BasedAI builds its products on open source AI, so we maintain this database for the same reason you are reading it: choosing an open-weight model means comparing licenses, activation sizes, context windows, and serving costs that are scattered across a dozen model cards and pricing pages. Here they are in one place, with a dedicated page per model and side-by-side comparisons for the matchups people actually weigh.
Every major open-weight model (July 2026)
| Model | Vendor | Parameters | Context | License | Released |
|---|---|---|---|---|---|
| DeepSeek V4 Pro | DeepSeek | 1.6T (49B active) | 1M tokens | MIT | 2026-04-24 |
| DeepSeek V4 Flash | DeepSeek | 284B (13B active) | 1M tokens | MIT | 2026-04-24 |
| DeepSeek V3.2 | DeepSeek | 671B (37B active) | 128K tokens | MIT | 2025-12-01 |
| DeepSeek R1 (0528) | DeepSeek | 671B (37B active) | 128K tokens | MIT | 2025-05-28 |
| Kimi K2.5 | Moonshot AI | 1T (32B active) | 256K tokens | Modified MIT | 2026-01-27 |
| Kimi K2 Thinking | Moonshot AI | 1T (32B active) | 256K tokens | Modified MIT | 2025-11-06 |
| Kimi K2.6 | Moonshot AI | 1T (32B active) | — | Modified MIT | 2026-04-15 |
| GLM-5.2 | Zhipu AI (Z.ai) | 753B (~40B active) | 1M tokens | MIT | 2026-06-13 |
| GLM-4.6 | Zhipu AI (Z.ai) | 357B (32B active) | 200K tokens | MIT | 2025-10-30 |
| Qwen3.5 (397B-A17B) | Alibaba (Qwen team) | 397B (17B active) | 262K tokens (native) | Apache 2.0 | 2026-02-16 |
| Qwen3-Coder (480B-A35B) | Alibaba (Qwen team) | 480B (35B active) | 256K tokens (1M with extrapolation) | Apache 2.0 | 2025-07-22 |
| Llama 4 (Scout & Maverick) | Meta | Scout 109B / Maverick 400B (17B (both) active) | Maverick 1M tokens; Scout up to 10M (claimed) | Llama Community | 2025-04-05 |
| MiniMax-M2 | MiniMax | 230B (10B active) | ~200K tokens | MIT | 2025-10-27 |
| Mistral Small 4 | Mistral AI | 119B (6.5B active) | 256K tokens | Apache 2.0 | 2026-03-01 |
| gpt-oss (120B & 20B) | OpenAI | 117B (gpt-oss-120b) / 21B (gpt-oss-20b) (5.1B / 3.6B active) | 128K tokens | Apache 2.0 | 2025-08-05 |
Specs and pricing verified July 2026. Model pages carry fuller detail, sources, and FAQs.
Rumored and unreleased models we track
- DeepSeek R2 — official status, what is actually announced, and what to use in the meantime.
Open-weight vs open-source: what the terms actually mean
The industry says “open source models” and usually means open-weight models: the trained parameters are published — typically on Hugging Face — under a license that lets you download, run, fine-tune, and deploy them commercially. What is usually not published is the training data and the full training pipeline, which is what a strict open-source definition would require.
The distinction matters less than the license terms do. An MIT-licensed open-weight model (DeepSeek V4, GLM-5.2, MiniMax-M2) gives you effectively unrestricted commercial rights. Apache 2.0 (Qwen3.5, Mistral Small 4, gpt-oss) adds an explicit patent grant — often the preference of enterprise counsel. Community licenses (Llama 4) and modified licenses (Kimi K2.5) are free for almost everyone but carry thresholds worth reading before you build a business on them.
Why choose open weights at all? Four structural advantages no proprietary API can offer: you can self-host (data never leaves your infrastructure), fine-tune on proprietary data, fix your costs (serving cost scales with hardware, not vendor pricing), and never lose access — a model on your disks cannot be deprecated, repriced, or geo-blocked.
Where open models stand against proprietary flagships
The capability gap that defined 2023–2024 has largely closed. By mid-2026, open-weight releases sit within single-digit percentage points of proprietary flagships on most coding and reasoning benchmarks — Kimi K2.6 ties GPT-5.5 on headline coding evals, and GLM-5.2 markets GPT-5.5-tier coding at roughly one sixth the cost. What remains proprietary territory: frontier multimodality (audio, video generation), consumer product polish, and the last few points of reliability on the hardest tasks. See the head-to-heads: DeepSeek vs ChatGPT, Kimi K2 vs Claude, GLM-5.2 vs Claude.
Popular comparisons
DeepSeek vs ChatGPT
Specs, pricing, and verdict
DeepSeek vs Claude
Specs, pricing, and verdict
DeepSeek vs Gemini
Specs, pricing, and verdict
Kimi K2.5 vs ChatGPT
Specs, pricing, and verdict
Kimi K2 vs Claude
Specs, pricing, and verdict
Qwen vs ChatGPT
Specs, pricing, and verdict
GLM-5.2 vs Claude
Specs, pricing, and verdict
MiniMax-M2 vs Claude
Specs, pricing, and verdict
Kimi K2 vs DeepSeek
Specs, pricing, and verdict
Frequently asked questions
What is an open-weight model?
An open-weight model is an AI model whose trained parameters (weights) are published for anyone to download, run, fine-tune, and usually use commercially. The weights are open even when the training data and full training pipeline are not — which is why 'open-weight' is the precise term for models like DeepSeek V4, Kimi K2.5, and Llama 4.
What is the difference between open-weight and open-source models?
Strictly, 'open source' requires everything needed to reproduce the system — code, training data, and methodology — under an OSI-approved license. Almost no frontier model meets that bar. 'Open-weight' means the trained weights are downloadable under a license that permits use and modification. In practice the industry uses 'open source AI' loosely for both; this database records each model's exact license so you can judge for yourself.
What is the best open source LLM right now?
As of July 2026 there is no single answer — it depends on the job. DeepSeek V4 Pro leads on general reasoning and long context (1M tokens, MIT). Kimi K2.5 leads multimodal and agentic work. GLM-5.2 is the strongest coding-first model. Qwen3.5 wins on efficiency, licensing, and family breadth. MiniMax-M2 is the best lean tool-use agent. Our comparison pages break down each matchup.
Are open-weight models free for commercial use?
Mostly yes, but licenses differ. MIT (DeepSeek, GLM, MiniMax) and Apache 2.0 (Qwen, Mistral, gpt-oss) are unrestricted. Moonshot's Modified MIT adds an attribution requirement above 100M monthly users or $20M monthly revenue. Meta's Llama Community License requires a separate agreement above 700M monthly active users. Each model page lists the exact terms.
Can open-weight models match ChatGPT and Claude?
On many everyday tasks, yes — mid-2026 evaluations put the gap between top open-weight models and proprietary flagships at single-digit percentage points on coding and reasoning benchmarks, while API costs run 4–10× lower and self-hosting is possible. Proprietary models retain edges in product polish, multimodal breadth, and peak reliability.