Models & Weights

DeepSeek-V4-Pro-0813

The heavyweight V4 finally goes GA — and it keeps Flash's MIT, LICENSE file and all

Open ✓Model

What it is

DeepSeek's flagship MoE reaches general availability as the 0813 build: the V4-Pro Preview architecture re-post-trained for agentic work, with the DSpark speculative-decoding module attached and a companion DSpark repo alongside. The card doesn't restate parameter counts for the GA build — the Preview's ~1.6T-total / 49B-active figures are the best public reference — and the safetensors shards total roughly 893GB, so 'Pro' means exactly what it says.

Why it's interesting

Two weeks ago V4-Flash walked DeepSeek off our fine-print-adjacent shelf; the open question was whether the heavyweight would follow. It does — MIT card tag plus a standalone 1.08 kB LICENSE file in the repo root, the same file Flash shipped. The agentic jump over Preview is the other story: 62.7 on DeepSWE against 12.8, 83.3 on CyberGym, 87.9 on Terminal-Bench 2.1, all self-published.

Use cases

  • Frontier-class self-hosted agents where the API's terms or latency won't do
  • Long-horizon coding and terminal work at the top of the open-weights range
  • Post-training and speculative-decoding research on an ungated MIT frontier base

Who it's for

Infra teams with serious multi-node hardware; everyone else via the API

Setup

Advanced. ~893GB of safetensors — plan on a multi-node FP8/FP4 deployment; unsloth and community quant repos are already carving it down

Limitations & cautions

Nearly a terabyte of weights puts self-hosting out of most orgs' reach, the GA card declines to restate parameter counts or context length, and the benchmark table is DeepSeek's own — independent evals were still landing at check time.

Editorial takeaway

Flash was the promise; Pro is the proof. When a lab MITs its flagship — not just the efficient sibling — the license change is policy, not accident.

Related & alternatives