AI library

Models

Open-weight models

Last reviewed 16 October 2026

Overview

Open-weight models publish their trained weights, so anyone can download, run, fine-tune, and self-host them. "Open-weight" is more accurate than "open-source": training data and code are usually not released, and licences range from fully permissive (Apache 2.0, MIT) to custom licences with usage limits.

Key concepts

  • Main families: Qwen (Alibaba), DeepSeek, Llama (Meta), Mistral, Gemma (Google), and others such as Kimi, GLM, and OpenAI's gpt-oss.
  • Mixture of experts: most large open models are MoE, quoted as total and active parameters (for example "675B total, 41B active").
  • Licences: check commercial-use terms, user limits, and attribution rules per model.
  • Running locally: tools such as Ollama, llama.cpp, and LM Studio run quantised models on laptops and desktops; vLLM and SGLang serve them on GPUs.
  • Hosted inference: many providers serve open models over an API.

How to use it

  • Choose open weights for data control, offline use, fine-tuning, or lowest cost at scale.
  • Expect frontier closed models to remain ahead on the hardest agentic tasks; test on your own workload.
  • Quantised 4-bit versions trade a little quality for large memory savings.

Main families

As of 2026-10-07, checked against official model cards and announcements:

Model Released Parameters (total / active) Context Licence
DeepSeek-V4.1-Flash 2026-09-10 552B / 8B prefill, 16B decode (encoder-decoder MoE) 1M MIT
DeepSeek-V4-Pro 2026-04-24 1.6T / 49B (MoE) 1M MIT
DeepSeek-V4-Flash 2026-04-24 284B / 13B (MoE) 1M MIT
Qwen3.8-Flash-Next 2026-08 125B / 6B (MoE), plus 51B n-gram embeddings 262K native, 1M with RoPE scaling Qwen Community 1.0 (custom)
Qwen3.5-397B-A17B 2026-02-16 397B / 17B (MoE) 262K native, ~1M with YaRN Apache 2.0
Gemma 4 (E2B to 31B) 2026-04-03 2.3B effective to 31B dense; 26B A4B MoE 128K–256K Apache 2.0
Mistral Large 4 (weights due end of Oct 2026) 2026-10-06 (API preview) 1T / 49B (MoE) not stated not stated
Mistral Large 3 2025-12-02 ~675B / 41B (MoE) 256K Apache 2.0
Llama 4 Scout / Maverick 2025-04-05 109B / 17B; 400B / 17B (MoE) Scout 10M Llama 4 licence (custom)

DeepSeek-V4-Pro is the largest open-weight model in the library. The newer, smaller V4.1-Flash beats it on DeepSeek's own benchmarks, and since 2026-09-14 DeepSeek's API routes deepseek-v4-pro requests to V4.1-Flash. DeepSeek's own news pages give the V4 timeline: V4-Pro and V4-Flash preview with open weights on 2026-04-24, V4-Pro general availability on 2026-08-13, V4-Flash-Vision-Exp on 2026-08-21, and V4.1-Flash on 2026-09-10. (A news report gave 2026-07-31 for V4-Flash; DeepSeek's date is used here.) Qwen3.8-Flash-Next previews the Qwen4 architecture. Licences are moving in opposite directions: Gemma 4 moved from Google's custom terms to Apache 2.0, while Qwen moved from Apache 2.0 (Qwen3.5) to a custom community licence for Qwen3.8-Flash-Next. Qwen3.5 uses linear attention in most layers, and Qwen3.8-Flash-Next adds sparse attention to that (see transformers).

Open weights are also spreading to non-language foundation models. Mistral's Timestral (2026-10-15) is a family of four time-series forecasting models (5M to 209M parameters, Modified MIT licence) that run zero-shot on a CPU.

Meta has not released open weights since Llama 4. Its newer Muse models are API-only: Muse Spark 1.1 (2026-07-09) and Muse Wonder / Muse Wonder Full (2026-10-12).

Sources disagree on Gemma 4: it was announced on 2026-04-03, but the model card cites a July 2026 technical report and lists a 12B size that launch coverage did not mention.

Prices for hosted open models can change quickly and may vary by time of day. DeepSeek's own API, as of 2026-10-05, per 1M tokens at peak: V4.1-Flash costs $0.30 for input and $1.20 for output, and V4-Pro costs $1.32 for input and $3.96 for output. Off-peak rates are half of those. See cost and caching.

Current state

  • 2026-10-15: Mistral released Timestral, its first time-series foundation models: four sizes (5M, 28M, 123M, 209M) with weights on Hugging Face under a Modified MIT licence, 32K-step context, and 21-quantile probabilistic forecasts. Mistral says the 209M model leads pre-trained zero-shot models on GIFT-Eval, TIME, and fev-bench, ahead of Chronos-2 and TimesFM-3 (vendor claims).
  • 2026-10-14: Google DeepMind and Singapore's government announced an on-device scam detector built on a fine-tuned lightweight Gemma, an example of small open models adapted for private, on-device use. The fine-tuned model itself is not being released.
  • 2026-10-12: Meta released Muse Wonder (Meta Model API and Muse Code) and Muse Wonder Full (Muse Code subscribers only), its strongest agent models, with cross-session "Active Memory". Again no open weights, licence, or prices; benchmark scores are deferred to a separate report.
  • 2026-10-06: Mistral Large 4 (1T total / 49B active, multimodal MoE) entered API preview as mistral-large-4-0 at $1.36 / $4.18 per 1M tokens. Mistral promises open weights at the end of October 2026, but has not stated the licence or context length. Mistral reports strong cyber and agentic scores, including 93% on Cybench.
  • 2026-10-06: Google released EmbeddingGemma 2 (740M parameters, Apache 2.0), an open multimodal embedding model small enough for phones. It is an embedding model, not a chat model, so it is not in the table; see embeddings.
  • 2026-10-04: Alibaba's Qwen 3.7 Max (May 2026) shipped as API-only without weights (secondary report, not yet confirmed from a primary source).

Open questions

  • Will leading labs keep releasing their largest models as open weights? DeepSeek still does (MIT). Meta's newest line is API-only. Qwen still releases weights, but under a custom licence. Mistral promises Large 4 weights; check at the end of October 2026 whether they ship and under which licence (Large 3 was Apache 2.0).
  • What does the Qwen Community 1.0 licence allow for commercial use? It has not been read yet.
  • Did Llama 4 Behemoth ever ship? No primary source has been found.
  • Add primary sources for Kimi, GLM, and gpt-oss.

Sources