# AI Model Release Timeline (2022-2026)

> Source: https://aiwiki.ai/wiki/ai_model_release_timeline
> Updated: 2026-07-07
> Categories: AI Models, Generative AI, Large Language Models
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution to "AI Wiki (aiwiki.ai)".

The pace of frontier AI model releases went from a single landmark launch in late 2022 to a new flagship roughly every one to two weeks by 2026. [ChatGPT](/wiki/chatgpt), released by [OpenAI](/wiki/openai) on November 30, 2022, opened the generative-AI era; sixteen months later every major lab was shipping GPT-4-class systems, and by mid-2026 they ship near-continuous point releases. OpenAI moved from [GPT-5](/wiki/gpt-5) to GPT-5.6 in under a year, and [Anthropic](/wiki/anthropic) went from Claude 3.7 to Claude Fable 5 in the same span. As of July 2026 the newest generally available flagships are OpenAI's [GPT-5.5](/wiki/gpt-5.5) (with GPT-5.6 in limited preview), Anthropic's Claude Fable 5 and Claude Sonnet 5, [Google DeepMind](/wiki/google_deepmind)'s [Gemini 3.5 Flash](/wiki/gemini_3_5_flash) and [Gemini 3.1 Pro](/wiki/gemini_3_1_pro), [xAI](/wiki/xai)'s [Grok 4.3](/wiki/grok_4_3), and, among open-weight models, Z.ai's GLM-5.2 and [DeepSeek V4](/wiki/deepseek_v4).

This page is a chronological index of major AI model launches from ChatGPT to the present, one row per release, each linking to its full article. For benchmark records over time (who held the top score), see the separate [LLM Benchmarks Timeline](/wiki/llm_benchmarks_timeline); this page tracks releases, not scores.

## How fast did AI model releases accelerate?

The cadence has compressed every year. This timeline lists one landmark launch in 2022, about a dozen in 2023, roughly fourteen in 2024, close to thirty in 2025, and more than twenty in just the first seven months of 2026. Two shifts drove the acceleration. First, the reasoning turn that began with OpenAI's [o1](/wiki/o1) in September 2024 added a second axis of progress, scaling test-time compute, that every lab then raced on. Second, the industry moved from big-bang annual versions toward rapid point releases: in a single year OpenAI shipped GPT-5.1 through GPT-5.6, and Anthropic shipped Claude Opus 4.1 through Opus 4.8. Open-weight Chinese labs (DeepSeek, [Alibaba](/wiki/alibaba)'s Qwen, [Moonshot AI](/wiki/moonshot_ai), and [Z.ai](/wiki/z_ai)) compressed the gap to the closed frontier from years to weeks over the same period.

## 2022: ChatGPT starts the race

The generative-AI era, for most people, begins with one launch. OpenAI shipped a research preview of [ChatGPT](/wiki/chatgpt) built on the GPT-3.5 line, and it reached an estimated 100 million users within two months, the fastest consumer-software adoption on record at the time.

| Date | Model | Developer | Why it mattered |
| --- | --- | --- | --- |
| 2022-11-30 | [ChatGPT](/wiki/chatgpt) (GPT-3.5) | [OpenAI](/wiki/openai) | Free conversational chatbot that triggered the generative-AI boom; about 100M users in two months [1] |

## 2023: the foundation-model land grab

Every major lab rushed a flagship to market. [GPT-4](/wiki/gpt_4) set the capability bar in March, [Meta](/wiki/meta_ai) opened the floodgates on open weights with [Llama 2](/wiki/llama_2) in July, and Google answered with [Gemini](/wiki/gemini_1_0) in December.

