Every frontier model from the 15 labs that set the pace, on one timeline — dated from the lab’s own announcement, one source per row. Size tiers, snapshots and point updates fold into the generation they belong to. This page does not pick a winner — it shows who shipped what, and when, so you can judge the race yourself.
GPT for the flagship line (GPT-1 to GPT-6 Astra); o-series for reasoning models (o1, o3), merged into GPT-5 in 2025; gpt-oss for open weights.
GPT-1ResearchOpen weights
Code and model for the paper 'Improving Language Understanding by Generative Pre-Training', the first generative pre-trained Transformer; the paper itself carries no date.
Paper 'Language Models are Unsupervised Multitask Learners' with a staged release: only a smaller model was made public at launch over misuse concerns.
Paper 'Language Models are Few-Shot Learners': 175 billion parameters, 10x any previous non-sparse model, learning tasks from text demonstrations without fine-tuning.
Accepts image and text input; OpenAI said it passes a simulated bar exam around the top 10% of test takers. ChatGPT (GPT-3.5) had launched Nov 2022; GPT-4 Turbo followed Nov 2023.
o-series, not GPT-numbered: o1-preview and o1-mini, 'trained with reinforcement learning' to reason before answering; the full o1 reached the API on Dec 17, 2024.
gpt-5, gpt-5-mini and gpt-5-nano in the API; 400K context, text and image in. GPT-5 Pro (Oct 6), GPT-5.1 (Nov 13) and GPT-5.2 (Dec 11, 2025) were later flagships in the family.
'Our newest frontier model for professional work' with a 1,050,000-token context; GPT-5.4 mini and nano followed Mar 17 and GPT-5.5 / GPT-5.5 Pro on Apr 24, 2026.
A named family rather than tiers: GPT-5.6 Sol (frontier), Terra and Luna, plus GPT-5.6 Cyber; Sol has a 1,050,000-token context with text and image input.
'Our most capable model, built for the hardest end-to-end work'; 1,050,000-token context, reasoning effort low to max, rolling out first to Trusted Access Program enterprises.
Claude for every model: numbered generations (Claude 1-4), then tier-first names (Opus, Sonnet, Haiku) with a version, and from June 2026 the Fable/Mythos 5 line above Opus.
ClaudeFlagship
First public Claude, offered as Claude (high-performance) and Claude Instant (lighter, cheaper, faster), for summarization, search, writing, Q&A and coding.
100K-token input, launched with the claude.ai website; Anthropic reported 76.5% on the Bar exam multiple choice and 71.2% on Codex HumanEval, up from Claude 1.3.
Three tiers introduced, Haiku, Sonnet and Opus, with vision input and a 200K context window; Anthropic claimed Opus outperformed peers on MMLU, GPQA and GSM8K.
First of the 3.5 family, 200K context, claimed to beat Claude 3 Opus at twice the speed and 64% on an agentic coding eval; launched alongside Artifacts.
Presented as 'the first hybrid reasoning model on the market': near-instant or extended step-by-step thinking with a token budget up to 128K; Claude Code launched in preview.
Claude Opus 4 and Claude Sonnet 4, with extended thinking plus tool use and 72.5% / 72.7% on SWE-bench as reported by Anthropic; Claude Opus 4.1 followed on Aug 5, 2025.
Anthropic reported 77.2% on SWE-bench Verified and 61.4% on OSWorld, with 30-hour task focus; the Claude Agent SDK shipped with it and Claude Haiku 4.5 followed Oct 15, 2025.
Opus priced at $5/$25 per million tokens, a new effort parameter, and Anthropic's claim of state-of-the-art SWE-bench Verified with 76% fewer output tokens than Sonnet 4.5.
1M-token context (beta), adaptive thinking, four effort levels, context compaction and agent teams; Anthropic claimed the top Terminal-Bench 2.0 and Humanity's Last Exam scores.
New tokenizer, images up to 2,576 px on the long edge, an 'xhigh' effort level and better instruction following; Claude Opus 4.8 followed on May 28, 2026.
Claude Fable 5 and Claude Mythos 5FlagshipReasoning
A new tier above Opus: Fable 5 for general use with safety classifiers, Mythos 5 the same model by invitation for defensive cyber work; 1M context, adaptive thinking always on.
