DeepSeek-V4.1-Flash
Multimodal MoE: 552B backbone params, 8B/16B activated, 1M context, MIT license. No Jinja chat template ships; local runs today mostly via antirez's DwarfStar SSD-streaming engine on Macs.
Recommended quants
- DwarfStar Q2 (340.6 GiB file, 151.8 GiB main weights)
antirez/deepseek-v4.1-flash-ggufIncludes 188.8 GiB FP8 Engram tables read from SSD; use --ssd-streaming on a single 128 GB Mac. Not a standard llama.cpp GGUF. - DwarfStar Q4 (483 GiB file, 294 GiB main weights)
antirez/deepseek-v4.1-flash-ggufResident only on a 512 GB Mac; SSD streaming on smaller. - Official FP8 weights (~763 GB repo)
deepseek-ai/DeepSeek-V4.1-FlashServer GPUs.
Chat template Caveats
No Jinja chat template in the release. DeepSeek ships a Python reference encoder (encoding/encoding.py) and the deepseek-recipe Rust/Python libs.
Fix: Use a runtime that implements the DeepSeek encoding; generic GGUF templates may be wrong.
Sampling settings
| Official | temp 1.0, top_p 0.95 or 1.0 |
|---|
Runtimes
| vLLM | Unverified | |
|---|---|---|
| llama.cpp | Unverified | |
| Ollama | Unverified | Library page deepseek-v4.1-flash exists; may be cloud-only. |
| LM Studio | Unverified | |
| MLX | Unverified |
Tool calling in agent tools
| Claude Code (local endpoint) | Unverified | |
|---|---|---|
| OpenCode | Unverified | |
| Cline | Unverified | |
| Aider | Unverified |
Known issues
Sources checked
- https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash
- https://huggingface.co/antirez/deepseek-v4.1-flash-gguf
Spotted something wrong or a new fix? Corrections are reviewed before publishing. (Submission form coming soon.)