| Date | Model | Developer | Why it mattered |
| --- | --- | --- | --- |
| 2023-02-24 | [LLaMA](/wiki/llama) (Llama 1) | [Meta](/wiki/meta_ai) | Research-only weights that leaked and seeded the open-model ecosystem |
| 2023-03-14 | [GPT-4](/wiki/gpt_4) | [OpenAI](/wiki/openai) | Large multimodal jump; about 90th-percentile bar-exam performance [2] |
| 2023-03-14 | [Claude](/wiki/claude_1) (Claude 1) | [Anthropic](/wiki/anthropic) | Anthropic's first assistant, built with Constitutional AI |
| 2023-03-21 | [Bard](/wiki/bard) | [Google](/wiki/google) | Google's first ChatGPT rival, initially on [LaMDA](/wiki/lamda) |
| 2023-05-10 | [PaLM 2](/wiki/palm_2) | Google | Powered Bard and Google's products before Gemini |
| 2023-07-11 | [Claude 2](/wiki/claude_2) | Anthropic | Public access and a then-large 100K-token context window |
| 2023-07-18 | [Llama 2](/wiki/llama_2) | Meta | Commercially licensed open weights (7B to 70B) that catalyzed the open ecosystem |
| 2023-09-27 | [Mistral 7B](/wiki/mistral_7b) | [Mistral AI](/wiki/mistral_ai) | Best-in-class small model under Apache-2.0 |
| 2023-11-06 | [GPT-4 Turbo](/wiki/gpt_4_turbo) | OpenAI | 128K context and lower prices, announced at the first DevDay |
| 2023-11-21 | [Claude 2.1](/wiki/claude_2_1) | Anthropic | Doubled context to 200K tokens |
| 2023-12-06 | [Gemini 1.0](/wiki/gemini_1_0) | [Google DeepMind](/wiki/google_deepmind) | Google's natively multimodal flagship (Ultra, Pro, Nano) |
| 2023-12-11 | [Mixtral 8x7B](/wiki/mixtral) | Mistral AI | Open sparse [mixture-of-experts](/wiki/mixture_of_experts) matching GPT-3.5 quality |

Also in 2023: [xAI](/wiki/xai) announced [Grok](/wiki/grok)-1 (November 4), and Alibaba open-sourced the first [Qwen](/wiki/qwen) models, starting what became the most prolific open-weight family.

## 2024: multimodal, agents, and the reasoning turn

2024 delivered real-time multimodality with [GPT-4o](/wiki/gpt_4o), the open community's first GPT-4-class model in [Llama 3.1 405B](/wiki/llama_3_1), and, in September, the pivot that defined the next two years: [o1](/wiki/o1), the first model trained to reason with reinforcement learning before answering. The year closed with [DeepSeek-V3](/wiki/deepseek_v3) proving a frontier model could be trained for under 6 million USD.

| Date | Model | Developer | Why it mattered |
| --- | --- | --- | --- |
| 2024-02-15 | [Sora](/wiki/sora) (preview) | [OpenAI](/wiki/openai) | Text-to-video demo that reset expectations for generative video |
| 2024-02-21 | [Gemma](/wiki/gemma) | [Google DeepMind](/wiki/google_deepmind) | Google's first open-weight small models |
| 2024-03-04 | [Claude 3](/wiki/claude_3) (Opus, Sonnet, Haiku) | [Anthropic](/wiki/anthropic) | First family to seriously challenge GPT-4's lead; Opus took the top spot |
| 2024-04-18 | [Llama 3](/wiki/llama_3) (8B, 70B) | [Meta](/wiki/meta_ai) | Strong, widely adopted open weights |
| 2024-05-13 | [GPT-4o](/wiki/gpt_4o) | OpenAI | "Omni" model with real-time voice and vision, free to consumers [3] |
| 2024-06-20 | [Claude 3.5 Sonnet](/wiki/claude_3_5_sonnet) | Anthropic | Beat GPT-4o on many tasks; introduced Artifacts |
| 2024-07-23 | [Llama 3.1 405B](/wiki/llama_3_1) | Meta | First open model at GPT-4 class |
| 2024-09-12 | [o1-preview](/wiki/o1) and [o1-mini](/wiki/o1_mini) | OpenAI | First reasoning model; scaled test-time compute via RL chain-of-thought [4] |
| 2024-09-19 | [Qwen2.5](/wiki/qwen2_5) | [Alibaba](/wiki/alibaba) | Broad, strong open family that became the open community's default base |
| 2024-10-22 | [Claude 3.5 Sonnet](/wiki/claude_3_5_sonnet) (upgraded) | Anthropic | Added computer use, an early agentic capability |
| 2024-12-05 | [o1](/wiki/o1) (full) and o1 pro | OpenAI | General-availability reasoning flagship [4] |
| 2024-12-06 | [Llama 3.3 70B](/wiki/llama_3_3) | Meta | 70B model matching the earlier 405B on many tasks |
| 2024-12-11 | [Gemini 2.0 Flash](/wiki/gemini_2_0_flash) | Google DeepMind | Opened Google's "agentic era" with native tool use |
| 2024-12-26 | [DeepSeek-V3](/wiki/deepseek_v3) | [DeepSeek](/wiki/deepseek) | 671B open MoE trained for about 5.6 million USD; frontier quality at low cost [5] |