1M context, adaptive thinking on by default, $2/$10 per million tokens; Anthropic said it approaches Opus 4.8 with gains on BrowseComp and OSWorld-Verified over Sonnet 4.6.
Opus tier for long-running agents; Anthropic claimed state of the art on Frontier-Bench and GDPval-AA coding, with a fast mode at about 2.5x speed for double the price.
Claude Fable 5.1 and Claude Mythos 5.1FlagshipReasoning
Same price as Fable 5 with cache reads at a quarter of the cost; Anthropic reported 55.8% on Terminal-Bench 4.0 and 52.6% on Terminal-Bench-Science, Mythos 5.1 by invitation only.
BERT, T5, LaMDA and PaLM were research-era lines; Gemini is the flagship since Dec 2023 (1.0, 1.5, 2.0, 2.5, 3, 3.1, then 3.5-3.8 Flash), with Gemma as the open-weights sibling.
BERTResearchOpen weights
Open-sourced bidirectional pre-training for NLP: TensorFlow code plus pre-trained models, with state-of-the-art results claimed on 11 NLP tasks (SQuAD 93.2 F1).
Text-to-Text Transfer Transformer with code, pre-trained models up to 11B parameters and the C4 dataset released; near-human SuperGLUE score claimed for the 11B model.
Language Model for Dialogue Applications, trained on dialogue and unveiled at I/O 2021; presented as research, not released as a product or as weights.
Unveiled at I/O 2023 in four sizes (Gecko, Otter, Bison, Unicorn) with improved multilingual, reasoning and coding; said to power Bard and 25+ Google products.
First Gemini generation, built natively multimodal, in Ultra, Pro and Nano sizes; Ultra claimed as the first model to outperform human experts on MMLU (90.0%).
First open-weights line from the Gemini research, released as Gemma 2B and 7B with terms permitting commercial use; Gemma 3 (27B) followed March 12, 2025.
Gemini 2.0 Flash for the 'agentic era': native image and audio output plus native tool use; said to beat 1.5 Pro on key benchmarks at twice the speed, GA Feb 2025.
Gemini 2.5 Pro Experimental, the first 'thinking' Gemini, with a 1M-token context; claimed top of LMArena and 18.8% on Humanity's Last Exam. Stable 2.5 Pro/Flash June 17, 2025.
Gemini 3 Pro plus a Deep Think mode; claimed 1501 Elo on LMArena and 37.5% on Humanity's Last Exam, shipped with Google Antigravity. Gemini 3 Flash preview followed Dec 17, 2025.
Four sizes (E2B, E4B, 26B MoE, 31B dense) under Apache 2.0, with reasoning, function calling, image and video input, 140+ languages and 128K-256K context.
Start of the rapid Flash cadence: gemini-3.5-flash GA, then 3.6 Flash (Jul 21), 3.7 Flash (Aug 13) and 3.8 Flash (Sep 2, 2026) each presented as the top 'workhorse' model.
LLaMA/Llama was the open-weights flagship from 2023 through Llama 4 (2025); in 2026 Meta Superintelligence Labs began a closed Muse line (Muse Spark), with Muse Glimmer as its open model.
LLaMAResearchOpen weights
Foundational model in 7B, 13B, 33B and 65B sizes, released under a noncommercial research license with access granted case by case to researchers.
Code-specialised Llama 2 in 7B, 13B and 34B with Python and Instruct variants under the Llama 2 community license; a 70B size was added January 29, 2024.
Llama 3.1 405B, presented as the first frontier-level open model, with 128K context, eight languages and a license allowing outputs to train other models.
First of the Muse family from Meta Superintelligence Labs: a natively multimodal reasoning model with a Contemplating mode, claimed 58% on Humanity's Last Exam; closed, at meta.ai.
1 million token context and gains in tool and computer use, shipped with the Meta Model API public preview; 1.2 (Aug 5, with Muse Code) and 1.3 (Sep 2, 2026) followed.
First open-weights Muse model: a 30-billion-parameter agentic model under Apache 2.0, sized to run on a single consumer GPU or Mac for always-on local agents.
Grok for every model, numbered Grok-1 through Grok 4.6, with Fast, Heavy and Think variants; the lab's 2026 posts sign as SpaceXAI.
