DeepSeek shipped V4-Pro-0813 on August 12 with 1.7T parameters, 893 GB of MIT-licensed weights, DSpark speculative decoding, and benchmark parity with Opus 4.8 across agentic coding. The Pro tier is now broadly competitive with the strongest proprietary models — and it is downloadable.

DeepSeek V4-Pro-0813 Drops As MIT-Licensed Open Weights. The Pro Tier Has Finally Caught Up To Its Own Flash.

Hey guys, Mr. Technology here.

It is Friday, August 14, 2026, and 48 hours ago DeepSeek shipped V4-Pro-0813 — the official general-availability release of the Pro tier, with MIT-licensed open weights on Hugging Face, DSpark speculative decoding built in, and benchmark parity with Opus 4.8 across agentic coding. 1.7T parameters. 893 GB of weights. MIT license. vLLM and SGLang recipes shipped the same day. If you are still arguing open-weights versus API at the highest capability tier, the argument just lost its factual basis.

This is the biggest single open-weights model release of 2026, and here is why it matters.

What Shipped On August 12

deepseek-ai/DeepSeek-V4-Pro-0813 — 1.7T parameter MoE, 893 GB of MIT-licensed weights on Hugging Face. The official replacement for the V4-Pro Preview build from April.
DSpark speculative decoding baked in. A single vLLM flag enables it; draft and target weights come from the same checkpoint, so no separate draft model to manage.
Three-tier reasoning control — low, high, and max are first-class parameters on reasoning_effort.
vLLM 0.20 and SGLang 0.5.13+ support day one. Single-node FP4 on a 4xGB300 node is the documented recipe.
DeepSeek Harness is now in developer preview for agent harness integrators — the same reference runtime that powered the Flash-0731 numbers is the Pro tier's reference runtime now.

The Numbers, Honestly Read

From the Hugging Face model card:

Benchmark                 V4-Pro-0813  V4-Flash-0731  Opus-4.8  GLM-5.2  Kimi-K3
HLE (wo / w tools)        42.7 / 60.0  37.8 / 51.5    49.8 / 57.9  40.5/54.7   43.5/56.0
Terminal Bench 2.1        87.9         82.7           85.0         81.0        88.3
Cybergym                  83.3         76.7           78.3          —          80.0
DeepSWE                   62.7         54.4           58.0         46.2        67.5
Toolathlon-Verified       74.1         70.3           76.2         59.9        76.5
AutomationBench            31.8         25.1           27.2         12.9        30.8
DSBench-FullStack         71.1         68.7           71.6         61.8        73.7
DSBench-Hard              67.2         59.6           71.7         54.5        63.0

Two readouts from that table.

One: Pro has caught up to Flash on agentic coding. V4-Pro-0813 beats V4-Flash-0731 on every benchmark. Terminal Bench 2.1: 87.9 vs 82.7. Cybergym: 83.3 vs 76.7. DeepSWE: 62.7 vs 54.4. NL2Repo: 61.5 vs 54.2. The Pro tier is now strictly better than the Flash tier for the work people pay for. For three weeks that was not true. Now it is.

Two: Pro is broadly competitive with the proprietary frontier. Against Opus 4.8, V4-Pro-0813 wins on Terminal Bench 2.1, Cybergym, DeepSWE, and HLE with tools (60.0 vs 57.9). Toolathlon-Verified is close at 74.1 vs 76.2. Against GLM-5.2 it wins almost everything. Against Kimi-K3 the picture is mixed. Against Fable-5 with fallback it loses or ties, and Fable-5 with fallback is a research artifact, not a product.

Honest read: V4-Pro-0813 is the highest-capability open-weights model of 2026, within roughly two points of the strongest proprietary model on every agentic coding benchmark DeepSeek published. When a $2,500/month Opus subscription and an MIT-licensed 893 GB download end up within two points across nine evals, the procurement conversation changes.

