The best open models, by job
thecontinuitylayer.com/open/best-nowChecked 22 October 2026. US dollars. No affiliate links.
Updated monthly. The overall ranking comes from Artificial Analysis's open-weights index; we translate it into plain answers. Each month's changes go into Did it change?
Overall ranking
| Model | Index score | Runs at home? |
|---|---|---|
| MiMo-V2.6-Pro (Xiaomi) | 46 | No |
| DeepSeek V4.1 Flash | 45 | No |
| GLM-5.3 (Z.ai) | 45 | No |
| Kimi K3 (Moonshot) | 44 | No |
| GLM-5.3-Flash | 42 | No |
| Qwen3.8-27B (Alibaba) | 34 | Yes, on a 32GB machine |
For comparison, the best paid model scores about 58 on the same index. Only the last one runs on a home machine; to see what fits yours, use the model finder.
The scores above are Artificial Analysis's. What the record grades itself is the paperwork behind each model: the license, whether the weights download, the context window, what it takes to run locally, the maker's hosted price, whether your prompts train it, and whether you can fine-tune it and keep the result, read from the makers' own pages and dated, on the Open models shelf.
Our pick for each job
| Job | Pick | Runs at home? | License | Gap vs paid |
|---|---|---|---|---|
| Chat and writing | GLM-5.3 or DeepSeek V4.1 Flash | No: use a plan or router | 🟢 / 🟡 | Good enough |
| Coding help | DeepSeek V4.1 Flash, GLM-5.3, Kimi K3 | No | 🟢 / 🟡 | Close on routine work |
| “Do it for me” agents | GLM-5.3 | No | 🟡 | Clearly behind |
| Hard reasoning | MiMo-V2.6-Pro | No | Not yet checked | Behind |
| Reading images | Qwen3.8-27B | Yes, 24–32GB | 🟢 | Usable |
| Making images | Z-Image-Turbo | Yes, 16–24GB graphics card | 🟢 | Competitive |
| Transcription | NVIDIA Canary, Whisper v3 | Yes, on a laptop | 🟢 | Equal |
| Laptop default | Qwen3.8-27B | 32GB machine | 🟢 | Fine for drafts |
| 8GB machine or phone | Qwen3.5 4B, Gemma 4 small | Yes | 🟢 | Basic help only |
License lights
🟢 Use for anything. 🟡 Conditions apply only to very large companies; they don't affect individuals or small businesses. 🔴 Non-commercial: FLUX.2 [dev] and Qwen-Image-2.1 can't be used for business images without a separate deal. That's the trap to watch.
“Open-weight” means you can download the file. It doesn't mean it's small enough to run at home.
The honest gap
The best open model trails the best paid model by about four months of progress. It's good enough for everyday chat, writing and routine coding, and clearly behind on long “do it for me” agent work. What runs on a normal laptop trails much further.
Checked 22 October 2026. Next check: 22 November 2026.