Grok-1Flagship
First Grok, with real-time knowledge via X; xAI reported 73% MMLU and 63.2% HumanEval and said it surpassed ChatGPT-3.5 in its compute class, after a 33B Grok-0 prototype.
Base-model weights and architecture of Grok-1 released under Apache 2.0: a 314B-parameter mixture-of-experts with 25% of weights active per token, not fine-tuned for dialogue.
Grok-2 and Grok-2 mini in beta, text and vision, with xAI citing 56.0% GPQA, 76.1% MATH and 88.4% HumanEval; image generation via Black Forest Labs' FLUX.1 on X.
Grok 3 and Grok 3 mini plus a reasoning mode, Grok 3 (Think), and the DeepSearch agent; xAI claimed a 1M-token context and 93.3% on AIME 2025 for Think.
Cost-efficient model uniting reasoning and non-reasoning modes with a 2M-token context; xAI claimed Grok 4-level benchmark results at a 98% lower price.
Grok 4.1 Thinking and a non-reasoning mode, with xAI citing #1 and #2 on LMArena, better emotional intelligence and creative writing, and a lower hallucination rate.
Built for coding, agentic tasks and knowledge work with Cursor; xAI claimed 62.0% on DeepSWE 1.0 at 80 tokens per second, priced $2/$6 per million tokens.
Aimed at long-running agents and interactive and visual work; xAI said it matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index; a fast variant costs double.
Open-weights Mistral 7B and Mixtral first, then a Large/Medium/Small ladder plus Codestral/Devstral (code), Pixtral (vision) and Magistral (reasoning), unified in Mistral 3, Small 4 and Medium 3.5.
Mistral 7BOpen weightsSmall
First release: a 7B model under Apache 2.0 using grouped-query and sliding-window attention, claimed to outperform Llama 2 13B on all benchmarks; shipped with 7B Instruct.
Sparse mixture-of-experts (46.7B total, 12.9B active per token) under Apache 2.0 with 32k context; claimed to beat Llama 2 70B and match or beat GPT-3.5 on most benchmarks.
First closed flagship, API-only via la Plateforme and Azure, 32K context, five European languages; claimed second-ranked API model after GPT-4. Launched with Le Chat.
First code model: 22B, 80+ programming languages, 32k context, fill-in-the-middle; weights under the Mistral AI Non-Production License. Codestral 25.08 update July 30, 2025.
123B dense model with 128k context, 84.0% MMLU claimed, weights on Hugging Face under the Mistral Research License (commercial license required for self-deployment).
First natively multimodal model: a 400M vision encoder plus 12B decoder, 128k context, Apache 2.0, 52.5% MMMU claimed; Pixtral Large followed November 18, 2024.
Enterprise mid-tier model claimed at or above 90% of Claude Sonnet 3.7 at $0.4/$2 per M tokens, deployable on four GPUs; closed weights, hybrid and on-prem options.
First reasoning model, released as Magistral Small (24B, Apache 2.0) and Magistral Medium (enterprise); Medium claimed 73.6% on AIME 2024, 90% with majority voting.
119B-total / 6B-active MoE under Apache 2.0 that merges Magistral reasoning, Pixtral vision and Devstral coding into one model with configurable reasoning effort.
Called the first flagship merged model: dense 128B, 256k context, instruction, reasoning and coding in one set of weights, modified MIT license; 77.6% SWE-Bench Verified claimed.
DeepSeek-numbered flagship line (LLM 67B, V2, V3, V4) with R1 as the reasoning branch folded into the V3.x hybrid-thinking models; DeepSeek Coder was a parallel code line.
DeepSeek CoderCodeOpen weights
Code models from 1B to 33B trained from scratch on 2T tokens (87% code), 16K window for project-level completion and infilling; commercial use permitted under the Model License.
First general model: 7B and 67B base and chat trained on 2T tokens; the lab reports 67B Chat at 73.78 HumanEval and 84.1 GSM8K, commercial use permitted.
236B MoE with 21B active, 128K context, MLA attention and DeepSeekMoE; the lab claims 42.5% lower training cost and 93.3% smaller KV cache than DeepSeek 67B; V2-Lite on 2024-05-16.
Code MoE continued from V2 on 6T more tokens, 16B-Lite and 236B variants, 338 languages, 128K context; the paper claims performance comparable to GPT4-Turbo on code tasks.
