AI Models & Open-Weight LLMs
Short answer
This category covers open-weight language models you can download and run yourself, not chat apps. In October 2026 the main families are Qwen3.8 (Alibaba), Kimi K3 (Moonshot AI), DeepSeek, Gemma (Google DeepMind) and gpt-oss (OpenAI). The biggest practical difference between them is the licence: gpt-oss is Apache 2.0 and DeepSeek's repositories are MIT, while Kimi K3 and Gemma weights carry their own terms. Llama 2 is listed as a legacy model. All facts below come from the vendors' own repositories, checked 5 October 2026.
What this category covers: Models, not chat apps. For assistants see Chat.
AI Models & Open-Weight LLMs compared
| Tool | What it does | Pricing model | Status | Last checked |
|---|---|---|---|---|
| DeepSeek | Chinese AI lab DeepSeek's open-weight language models, released under the MIT licence, plus a chat app and API. | Open weights | Active | 5 October 2026 |
| Qwen | Alibaba's Qwen family of language models; Qwen3.8 was released as open weights in August 2026, including a 2.4T-parameter flagship. | Open weights | Active | 5 October 2026 |
| Kimi | Moonshot AI's Kimi models; Kimi K3 is a 2.8T-parameter open-weight model with native vision and a 1M-token context window. | Open weights | Active | 5 October 2026 |
| gpt-oss | OpenAI's open-weight reasoning models, gpt-oss-120b and gpt-oss-20b, released under the Apache 2.0 licence. | Open weights | Active | 5 October 2026 |
| Gemma | Google DeepMind's family of open-weight language models; Gemma 3n is the mobile-first member and Gemma 4 is the newest generation. | Open weights | Superseded | 5 October 2026 |
| Llama 2 | Meta's 2023 open-weight language model family. It is a legacy release, superseded by newer Llama and other open-weight models. | Open weights | Legacy | 5 October 2026 |
All tools in this category
DeepSeek
Chinese AI lab DeepSeek's open-weight language models, released under the MIT licence, plus a chat app and API.
Qwen
Alibaba's Qwen family of language models; Qwen3.8 was released as open weights in August 2026, including a 2.4T-parameter flagship.
Kimi
Moonshot AI's Kimi models; Kimi K3 is a 2.8T-parameter open-weight model with native vision and a 1M-token context window.
gpt-oss
OpenAI's open-weight reasoning models, gpt-oss-120b and gpt-oss-20b, released under the Apache 2.0 licence.
Gemma
Google DeepMind's family of open-weight language models; Gemma 3n is the mobile-first member and Gemma 4 is the newest generation.
Llama 2
Meta's 2023 open-weight language model family. It is a legacy release, superseded by newer Llama and other open-weight models.
How to choose
Evidence: Editorial assessmentRead the licence first
Open weights are not all open source. Apache 2.0 (gpt-oss) and MIT (DeepSeek's V3 repositories) allow broad commercial use. Kimi K3 uses its own Kimi K3 License. Gemma and Qwen publish terms with each model. Meta's Llama licences add an acceptable use policy. If you are building a product, have someone read the exact licence file shipped with the weights, not a summary.
Match the size to your hardware
Model size drives cost. Qwen3.8's flagship has 2.4 trillion parameters and Kimi K3 has 2.8 trillion: both are mixture-of-experts models that activate a fraction of their parameters per token, but they still need data-centre hardware or a hosting provider. Qwen3.8-27B, Gemma and gpt-oss-20b are realistic on much smaller machines.
Check context length and modality
Kimi K3 lists a 1,048,576-token context window and native vision. Qwen3.8's deployment examples use 262,144 tokens. If your use case is long documents or images, check these limits on the model card and test with your own inputs; advertised context and useful context are not the same.
Decide between self-hosting and an API
Self-hosting gives control over data and cost at scale but needs engineering time. Several vendors also offer hosted APIs (Moonshot AI offers an OpenAI- and Anthropic-compatible API for Kimi K3). Hosted APIs in another jurisdiction may conflict with data-residency rules; self-hosting the same weights avoids that.
What changed in this update
- Expanded from one tool to six: added Qwen, Kimi, DeepSeek and gpt-oss; moved Llama 2 here as legacy; Gemma page now covers Gemma 4.
- Renamed the category display name to 'AI Models & Open-Weight LLMs' (URL unchanged).
Common questions
What is the most permissively licensed open-weight model?
Of the models listed here, OpenAI's gpt-oss uses Apache 2.0 and DeepSeek's V3 repositories use MIT, both permissive licences. Always check the licence file shipped with the specific weights you download.
What is the largest open-weight model in 2026?
Kimi K3 lists 2.8 trillion total parameters (104 billion active) and Qwen3.8's flagship 2.4 trillion (95 billion active), according to their repositories.
Is Llama 2 still a good choice?
Not for new projects. Meta has consolidated its repositories around newer Llama releases, and newer open-weight families are more capable.
Where is Gemma 3n?
Our Gemma page now covers the whole family, including Gemma 3n and the newer Gemma 4.