## 2025: reasoning goes mainstream, GPT-5 and Claude 4 arrive

The pivotal year. [DeepSeek-R1](/wiki/deepseek_r1) proved open reasoning could rival o1 and triggered a market selloff; the [Claude 4](/wiki/claude_4) family made agentic coding the headline capability; and OpenAI shipped both its first open weights since GPT-2 ([gpt-oss](/wiki/gpt_oss)) and [GPT-5](/wiki/gpt-5). Google reset the frontier again with [Gemini 3](/wiki/gemini_3) in November.

| Date | Model | Developer | Why it mattered |
| --- | --- | --- | --- |
| 2025-01-20 | [DeepSeek-R1](/wiki/deepseek_r1) | [DeepSeek](/wiki/deepseek) | First open-weight (MIT) reasoning model rivaling o1; triggered a global AI selloff [6] |
| 2025-01-31 | [o3-mini](/wiki/o3) | [OpenAI](/wiki/openai) | Brought reasoning to the free tier |
| 2025-02-17 | [Grok 3](/wiki/grok_3) | [xAI](/wiki/xai) | Trained on the 200,000-GPU Colossus cluster; xAI's first frontier model |
| 2025-02-24 | [Claude 3.7 Sonnet](/wiki/claude_3_7_sonnet) | [Anthropic](/wiki/anthropic) | First hybrid reasoning model; launched [Claude Code](/wiki/claude_code) [7] |
| 2025-02-27 | [GPT-4.5](/wiki/gpt-4.5) (Orion) | OpenAI | Largest non-reasoning GPT; deprecated less than five months later |
| 2025-03-12 | [Gemma 3](/wiki/gemma_3) | [Google DeepMind](/wiki/google_deepmind) | Strong open multimodal models (1B to 27B) |
| 2025-03-25 | [Gemini 2.5 Pro](/wiki/gemini_2_5_pro) | Google DeepMind | Thinking flagship that topped the LMArena leaderboard [8] |
| 2025-04-05 | [Llama 4 Scout and Maverick](/wiki/llama_4_scout_maverick) | [Meta](/wiki/meta_ai) | First MoE, natively multimodal Llama; Scout advertised a 10M-token context [9] |
| 2025-04-14 | [GPT-4.1](/wiki/gpt-4.1) | OpenAI | 1M-token context, coding-focused, API-only |
| 2025-04-16 | [o3](/wiki/o3) and [o4-mini](/wiki/o4_mini) | OpenAI | Full o3; first reasoning models to use tools and "think with images" |
| 2025-04-29 | [Qwen3](/wiki/qwen_3) | [Alibaba](/wiki/alibaba) | Hybrid thinking/non-thinking open family; landmark open generation |
| 2025-05-07 | [Mistral Medium 3](/wiki/mistral_medium_3) | [Mistral AI](/wiki/mistral_ai) | Frontier-class performance at roughly 8x lower cost |
| 2025-05-22 | [Claude Opus 4](/wiki/claude_opus_4) and [Sonnet 4](/wiki/claude_sonnet_4) | Anthropic | Claude 4 generation; "world's best coding model," first ASL-3 deployment [10] |
| 2025-06-16 | [MiniMax-M1](/wiki/minimax_m1) | [MiniMax](/wiki/minimax) | Open hybrid-attention reasoning model, 456B MoE, 1M context |
| 2025-07-09 | [Grok 4](/wiki/grok_4) | xAI | xAI's most capable model; livestream launch drew about 1.5M viewers [11] |
| 2025-07-11 | [Kimi K2](/wiki/kimi_k2) | [Moonshot AI](/wiki/moonshot_ai) | First trillion-parameter open-weight agentic model |
| 2025-07-28 | [GLM-4.5](/wiki/glm_4_5) | [Z.ai](/wiki/z_ai) (Zhipu AI) | Landmark open agentic MoE (355B), MIT license |
| 2025-08-05 | [gpt-oss](/wiki/gpt_oss) | OpenAI | OpenAI's first open-weight models since GPT-2 in 2019; Apache-2.0 [12] |
| 2025-08-07 | [GPT-5](/wiki/gpt-5) | OpenAI | Unified model plus router; OpenAI's headline flagship [13] |
| 2025-08-19 | [DeepSeek-V3.1](/wiki/deepseek_v3_1) | DeepSeek | First single model combining thinking and non-thinking modes |
| 2025-09-24 | [Qwen3-Max](/wiki/qwen3_max) | Alibaba | Qwen's first model past 1 trillion parameters |
| 2025-09-29 | [Claude Sonnet 4.5](/wiki/claude_sonnet_4_5) | Anthropic | Billed as the best coding model in the world at launch |
| 2025-09-30 | [Sora 2](/wiki/sora_2) | OpenAI | "GPT-3.5 moment for video," with synchronized audio |
| 2025-10-15 | [Claude Haiku 4.5](/wiki/claude_haiku_4_5) | Anthropic | Near-frontier quality at low cost and high speed |
| 2025-11-12 | [GPT-5.1](/wiki/gpt-5.1) | OpenAI | First point release of the GPT-5 line |
| 2025-11-18 | [Gemini 3](/wiki/gemini_3) / [3 Pro](/wiki/gemini_3_pro) | Google DeepMind | New flagship generation; first model past 1500 on LMArena [14] |
| 2025-11-24 | [Claude Opus 4.5](/wiki/claude_opus_4_5) | Anthropic | Anthropic flagship that pushed SWE-bench Verified past 80% at launch [15] |
| 2025-12-02 | [Mistral Large 3](/wiki/mistral_large_3) | Mistral AI | Flagship returns as an open-weight (Apache-2.0) 675B MoE |
| 2025-12-11 | [GPT-5.2](/wiki/gpt-5.2) | OpenAI | Three-tier update (Instant, Thinking, Pro) to the GPT-5 line |