DSpark Is The Real Story

DSpark is now a first-class citizen on the Pro tier. The same speculative decoder that gave V4-Flash-0731 a roughly 2.5x throughput multiplier on long-context agent traffic is built into V4-Pro-0813 by construction.

bash

vllm serve deepseek-ai/DeepSeek-V4-Pro-0813 \
  --trust-remote-code --kv-cache-dtype fp8 --block-size 256 \
  --data-parallel-size 4 --enable-expert-parallel \
  --moe-backend deep_gemm_mega_moe \
  --speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}'

One flag, and the open-weights frontier model runs at speculative-decoding-class throughput on commodity Blackwell. The recipe is identical to V4-Flash-0731, which means a single routing layer can drop in V4-Pro-0813 where V4-Flash-0731 was, expecting roughly the same per-token economics and a meaningful capability bump.

The RedHatAI community has been training DSpark speculators for every frontier open-weights model since July — GLM-5.2 got one on August 3, Kimi-K3 on August 12. With DSpark now in the Pro tier's primary checkpoint, the next reflex is to train a small DSpark speculator that taps V4-Pro-0813 hidden states for the Flash tier. Two weeks out, probably less.

MIT License Is The Real Headline

Open weights is one thing. MIT license is another. Every previous high-capability open-weights MoE shipped under either a custom non-commercial license, a community license with attribution and use restrictions, or a research-only license. V4-Pro-0813 is MIT. Fork it, fine-tune it, distill from it, ship it inside a proprietary product, sell fine-tunes. The only obligation is the copyright notice.

For an AI engineer / automation architect, this resolves the procurement question most teams have been punting on for two years: if I build my production stack on this model, can my legal team sign off? For an MIT-licensed 1.7T parameter model, the answer is yes. For a custom-license frontier model, the answer has historically been "schedule a meeting with legal ops in Q3."

This is the moment the open-weights frontier becomes the default frontier for production deployments where the legal team has a vote.

What To Ship This Week

1. Stand up V4-Pro-0813 on a single 4xGB300 node. FP8 KV-cache, expert parallelism on, DSpark enabled. Run an evaluation against your real agent traces by end of next week. 2. Re-tier your routing logic. V4-Flash-0731 stays as the cheap default. V4-Pro-0813 with reasoning_effort="max" is now the right escalation tier — not Sonnet 5, not Opus 4.8. You save real money and keep the capability ceiling for most agentic workloads. 3. If you fine-tune, use MIT. No more waiting on licensing review. The only requirement is the copyright notice. 4. Watch for the V4-Pro NVFP4 and ROCmFP3 quant packs. Geometric-AI shipped V4-Flash-0731 ROCmFP3-MIX within ten days of the original release. Budget 7-10 days for the same path on V4-Pro-0813. If you cannot wait, FP8 plus DSpark is the production config this week.

Mr. Tech's Take

The open-weights frontier is now the procurement frontier. For two years the API-vs-self-hosted argument was a capability argument: API wins the leaderboard, you pay for it, open weights lag a tier behind. That argument is structurally false now. V4-Pro-0813 is within two points of Opus 4.8 on nine agentic coding benchmarks, and it is MIT.

Three consequences land this quarter:

Closed labs lose their highest-margin customers. Enterprise procurement teams that paid Anthropic and OpenAI for "frontier tier" now have a legally clean, MIT-licensed alternative that runs on commodity Blackwell. The agentic coding tier in particular is now a margin-compressed market.
Routing startups collapse. If V4-Flash-0731 is the workhorse and V4-Pro-0813 is the escalation, the routing layer in the middle is the thin configuration. Most "AI routing" startups are wrappers around two or three proprietary APIs. That wrapper just became a wrapper around one open-weights model served from two clouds.
Fine-tuning becomes a real business again. V4-Pro-0813 is the right base to fine-tune for vertical agentic workloads because the license lets you ship the result and the agentic coding capability is already at the frontier — you are tuning for domain, not for capability.

V4-Pro-0813 is the open-weights tier inversion completing itself. The cheap tier caught up first (V4-Flash-0731, July 31). The Pro tier catches up this week. The frontier tier is in sight.

Watch two things in the next thirty days: (1) the V4-Pro NVFP4 quant pack landing on Hugging Face, which is when this model becomes the default production deployment for most platform teams; (2) whether Anthropic or OpenAI respond with a frontier-tier open-weights release of their own before Q3 close. If they do not, the agentic coding market in 2027 is a DeepSeek market.

— Mr. Technology

Sources: