Source notes
Facts stay useful when their source boundary is visible.
These links are evidence notes, not signup or download CTAs. Always verify the exact model card and license before production use.
| Source | Type | Freshness | Used for |
|---|---|---|---|
| Open Weight DefinitionOpen Source Alliance | definition | Version 0.3, 2025-01-21 | Open Weight criteria include free redistribution, real model weights, derived works, non-discrimination, and technology neutrality. |
| Introducing gpt-ossOpenAI | official release | Published 2025-08-05; crawled 2026-08-04 | gpt-oss-120b and gpt-oss-20b are Apache 2.0 open-weight reasoning models; the 120b variant targets 80 GB memory and the 20b variant targets 16 GB memory. |
| OpenAI open-weight models (gpt-oss)OpenAI Help Center | official help | Updated 2026-07-30; crawled 2026-08-04 | gpt-oss runs on infrastructure controlled by the user or hosting providers and is not served through the OpenAI API or ChatGPT. |
| Qwen3: Think Deeper, Act FasterQwen Team | official release | Published 2025-04-29; crawled 2026-08-04 | Qwen3 includes MoE and dense open-weight models under Apache 2.0, with recommended deployment frameworks such as SGLang and vLLM and local tools including Ollama, LM Studio, MLX, llama.cpp, and KTransformers. |
| Models filtered by open-weightHugging Face | model index | Crawled 2026-08-04 | The Hugging Face model hub exposes open-weight as a model filter and shows downloadable model repositories across inference tools and providers. |
| What Are Open Models?NVIDIA | glossary | Crawled 2026-08-04 | Open models expose weights, data, or training recipes that developers may inspect, customize, and deploy on infrastructure they control. |
| llama-models READMEMeta Llama | official repository | Crawled 2026-08-04 | The Llama models table lists Llama 4 Scout and Maverick releases and requires license acceptance before model-weight download. |
| Under which license are Mistral's open models available?Mistral AI Help Center | official help | Updated 2026-08; crawled 2026-08-04 | Most Mistral open models use Apache 2.0; some models use a modified MIT license with a monthly revenue threshold. |
| Introducing Mistral 3Mistral AI | official release | Published 2025-12-02; crawled 2026-08-04 | Mistral 3 includes dense Ministral models and Mistral Large 3 under Apache 2.0, with optimized deployment paths through vLLM, NVIDIA, Red Hat, and managed providers. |
| MoonshotAI/Kimi-K3 READMEMoonshot AI | official repository | Crawled 2026-08-04 | Kimi K3 is described as an open-weight native multimodal agentic model with full weights released under the Kimi K3 License and deployment paths including vLLM, SGLang, and TokenSpeed. |
| DeepSeek-V4-Pro model cardDeepSeek AI | official model card | Crawled 2026-08-04 | DeepSeek-V4-Pro and DeepSeek-V4-Flash are open-weight MoE models with 1M context model-card notes and MIT-licensed repository/model weights. |
| OpenAI's new open weight models are goodHacker News | community discussion | Crawled 2026-08-04 | Community discussion highlights practical concerns around capability, filtering, and architecture interpretation. |