First reasoning model, live on chat.deepseek.com with a visible thought process; the lab claims o1-preview-level AIME and MATH results, with open weights and API 'coming soon'.
Reasoning model with MIT-licensed weights; the lab claims performance on par with OpenAI-o1 in math, code and reasoning, plus six distilled models, 32B and 70B on par with o1-mini.
Hybrid inference: one model with Think and Non-Think modes, 128K context on both API endpoints, post-training for tool use and agents; Base and instruct weights on Hugging Face.
Successor to V3.2-Exp (2025-09-29, DeepSeek Sparse Attention debut); first DeepSeek model with thinking inside tool use; API-only V3.2-Speciale claims IMO and IOI 2025 gold level.
Preview of V4-Pro (1.6T total / 49B active) and V4-Flash (284B / 13B), 1M context now default across DeepSeek services, open weights; claims open-source SOTA in agentic coding.
Experimental multimodal model keeping V4-Flash text ability, image input via base64, URL or the new Files API at up to 384 tokens per image; claims a big leap on multimodal agents.
Qwen (Tongyi Qianwen) is the whole line, numbered 1 to 1.5, 2, 2.5, 3, 3.5, 3.8 and mostly open-weights, with closed -Max/-Plus API tiers and QwQ as the first reasoning branch.
Qwen-7BFlagshipOpen weights
First open Qwen release, 7B base and chat on ModelScope and Hugging Face; Qwen-14B followed 2023-09-25, and Qwen-72B plus Qwen-1.8B (3T tokens, 32K context) on 2023-11-30.
Vision-language model on Qwen-7B taking image, text and bounding boxes as input; base and chat weights released with a paper, commercial use allowed; Qwen-VL-Max API 2024-01-18.
Base and chat models from 0.5B to 72B (32B, 110B and an MoE added later), all with 32K context; code merged into Hugging Face transformers 4.37 so no custom code is needed.
Five sizes 0.5B to 72B including a 57B-A14B MoE, 128K context on instruct models, 27 more languages; all but 72B under Apache 2.0, and the lab claims 72B beats Qwen1.5-110B.
Seven sizes 0.5B to 72B trained on up to 18T tokens, 128K context, shipped with Qwen2.5-Coder and Qwen2.5-Math; Apache 2.0 except 3B and 72B; API-only Qwen-Plus and Qwen-Turbo.
Experimental 32B reasoning model that reflects and self-questions; the lab reports 65.2% GPQA, 50.0% AIME and 90.6% MATH-500, weights on Hugging Face. QwQ-32B followed 2025-03-06.
Large-scale MoE pretrained on over 20T tokens with SFT and RLHF, API-only via Alibaba Cloud (qwen-max-2025-01-25) and Qwen Chat; the lab claims it beats DeepSeek V3 on Arena-Hard.
Two MoE (235B-A22B, 30B-A3B) and six dense models 0.6B to 32B with switchable thinking and non-thinking modes, 119 languages, up to 128K context, Apache 2.0; claims R1/o1 parity.
480B MoE with 35B active for agentic coding, 256K context natively and 1M with extrapolation, released with the open-source Qwen Code CLI; claims open-model SOTA on agentic coding.
Qwen3-Next-80B-A3B, an ultra-sparse MoE with hybrid Gated DeltaNet and gated attention, 262K native context extensible to 1M; said to match Qwen3-235B-A22B-Instruct-2507.
Qwen3-VL-235B-A22B Instruct and Thinking launched the series; 30B-A3B, 4B, 8B, 2B and 32B variants followed through 2025-10-21, with a technical paper on 2025-11-27.
Natively multimodal 397B-A17B MoE with early-fusion vision-language training, 262K context, thinking on by default, 201 languages, Apache 2.0; Qwen3.6 sizes followed in April 2026.
Qwen3.8-2.4T-A95B (2.4T total / 95B active, 262K context, text-only, thinking required, reasoning_effort levels), called Qwen3.8-Max in the blog; 27B multimodal on 2026-08-14.
Kimi names both the product and the models: Kimi Chat, then k1.5 for reasoning, the open K2 line (K2, K2 Thinking, K2.5, K2.6, K2.7-Code) and K3 in 2026.
Kimi k1.5ReasoningResearchMultimodal
Multimodal reasoning model trained with long-context RL; the paper claims 77.5 on AIME and 96.2 on MATH-500, and 'long2short' methods beating GPT-4o and Claude 3.5 Sonnet.
