Models & Weights

DeepSeek-V4-Flash-0731

DeepSeek's agent-tuned frontier MoE goes MIT — LICENSE file and all

Open ✓Model

What it is

DeepSeek's 284B-total / 13B-active MoE with 256 routed experts per layer, a CSA + HCA hybrid attention stack, a 1M-token context with 384K-token max output, and native FP8 weights — a 167GB download shipping with the DSpark speculative-decoding module and native Responses API support. The 0731 build is a re-post-training pass on the April V4 preview: architecture unchanged, agentic behavior overhauled.

Why it's interesting

The agentic jump is the story — 54.4 on DeepSWE against the preview's 12.8, 82.7 on Terminal-Bench 2.1, 76.7 on Cybergym — and it lands under plain MIT with an actual LICENSE file in the repo root. DeepSeek-V3 sits on our Fine Print-adjacent shelf for its Model License riders; V4-Flash walks off it.

Use cases

  • Self-hosted long-horizon coding and tool-use agents with 1M-token context
  • Frontier-adjacent inference at commodity API prices or on your own node
  • Post-training and speculative-decoding research on an ungated MIT base

Who it's for

Agent builders and infra teams ready for a 167GB download

Setup

Advanced. ~167GB native FP8; roughly 110GB resident at 3-bit quantization, or a single 4×GB300 node at full precision. Community GGUF quants via unsloth

Limitations & cautions

The download is 167GB and serving wants FP4/FP8-era infrastructure; the nine agentic benchmark wins are self-published with independent evaluation still early; and 'Flash' means efficiency-tuned — the heavier V4-Pro line remains unreleased.

Editorial takeaway

Two batches ago we shelved DeepSeek-V3 for its license riders. V4-Flash answers the only way that counts: an MIT file in the repo root.

Related & alternatives