Source: DeepSeek V4 model cards (DeepSeek, Hugging Face)

Model cards for DeepSeek-V4-Pro-0813 and DeepSeek-V4.1-Flash-1017: MIT-licensed MoE models with 1M context and built-in speculative decoding.

Contents
  • What: Hugging Face model cards
  • Publisher: DeepSeek-AI
  • Date: 2026-08-13 (V4-Pro-0813) and 2026-10-17 (V4.1-Flash-1017); accessed 2026-11-14
  • Original: DeepSeek-V4-Pro-0813

Summary

Release notes for DeepSeek’s V4 generation. V4-Pro-0813 is the official release superseding the preview, with a DSpark speculative-decoding module; V4.1-Flash-1017 is a minor update that improves benchmarks while shortening outputs. Also covers V4.1-Flash-1017.

Key claims

  • Both are MIT-licensed, weights and code.
  • Hugging Face metadata reports ~1.65T parameters for V4-Pro-0813 and ~763B for V4.1-Flash-1017 (stored tensors; active counts not stated on the cards).
  • No Jinja chat template: prompts are built with a Python encoding reference implementation (and a Rust deepseek-recipe library for V4.1).
  • reasoning_effort supports low, high and max on V4-Pro.
  • Serving recipes target a single 4xGB300 node with vLLM or SGLang, FP8 KV cache, expert parallelism and DSpark speculative decoding.
  • Recommended V4.1-Flash settings: temperature 1.0, top_p 1.0, 1M context, max_tokens of at least 256K.
  • The V4.1-Flash technical report is titled around pushing the limits of KV cache compression.
  • Vendor-reported benchmarks place V4-Pro near (but mostly below) the leading proprietary models of mid-2026.

Feeds

Open-weight models

Suggest changes →