Also shipped in 2025: [QwQ-32B](/wiki/qwq) (Alibaba's open reasoning model), [Cohere](/wiki/cohere) [Command A](/wiki/command_a), [Magistral](/wiki/magistral) (Mistral's first reasoning models), [DeepSeek-V3.2](/wiki/deepseek_v3_2)-Exp with sparse attention, [GLM-4.6](/wiki/glm_4_6), [MiniMax-M2](/wiki/minimax_m2), Kimi K2 Thinking, [Grok 4.1](/wiki/grok_4_1), [Gemini 2.5 Deep Think](/wiki/gemini_2_5_deep_think), [ERNIE 4.5](/wiki/ernie_4_5) ([Baidu](/wiki/baidu)), [Hunyuan-A13B](/wiki/hunyuan_a13b) ([Tencent](/wiki/tencent)), and [Gemini 3 Flash](/wiki/gemini_3_flash).

## 2026: the current frontier

In its first seven months, 2026 already rivals a full earlier year in launch volume. Anthropic opened a new "5" tier (Claude Fable 5 and the restricted Mythos 5), OpenAI retrained its base model for GPT-5.5 and previewed GPT-5.6, and open-weight leaders GLM-5.2 and DeepSeek V4 pushed the open frontier to within striking distance of the closed labs. [Meta](/wiki/meta_ai) shifted its flagship off fully-open weights with the closed [Muse Spark](/wiki/muse_spark).

