All tasks / Text & knowledge / Writing
Best local AI for writing
Open-weights models and tools that run entirely on your own hardware, ranked from public benchmarks. Local tools are rated by the best open model they run.
ATop ranked: Kimi K3 (Moonshot AI), 1,461 on LMArena Text Arena: Creative Writing (style control).29 options listed · data verified 19 Oct 2026
Ranked list
| Tier | Name | Approach | Runs | Rated by | Price | Skill | Compare |
|---|---|---|---|---|---|---|---|
| A |  Kimi K3Moonshot AI | Models | Local | 1,461 on LMArena Text Arena: Creative Writing (style control) | ~1600 GB VRAM | — | |
| A |  GLM-5.3Z.ai | Models | Local | 1,453 on LMArena Text Arena: Creative Writing (style control) | ~425 GB VRAM | — | vs Kimi K3 |
| A |  DeepSeek-V4-Pro-0813DeepSeek | Models | Local | 1,446 on LMArena Text Arena: Creative Writing (style control) | ~920 GB VRAM | — | |
| A |  MiMo-V2.6-ProXiaomi | Models | Local | 1,445 on LMArena Text Arena: Creative Writing (style control) | ~585 GB VRAM | — | |
| A | DeepSeek-V4.1-Flash-1017 DeepSeek | Models | Local | 1,440 on LMArena Text Arena: Creative Writing (style control) | ~320 GB VRAM | — | |
| A | GLM-5.3-Flash Z.ai | Models | Local | 1,434 on LMArena Text Arena: Creative Writing (style control) | ~185 GB VRAM | — | |
| A |  OllamaOllama | Runtimes & gateways | Local | via Kimi K3 | Free | intermediate | |
| A |  Open WebUIOpen WebUI Inc. | Apps | Local | via Kimi K3 | Free | intermediate | |
| A |  llama.cppggml-org | Runtimes & gateways | Local | via Kimi K3 | Free | advanced | |
| A |  TextGen (text-generation-webui)oobabooga | Apps | Local | via Kimi K3 | Free | intermediate | |
| A |  JanMenlo Research | Apps | Local | via Kimi K3 | Free | beginner | |
| A |  SillyTavernSillyTavern (open-source project) | Apps | Local | via Kimi K3 | Free | advanced | |
| A |  KoboldCppLostRuins (open-source project) | Apps + Runtimes & gateways | Local | via Kimi K3 | Free | intermediate | |
| A |  LM StudioElement Labs | Apps | Local | via Kimi K3 | Free | beginner | |
| A |  Msty StudioMsty | Apps | Local | via Kimi K3 | Free | beginner | |
| B |  Gemma 4 (31B)Google DeepMind | Models | Local | 1,419 on LMArena Text Arena: Creative Writing (style control) | ~19 GB VRAM | — | |
| B |  MiniMax-M3MiniMax | Models | Local | 1,405 on LMArena Text Arena: Creative Writing (style control) | ~245 GB VRAM | — | |
| B |  Step 5 PreviewStepFun | Models | Local | 1,403 on LMArena Text Arena: Creative Writing (style control) | ~346 GB VRAM | — | |
| B |  Mistral Medium 3.5Mistral AI | Models | Local | 1,400 on LMArena Text Arena: Creative Writing (style control) | ~75 GB VRAM | — | |
| B | MiMo-V2.6-Flash Xiaomi | Models | Local | 1,390 on LMArena Text Arena: Creative Writing (style control) | ~178 GB VRAM | — | |
| B |  InklingThinking Machines Lab | Models | Local | 1,381 on LMArena Text Arena: Creative Writing (style control) | ~557 GB VRAM | — | |
| B |  Qwen3.8-27BAlibaba (Qwen) | Models | Local | 1,366 on LMArena Text Arena: Creative Writing (style control) | ~17 GB VRAM | — | |
| B |  Muse Glimmer 30BMeta | Models | Local | 1,364 on LMArena Text Arena: Creative Writing (style control) | ~19 GB VRAM | — | |
| C | Inkling Small Thinking Machines Lab | Models | Local | 1,314 on LMArena Text Arena: Creative Writing (style control) | ~159 GB VRAM | — | |
| C |  Llama 4 MaverickMeta | Models | Local | 1,307 on LMArena Text Arena: Creative Writing (style control) | ~230 GB VRAM | — | |
| C |  gpt-oss-120bOpenAI | Models | Local | 1,279 on LMArena Text Arena: Creative Writing (style control) | ~80 GB VRAM | — | |
| C |  Hemmingway-1The Commons Lab (Altworld) | Models | Local | 1,906 on EQ-Bench Creative Writing v3 | ~17 GB VRAM | — | |
| C | gpt-oss-20b OpenAI | Models | Local | 1,240 on LMArena Text Arena: Creative Writing (style control) | ~16 GB VRAM | — | |
| C | Qwen3.8-2.4T-A95B Alibaba (Qwen) | Models | Local | 1,843 on EQ-Bench Creative Writing v3 | ~1370 GB VRAM | — |
Tiers come only from public benchmarks (LMArena Text Arena: Creative Writing (style control), EQ-Bench Creative Writing v3); tools without a result of their own inherit the tier of the best model they offer. How we rank.