deepseek-flash open weights and deployment

Status: Available. Official weights are published on Hugging Face under the MIT License, with FP8 dense weights and FP4 MoE experts. DeepSeek ships a reference implementation rather than a production serving engine and no Jinja chat template, so self-hosting depends on a runtime that documents support for this release.

Agent deployment prompt

You are my deployment agent. Deploy the official deepseek-ai/DeepSeek-V4.1-Flash weights as a private, OpenAI-compatible service on this host. Use the publisher's model card as the source of truth: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash.

Preflight
- Inspect the OS, GPU model and count, free VRAM, driver and CUDA versions, free disk space, and available ports. The official release stores dense weights in FP8 and MoE experts in FP4: confirm that this GPU generation and the chosen runtime can execute both formats, and estimate whether the weights and the requested context fit with safe headroom. DeepSeek publishes no minimum hardware requirement.
- DeepSeek names no production serving engine: the repository's inference code is a reference implementation, and the release ships no Jinja chat template. Check the current release notes of vLLM and SGLang for explicit DeepSeek-V4.1-Flash support, including its chat encoding. If no runtime supports the model and its chat format on this hardware, stop before installing or launching anything and explain what is missing; do not patch kernels, convert the weights or substitute a community quantization.

Deployment
- If preflight passes, create an isolated environment, install the runtime release that documents DeepSeek-V4.1-Flash support, record the exact model revision you download, and serve the model through an OpenAI-compatible API. Generate an API key and store it in a protected environment variable. Bind to 127.0.0.1 first; do not expose the endpoint publicly.
- Start with a context length the preflight shows the hardware can hold. The model supports up to 1,048,576 tokens; raise the limit only if I need it.

Verification and handoff
- Send a real request to /v1/chat/completions and confirm that the model responds. Report the exact launch command, installed versions, model revision, local endpoint, test result, measured GPU memory and log location. Identify any reverse-proxy or firewall changes that still need my approval. Never print the API key or other secrets.
License
MIT
Hardware note
Not published by DeepSeek; the Agent recipe checks GPU support for FP8 and FP4 weights and memory headroom before serving.
Recommended runtime
Not named by DeepSeek; use a vLLM or SGLang release that documents DeepSeek-V4.1-Flash support
Verified date
2026-09-24

Official repository

Published model card

Parameter count
552B backbone parameters (MoE); 8B active per token for input and 16B for output
Context window
1,048,576 tokens (1M)
Maximum output
384K tokens (393,216) on the DeepSeek API; the model card sets no hard maximum
Input/output modalities
Text and image input; text output

Official sources