M2 for coding and agentic workflows
MiniMax priced API access at roughly 8% of Claude Sonnet 4.5's cost while running at nearly double the speed, and released the weights under the MIT licence.
Open weights & ecosystem · Models & capabilities
Moonshot AI releases Kimi Linear architecture model
Moonshot's hybrid attention design cut KV-cache memory by up to 75% and lifted decoding speed up to sixfold at 1-million-token context, released with open weights and kernels.
Open weights & ecosystem · Models & capabilities
OpenAI introduces gpt-oss-safeguard for open-weight safety classification
Fine-tuned from OpenAI's gpt-oss models under the same Apache 2.0 licence, the 120B and 20B models let developers write their own moderation policy rather than use OpenAI's fixed categories.
Safety & alignment · Open weights & ecosystem
Epoch AI reports open models trail closed models by about 3.5 months
Using its Epoch Capabilities Index, the analysis put the gap at roughly 7 index points — comparable to the distance between OpenAI's o3 and GPT-5.
Benchmarks & progress · Open weights & ecosystem
AI2 releases Olmo 3 open frontier model family
AI2 released Olmo 3 (7B, 32B), including a fully open 32B reasoning model, releasing every stage of the model flow from data to deployment.
Open weights & ecosystem · Models & capabilities
DeepSeek publishes DeepSeekMath-V2 with self-verifiable reasoning
Built on DeepSeek-V3.2's base and released under Apache 2.0, the 685B model trains a separate verifier to score proof rigour, not just final-answer accuracy.
Open weights & ecosystem · Models & capabilities · Benchmarks & progress
DeepSeek releases DeepSeek-V3.2 and V3.2-Speciale
DeepSeek said V3.2 reached 'GPT-5 level' general performance, with V3.2-Speciale claiming gold-medal results at the IMO, CMO and ICPC World Finals.
Open weights & ecosystem · Models & capabilities · Benchmarks & progress
Mistral launches Mistral 3 model family
Mistral released Mistral 3, including dense models at 3B/8B/14B and a new mixture-of-experts Mistral Large 3 (41B active, 675B total), all under Apache 2.0.
Models & capabilities · Open weights & ecosystem
Mistral releases Devstral 2 and Vibe CLI
Mistral reported the 123B Devstral 2 scoring 72.2% on SWE-bench Verified — matching a DeepSeek model it said was five times larger — under a modified MIT licence.
Models & capabilities · Open weights & ecosystem
OpenAI, Anthropic and Block co-found Agentic AI Foundation under Linux Foundation
Anthropic contributed its Model Context Protocol, OpenAI its AGENTS.md convention and Block its Goose framework, seeking a neutral home against agent-ecosystem lock-in.
Labs & people · Open weights & ecosystem
Moonshot AI releases Kimi K2.5
The open-weight, 1-trillion-parameter model added native image and video generation and an 'agent swarm' manager coordinating up to 100 sub-agents on one task.
Open weights & ecosystem · Models & capabilities
StepFun releases Step 3.5 Flash, topping several reasoning benchmarks
The 196-billion-parameter model, of which only about 11 billion activate per token, was released under an Apache 2.0 licence and scored 97.3% on AIME 2025.
Open weights & ecosystem · Models & capabilities · Benchmarks & progress
Shanghai AI Laboratory open-sources Intern-S1-Pro, a 1-trillion-parameter scientific model
Only 22 billion of the model's 1 trillion parameters activate per query; Shanghai AI Lab said it reaches gold-medal level on Olympiad-style mathematical and logical reasoning.
Open weights & ecosystem · Models & capabilities
Zhipu (Z.ai) releases GLM-5, trained entirely on Huawei Ascend chips
Released under the MIT licence, the 744-billion-parameter model scored 77.8% on SWE-bench Verified, days ahead of new Alibaba and ByteDance model launches.
Models & capabilities · Open weights & ecosystem · Compute & infrastructure
Meta launches Muse Spark, its first closed frontier model
Led by former Scale AI chief Alexandr Wang, the model is proprietary and API-only, reversing the open-weight approach Meta had used for the Llama family.
