What this area covers
Open-weights progress tracks the release, capability trajectory, and ecosystem around AI models whose weights are publicly available — meaning anyone can download, run, fine-tune, or redistribute them. The key players are Meta (Llama), Mistral AI, Alibaba's Qwen team, DeepSeek, Moonshot AI, Google DeepMind (Gemma), and, as of 2025, OpenAI itself. Hugging Face serves as the central distribution and tooling hub.
The defining question of this thread is the gap to closed frontier models: how far behind are the best open-weights models, and is that gap closing? The answer, as of mid-2026, is that it has nearly closed on several important dimensions — and the implications are reshaping policy, safety research, and competitive strategy across the industry.
---
How it evolved
Phase 1 — The early open frontier (2022–2023)
The modern open-weights era began in earnest with BLOOM (176B, July 2022), a collaborative multilingual model from Hugging Face and the BigScience workshop, and accelerated sharply with Meta's Llama 2 (July 2023). Llama 2 was the first open-weights model from a Tier 1 lab to be distributed at scale with both base and instruction-tuned variants, and it established Hugging Face as the default distribution channel.
Mistral AI entered in September 2023 with Mistral 7B — a 7.3B dense model under Apache 2.0 that outperformed Llama 2 13B across benchmarks using Grouped-Query Attention and Sliding Window Attention. Three months later, Mixtral 8x7B introduced sparse Mixture-of-Experts (MoE) to the open-weights world: 46.7B total parameters, only 12.9B active per token, matching or exceeding GPT-3.5 at a fraction of the inference cost. MoE would become the dominant architectural pattern for open frontier models going forward.
Phase 2 — Scaling and specialization (2024)
2024 saw a proliferation of model families and a rapid climb in scale. Meta's Llama 3 (April 2024) and Llama 3.1 (July 2024) brought the 405B parameter mark — the first open-weights model credibly positioned as frontier-class at GPT-4 scale, with multilingual support and extended context. Llama 3.2 (September 2024) added vision-capable multimodal models and lightweight 1B/3B variants for edge deployment.
Alibaba's Qwen team emerged as a prolific open-weights contributor. Qwen2 (June 2024) introduced a 57B MoE variant and 128K context. Qwen2.5 (September 2024) expanded to seven sizes from 0.5B to 72B. Qwen2.5-Coder (November 2024) claimed parity with GPT-4o on coding benchmarks at 32B scale. Qwen2.5-VL (January 2025) extended the family to vision-language tasks.
Mistral continued its cadence: Mixtral 8x22B (April 2024, Apache 2.0, 141B total / 39B active, 64K context) and Mistral Large 2 (July 2024, 123B, 128K context) kept the European lab competitive with the largest open models from US and Chinese labs.
Phase 3 — Reasoning, agents, and closing the gap (2025–2026)
The 2025–2026 period is defined by three shifts: the arrival of reasoning models in the open-weights space, the MoE scaling race pushing total parameter counts into the trillions, and the geopolitical and safety complications that come with models this capable being freely distributable.
DeepSeek became the most disruptive force. DeepSeek-R1 (MIT license, including weights and outputs) claimed performance parity with OpenAI o1 on math, code, and reasoning benchmarks, with API pricing at $0.55/M input tokens — a fraction of comparable closed-model costs. Six distilled smaller variants (up to 70B) reportedly matched o1-mini. DeepSeek-V3 (671B MoE, 37B active, 14.8T training tokens) ran at 60 tokens/second with API pricing at $0.27/M input. The V3 line continued through V3.1 (hybrid think/non-think, agent-focused), V3.2-Exp (DeepSeek Sparse Attention, 50%+ API price cut), and V3.2/V3.2-Speciale (chain-of-thought integrated into tool-use, gold-medal math competition claims). DeepSeek V4 (the current flagship) ships as V4-Pro (1.6T total / 49B active) and V4-Flash (284B total / 13B active), both with 1M token context by default via Token-wise compression and DeepSeek Sparse Attention.
Qwen3 (April 2025) brought a 235B MoE flagship claiming competitive benchmarks against DeepSeek-R1, o1/o3-mini, Grok-3, and Gemini 2.5 Pro. Qwen3-Coder-480B-A35B (July 2025) targeted agentic coding with 256K native context and performance described as comparable to Claude Sonnet 4.