16B MoE vision-language model with 2.8B active, MoonViT native-resolution encoder, 128K context and a Thinking variant under MIT; the lab reports 61.7 MMMU for the thinking model.
Kimi-Dev-72B, an open coding model built on Qwen2.5-72B and optimised with large-scale RL to patch repositories; the lab claims 60.4% on SWE-bench Verified, MIT licence.
1T-parameter MoE with 32B active, optimised for agentic tool use as a non-thinking model, released as K2-Base and K2-Instruct; the 0905 update (2025-09-05) brought 256K context.
Thinking agent on K2 that reasons step by step across 200-300 sequential tool calls, 256K context, native INT4 QAT; claims 44.9% HLE with tools and 71.3% SWE-bench Verified.
Native multimodal agentic model (1T total / 32B active, MoonViT encoder, 256K context) with thinking and instant modes and an agent swarm of up to 100 sub-agents; Modified MIT.
Open-sourced update focused on long-horizon coding, agent swarm scaled to 300 sub-agents and better tool calling; the lab reports 58.6% SWE-Bench Pro and 66.7% Terminal-Bench 2.0.
2.8T-parameter model (104B active) with native vision and 1M-token context, max thinking effort by default; announced 2026-07-16, weights promised by 2026-07-27 (Kimi K3 License).
GLM for every generation: GLM-130B, then the ChatGLM-6B open chat models, then GLM-4 through GLM-4.7 and GLM-5.x; GLM-Z1 was the first reasoning line, GLM-4.5V/4.6V the vision variants.
GLM-130BResearchOpen weights
130B-parameter bilingual (English/Chinese) pre-trained model released with weights on application, presented as runnable on a single A100 server; paper followed 2022-10-05.
First ChatGLM: the aligned ChatGLM-130B went live at chatglm.cn and its 6.2B open-weights version ChatGLM-6B, runnable in 6GB GPU memory, shipped the same day.
Second-generation 6B chat model: context 2K to 32K, MMLU +23% and GSM8K +571% over ChatGLM-6B per the lab, 42% faster inference; a 32K variant followed 2023-07-31.
First closed GLM-4 (0116) via API, claimed to match GPT-4 Turbo (128K) on long context; GLM-4 (0520), GLM-4-Air and the open GLM-4-9B (128K/1M) followed by 2024-06-05.
First GLM reasoning model, built on GLM-4-32B-0414 with cold start and extended RL on math, code and logic; shipped MIT-licensed with the GLM-4-32B-0414 series and a 9B variant.
355B/32B-active MoE with thinking and non-thinking modes, 128K context, MIT licence, ranked 3rd across 12 benchmarks per the lab; shipped with GLM-4.5-Air, GLM-4.5V followed 2025-08-11.
Context expanded from 128K to 200K, higher code-benchmark scores and tool use during inference, claimed gains over GLM-4.5 on eight benchmarks; GLM-4.6V vision followed 2025-12-08.
355B/32B-active MIT model with interleaved and preserved thinking, claiming 73.8% SWE-bench Verified; the 30B/3B-active GLM-4.7-Flash followed 2026-01-19.
Scaled to 744B/40B-active on 28.5T tokens with DeepSeek Sparse Attention, MIT licence, claimed best among open models on reasoning, coding and agentic tasks.
754B MIT model for agentic engineering, 200K context, presented as able to work autonomously on one task for up to 8 hours across thousands of tool calls.
753B model with 1M context, claimed 50% better coding than GLM-5.2 on Z.ai Code Bench and emergent cybersecurity capability; GLM-5.3-Flash with native vision followed 2026-08-26.
ERNIE for everything: research models ERNIE 1.0-3.0, the ERNIE Bot product on ERNIE 3.5/4.0, then ERNIE 4.5 (open-sourced June 2025), X1 for reasoning, and ERNIE 5.0/5.1.
ERNIEResearch
First ERNIE (Enhanced Representation through Knowledge Integration): a BERT-style Chinese model with entity- and phrase-level masking, claimed to beat BERT on Chinese NLP benchmarks.
10B-parameter model fusing auto-regressive and auto-encoding networks trained on 4TB of text plus a knowledge graph, claimed to surpass human performance on SuperGLUE.