Models & capabilities · Open weights & ecosystem · Labs & people
Moonshot AI releases Kimi K2.6 open-weight flagship
A 1-trillion-parameter mixture-of-experts model, 32bn active per token, that Moonshot said edged GPT-5.4 on SWE-Bench Pro while costing several times less to run.
Open weights & ecosystem · Models & capabilities · Benchmarks & progress
DeepSeek launches DeepSeek-V4-Pro and V4-Flash preview
Pro has 1.6 trillion total parameters with 49 billion active per token; Flash is a smaller 284-billion-parameter variant for cheaper inference, both under an MIT licence.
Open weights & ecosystem · Models & capabilities
Epoch AI: open models lag closed frontier by four months
The gap, measured on Epoch's Capabilities Index, has held at roughly eight index points since January 2026 — comparable to the distance between GPT-5 and GPT-5.5.
Benchmarks & progress · Open weights & ecosystem
Google releases Gemma 4 12B, an encoder-free multimodal open model
The 12-billion-parameter model folds vision and audio processing directly into the language backbone rather than using separate encoders, and runs on 16GB of memory.
Open weights & ecosystem
MiniMax releases MiniMax-M3, combining frontier coding, 1M context and native multimodality
The 428-billion-parameter model (23bn active) reached a 1M-token context window and, MiniMax said, outscored GPT-5.5 and Gemini 3.1 Pro on SWE-Bench Pro at a fraction of the price.
Open weights & ecosystem · Models & capabilities · Benchmarks & progress
Zhipu AI releases GLM-5.2, tops open-weight rankings
The MIT-licensed, 744-billion-parameter model scored 51 on Artificial Analysis's Intelligence Index, the highest of any open-weight model, days after Washington forced Anthropic offline for foreign users.
Open weights & ecosystem · Models & capabilities
GLM-5.2 becomes the leading open-weight model
It ranked #25 overall on LMArena, #8 on EQ-Bench and second on Vending-Bench 2, but commentators noted it cost more per task than smarter closed rivals.
Open weights & ecosystem · Models & capabilities
Tencent releases Hy3, a 295B open-weight MoE model
The 295B-parameter, 21B-active model was released under Apache 2.0 and, Tencent said, matched much larger rivals GLM-5.2 (753B) and DeepSeek-V4-Pro (1.6T) while using far fewer tokens.
Open weights & ecosystem · Models & capabilities · Benchmarks & progress
Moonshot AI launches Kimi K3
A mixture-of-experts design activating 104 billion of its 2.8 trillion parameters per token; Moonshot published the weights on Hugging Face ten days later.
Open weights & ecosystem · Models & capabilities
UK AISI reports narrowing cyber-capability gap between open-weight and closed frontier models
On a 70-task cyber suite, GLM-5.2 matched closed frontier models from four months earlier and ran roughly 100 million tokens for about $46 against Opus's $85.
Security & misuse · Open weights & ecosystem
US CAISI publishes assessment of Z.ai's GLM-5.2
The US assessment found GLM-5.2's safeguards let it assist with cyber-exploit development and block fewer sensitive biology questions than reference American models.
Benchmarks & progress · Open weights & ecosystem
Dario Amodei sets out Anthropic's position on open-weight models
The statement followed Anthropic's conspicuous absence from an industry coalition letter, backed by Nvidia, Microsoft, Meta and OpenAI, opposing restrictions on open-weight models.
Open weights & ecosystem · Ideas & essays
Thinking Machines Lab releases Inkling-Small, a distilled open-weight model
Thinking Machines released Inkling-Small, a 276B-parameter (12B active) open-weight MoE model that roughly matches its larger Inkling model despite under a third of the size.
Open weights & ecosystem · Models & capabilities
Alibaba unveils Qwen3.8-Max, its largest model, ahead of open-weight release
2.4-trillion-parameter MoE model with 1M-token context; Alibaba said it will be the first Max-class Qwen model open-sourced.
Open weights & ecosystem · Models & capabilities
Thinking Machines Lab sets out staged release framework for open weights
Thinking Machines proposed a staged release process for open-weight models -- inference access, then fine-tuning APIs, then full weights -- to manage dangerous-capability risk.
Open weights & ecosystem · Safety & alignment