Mistral shipped its most ambitious lineup yet: Magistral (June 2025, first open reasoning model, 24B Apache 2.0 variant scoring 70.7% on AIME2024), Devstral 2 (December 2025, 123B, 72.2% SWE-bench Verified), Mistral Large 3 (January 2026, 675B MoE / 41B active, Apache 2.0, #2 on LMArena OSS non-reasoning leaderboard), Mistral Small 4 (March 2026, 119B MoE / 6B active, unified multimodal + reasoning + coding under Apache 2.0), and Mistral Medium 3.5 (April 2026, 128B dense, 77.6% SWE-Bench Verified, vision encoder trained from scratch). Voxtral (July 2025) extended the family to open-weight speech understanding.
OpenAI made its most significant open-weights move in August 2025: gpt-oss-120b and gpt-oss-20b, released under Apache 2.0, with a dedicated safety evaluation methodology (malicious fine-tuning / MFT) published alongside. This marked a historic reversal for the lab that had kept all its frontier models proprietary since GPT-2.
Google DeepMind released Gemma 4 (April 2026), described as their most capable open models to date, purpose-built for reasoning and agentic workflows.
Moonshot AI's Kimi K3 (July 2026) raised the ceiling further: 2.8T total parameters, 50B active, described as the largest open model ever released, with performance comparable to Claude Opus 4.8 at Sonnet 5 pricing.
---
The ecosystem: fine-tuning and serving infrastructure
The model releases are only half the story. The open-weights ecosystem depends on a stack of tooling that has matured in parallel:
- Hugging Face hosts virtually every major open-weights release and provides the transformers library, model cards, and dataset infrastructure that make the ecosystem function.
- llama.cpp and GGML enable quantized local inference on consumer hardware — and in February 2026, both projects joined Hugging Face, consolidating the local-inference stack under a single steward.
- vLLM, SGLang, and llama.cpp are the primary serving frameworks; Mistral's releases consistently ship with day-0 vLLM support and NVIDIA NIM availability.
- NVIDIA co-optimization has become a standard feature of major Mistral releases, with Blackwell/Hopper kernel tuning and NVFP4 format support baked in.
- LoRA and QLoRA (parameter-efficient fine-tuning) are the default customization path, enabling task-specific adaptation of large open models on single-GPU hardware.
---
The complications: distillation, safety, and geopolitics
As open-weights models approach closed-frontier capability, the stakes of their free distribution have risen sharply.
Distillation attacks: Anthropic publicly identified DeepSeek, Moonshot AI, and MiniMax as conducting coordinated large-scale distillation attacks against Claude — generating over 16 million exchanges through approximately 24,000 fraudulent accounts to train on Claude's most differentiated capabilities (agentic reasoning, tool use, coding, chain-of-thought). Anthropic framed this as a national security concern, arguing distilled models strip safety safeguards and undermine US export controls.
Export controls and access restrictions: Andrew Ng's commentary on Claude Fable 5 (a guardrailed Anthropic model) noted that US government export controls forced Anthropic to disable global access — an event Ng argued accelerates global interest in AI sovereignty and open-source alternatives. The episode illustrates how open-weights models serve as a hedge against sudden access restrictions by closed-model providers.
Meta's pivot: Meta released Muse Spark (April 2026) as a closed-weights model from its new Superintelligence Labs — a notable departure from its open Llama strategy, signaling that even the most committed open-weights lab may selectively close models it considers strategically sensitive.
OpenAI's safety methodology: Alongside gpt-oss, OpenAI published a malicious fine-tuning (MFT) evaluation framework to assess worst-case risks of open-weight releases in biology and cybersecurity domains — an acknowledgment that open-weights releases require a different safety posture than API-gated models.
---
Where it's heading
The open-weights frontier is no longer defined by whether open models can match closed ones — on many benchmarks, they can. The defining questions now are:
1. Distillation and safety: Can open-weights releases be structured to preserve safety properties that closed-model providers invest in? The distillation-attack dynamic suggests the answer is currently no. 2. Geopolitical fragmentation: Export controls and access restrictions are accelerating demand for sovereign open-weights alternatives, particularly outside the US. 3. Infrastructure consolidation: The GGML/llama.cpp acquisition by Hugging Face suggests the serving stack is consolidating even as the model landscape fragments. 4. MoE scaling ceiling: With Kimi K3 at 2.8T total parameters and DeepSeek V4-Pro at 1.6T, the question is whether further total-parameter scaling yields returns, or whether active-parameter efficiency (50B active out of 2.8T) becomes the primary optimization axis.