Baidu's first chatbot product, built on ERNIE and PLATO, presented with literary creation, business writing, maths, Chinese understanding and multi-modal generation.
Presented at Baidu World as the most advanced ERNIE foundation model, with upgraded understanding, generation, reasoning and memory to power AI-native applications.
Native multimodal foundation model, free to individuals in ERNIE Bot and priced from RMB 0.004 per 1K input tokens via API; ERNIE 4.5 Turbo followed 2025-04-25.
Baidu's first deep-thinking reasoning model with tool use, launched alongside ERNIE 4.5 at half its price; X1 Turbo followed 2025-04-25 and X1.1 on 2025-09-09.
Natively omni-modal 2.4T-parameter model jointly handling text, image, audio and video, in public preview via ERNIE Bot and Qianfan; the full-release blog followed 2026-02-06.
Compresses total parameters to about a third of ERNIE 5.0 while claiming leading performance at lower cost; preceded by ERNIE-5.1-Preview on LMArena 2026-04-30.
Command for the enterprise line: Command, then Command R and R+ (2024), Command A (2025) with Vision, Reasoning and Translate variants, and Command A+ (2026); Aya is the separate multilingual research line.
Command RFlagshipOpen weights
35B model built for production-scale retrieval-augmented generation with 128K context, released with CC-BY-NC research weights on Hugging Face; refreshed as command-r-08-2024.
104B RAG-optimised model for enterprise workloads with 128K context and multi-step tool use, first available on Microsoft Azure, CC-BY-NC weights on Hugging Face; refreshed 08-2024.
111B model with 256K context claimed to match or beat GPT-4o and DeepSeek-V3 on enterprise agentic tasks at far lower compute; CC-BY-NC weights on Hugging Face.
Cohere's first image-capable model, 112B, for enterprise chart, graph and diagram analysis at a low compute footprint; CC-BY-NC weights on Hugging Face.
Cohere's first reasoning model, 111B with 256K context and a controllable thinking budget for agents; open-weights research release under CC-BY-NC; Command A Translate shipped the same month.
Cohere's first Mixture-of-Experts model, 218B total/25B active, 128K context, vision input, reasoning and 48 languages in one model, released under Apache 2.0.
MiniMax-Text-01 (2025) began the open line, then M1 for reasoning (2025), the agent-focused M2 line (M2, M2.1, M2.5, M2.7; 2025-26) and the multimodal M3 in 2026.
MiniMax-Text-01FlagshipOpen weights
456B-parameter mixture-of-experts language model (45.9B active) using Lightning Attention with a 4 million token context, released with MiniMax-VL-01.
Open-weight hybrid-attention reasoning model, 456B parameters (45.9B active), 1 million token context and up to 80,000 tokens of reasoning output; 40k and 80k variants.
M2 line update (with M2.5-highspeed) aimed at programming, tool calling, search and office work; only the month was stated. M2.1 (Dec. 22, 2025) preceded it.
Natively multimodal model, about 428B parameters (23B active) with MiniMax Sparse Attention and up to 1M token context; API on release, weights on Hugging Face later.
Step for every generation: Step-1 and Step-2 first, then the open-weights Step 3 (2025), Step 3.5 Flash and 3.7 Flash (2026), and Step 5 Preview (2026), API-first with weights announced for October.
Step 3FlagshipReasoningMultimodalOpen weights
321B-parameter (38B active) mixture-of-experts vision-language model with 65,536-token context; Apache-2.0 weights; only the month is stated here (paper submitted 2025-07-25).
About 198B-parameter sparse mixture-of-experts vision-language model with 256K context, Apache-2.0 weights and three selectable reasoning levels; only the month is confirmed.
600B-parameter, 27B-active sparse mixture-of-experts model with 1M-token context and text, image and video input, API-only at launch with open weights announced for October 15, 2026.
EXAONE 3.0 (2024, first open release), 3.5 (2024), Deep (2025, reasoning), 4.0 (2025, hybrid reasoning), 4.5 (2026, vision-language) and the sovereign-project K-EXAONE line, K-EXAONE 2.0 in 2026.
EXAONE 3.0FlagshipOpen weights
First open EXAONE: a 7.8B instruction-tuned bilingual (English/Korean) model released on Hugging Face for research use.
OpenAI