| Date | Model | Developer | Why it mattered |
| --- | --- | --- | --- |
| 2026-01-27 | [Kimi K2.5](/wiki/kimi_k2_5) | [Moonshot AI](/wiki/moonshot_ai) | Extended the trillion-parameter Kimi line to native multimodality |
| 2026-02-05 | [Claude Opus 4.6](/wiki/claude_opus_4_6) | [Anthropic](/wiki/anthropic) | Long-horizon agentic and enterprise flagship |
| 2026-02-11 | [GLM-5](/wiki/glm_5) | [Z.ai](/wiki/z_ai) (Zhipu AI) | First frontier-tier open model from a publicly listed Chinese lab |
| 2026-02-16 | [Qwen3.5](/wiki/qwen3_5) | [Alibaba](/wiki/alibaba) | Natively multimodal open "agent" flagship; 201 languages |
| 2026-02-19 | [Gemini 3.1 Pro](/wiki/gemini_3_1_pro) | [Google DeepMind](/wiki/google_deepmind) | Point-release upgrade; Google's shipping reasoning flagship |
| 2026-03-05 | [GPT-5.4](/wiki/gpt-5.4) | [OpenAI](/wiki/openai) | Mainline update, after the [GPT-5.3](/wiki/gpt-5.3) Codex and Instant models |
| 2026-04-02 | [Gemma 4](/wiki/gemma_4) | Google DeepMind | Gemma's first release under Apache-2.0 |
| 2026-04-07 | [GLM-5.1](/wiki/glm_5_1) | Z.ai (Zhipu AI) | 754B open agentic model, SOTA open on SWE-bench Pro |
| 2026-04-08 | [Muse Spark](/wiki/muse_spark) | Meta (Superintelligence Labs) | Meta's new frontier flagship, but closed and proprietary, not Llama-branded |
| 2026-04-16 | [Claude Opus 4.7](/wiki/claude_opus_4_7) | Anthropic | Most capable generally available Anthropic model at launch |
| 2026-04-20 | [Kimi K2.6](/wiki/kimi_k2_6) | Moonshot AI | Ranked the strongest open-weight model by Artificial Analysis |
| 2026-04-23 | [GPT-5.5](/wiki/gpt-5.5) | OpenAI | Current GA flagship; first fully retrained base since GPT-4.5 [16] |
| 2026-04-24 | [DeepSeek V4](/wiki/deepseek_v4) | [DeepSeek](/wiki/deepseek) | Next-gen open flagship (V4-Pro 1.6T MoE); math and coding leader [20] |
| 2026-04-30 | [Grok 4.3](/wiki/grok_4_3) | [xAI](/wiki/xai) | xAI's current public flagship [21] |
| 2026-04-30 | [Mistral Medium 3.5](/wiki/mistral_medium_3_5) | [Mistral AI](/wiki/mistral_ai) | Unified open-weight chat, reasoning, and code; 77.6% SWE-bench Verified |
| 2026-05-19 | [Gemini 3.5 Flash](/wiki/gemini_3_5_flash) | Google DeepMind | Newest generally available Gemini and the default fast model |
| 2026-05-20 | [Qwen3.7-Max](/wiki/qwen3_7_max) | Alibaba | Alibaba's current top (proprietary) flagship; 1M context |
| 2026-05-28 | [Claude Opus 4.8](/wiki/claude_opus_4_8) | Anthropic | Most capable Claude 4 model; 41 days after Opus 4.7 |
| 2026-06-01 | [MiniMax M3](/wiki/minimax_m3) | [MiniMax](/wiki/minimax) | Frontier coding plus 1M context plus multimodality, open weights |
| 2026-06-09 | Claude Fable 5 and Claude Mythos 5 | Anthropic | Anthropic's most capable public model (Fable 5); briefly export-controlled [17] |
| 2026-06-13 | GLM-5.2 | Z.ai (Zhipu AI) | Current strongest open-weight model; beats GPT-5.5 on some coding tasks at about 1/6 the cost [23] |
| 2026-06-26 | GPT-5.6 (Sol, Terra, Luna) | OpenAI | Limited preview of OpenAI's next tier [19] |
| 2026-06-30 | Claude Sonnet 5 | Anthropic | Near-Opus performance at lower cost; default for Free and Pro plans [18] |

Also shipped in 2026: Kimi K2.7-Code (Moonshot, June), [Qwen3.6](/wiki/qwen3_6) (Alibaba), and the Mistral 3 family (open-weight). Meta's [Llama 4 Behemoth](/wiki/llama_4_behemoth) and a DeepSeek "R2" were previewed or expected but never shipped under those names.

## Which is the newest AI model? (July 2026)

As of July 2026, the current generally available flagships by developer are:

- OpenAI: [GPT-5.5](/wiki/gpt-5.5) (April 23, 2026), with GPT-5.6 in limited preview [16][19].
- Anthropic: Claude Fable 5 (June 9, 2026) is the most capable public model, with Claude Sonnet 5 (June 30, 2026) as the default consumer model and [Claude Opus 4.8](/wiki/claude_opus_4_8) still in the lineup [17][18].
- Google DeepMind: [Gemini 3.5 Flash](/wiki/gemini_3_5_flash) (May 19, 2026) is the newest release, alongside the [Gemini 3.1 Pro](/wiki/gemini_3_1_pro) reasoning flagship [14].
- xAI: [Grok 4.3](/wiki/grok_4_3) (April 30, 2026) [21].
- Meta: the closed [Muse Spark](/wiki/muse_spark) (April 8, 2026); the open-weight flagship remains [Llama 4 Maverick](/wiki/llama_4_scout_maverick) [22].
- Open-weight leaders: Z.ai's GLM-5.2 (June 13, 2026) tops the open-weight rankings, ahead of [DeepSeek V4](/wiki/deepseek_v4), [Kimi K2.6](/wiki/kimi_k2_6), and [MiniMax M3](/wiki/minimax_m3) [23].

The clearest signal of the accelerating cadence: at least six of these current flagships are so new that their releases postdate most reference material. Expect this timeline to keep growing by the month.

## References

1. OpenAI, "Introducing ChatGPT," openai.com/blog/chatgpt, November 30, 2022.
2. OpenAI, "GPT-4," openai.com/index/gpt-4-research, March 14, 2023.
3. OpenAI, "Hello GPT-4o," openai.com/index/hello-gpt-4o, May 13, 2024.
4. OpenAI, "Introducing OpenAI o1-preview," openai.com/index/introducing-openai-o1-preview, September 12, 2024 (full o1 released December 5, 2024).
5. DeepSeek-AI, "DeepSeek-V3 Technical Report," arxiv.org/abs/2412.19437, December 26, 2024. 671B MoE (37B active), about 5.6 million USD training cost.
6. DeepSeek, "DeepSeek-R1 Release," api-docs.deepseek.com/news/news250120, January 20, 2025.
7. Anthropic, "Claude 3.7 Sonnet and Claude Code," anthropic.com/news/claude-3-7-sonnet, February 24, 2025.
8. Google DeepMind, "Gemini 2.5: Our most intelligent AI model," blog.google/technology/google-deepmind, March 25, 2025.
9. Meta AI, "The Llama 4 herd," ai.meta.com/blog/llama-4-multimodal-intelligence, April 5, 2025.
10. Anthropic, "Introducing Claude 4," anthropic.com/news/claude-4, May 22, 2025.
11. xAI, "Grok 4," x.ai/news/grok-4, July 9, 2025.
12. OpenAI, "Introducing gpt-oss," openai.com/index/introducing-gpt-oss, August 5, 2025. First OpenAI open weights since GPT-2 (2019).
13. OpenAI, "Introducing GPT-5," openai.com/index/introducing-gpt-5, August 7, 2025.
14. Google, "A new era of intelligence with Gemini 3," blog.google/products/gemini/gemini-3, November 18, 2025.
15. Anthropic, "Claude Opus 4.5," anthropic.com/news/claude-opus-4-5, November 24, 2025.
16. OpenAI, "Introducing GPT-5.5," openai.com/index/introducing-gpt-5-5, April 23, 2026.
17. Anthropic, "Claude Fable 5 and Claude Mythos 5," anthropic.com/news/claude-fable-5-mythos-5, June 9, 2026; "Redeploying Claude Fable 5," anthropic.com/news/redeploying-fable-5 (available globally from July 1, 2026 after export controls were lifted).
18. Anthropic, "Introducing Claude Sonnet 5," anthropic.com/news/claude-sonnet-5, June 30, 2026.
19. OpenAI, "Previewing GPT-5.6 Sol," openai.com/index (limited preview), June 26, 2026.
20. DeepSeek, "DeepSeek-V4 (Preview)," api-docs.deepseek.com/news/news260424, April 24, 2026.
21. xAI, "Grok 4.3," x.ai/news, released via API April 30, 2026 with general availability the following week.
22. Meta, "Introducing Muse Spark, Meta Superintelligence Labs," about.fb.com/news, April 8, 2026 (closed, proprietary flagship).
23. Z.ai (Zhipu AI), GLM-5.2 model card and Artificial Analysis open-weight Intelligence Index, artificialanalysis.ai/models/open-source (accessed July 2026). GLM-5.2 leads open-weight models.