DeepSeek Releases DeepSeek-V4-Pro-0813, a 1.7T-Parameter Model with DSpark Speculative Decoding

DeepSeek has released DeepSeek-V4-Pro-0813, the official production version of DeepSeek-V4-Pro, replacing the earlier preview build. The model weighs in at 1.7 trillion parameters, is distributed in BF16/F8_E4M3/I8/I64/F32 tensor formats, and is licensed under MIT, according to the model card on Hugging Face.

What changed from the preview

DeepSeek-V4-Pro-0813 is built on the same DeepSeek-V4-Pro (Preview) architecture but adds a DSpark speculative decoding module, which the company says improves inference throughput without a separate draft model — the target and draft weights come from the same checkpoint. DeepSeek states the new release shows the largest gains in "production environments" and in agentic task performance compared to the preview.

The model also expands the reasoning_effort parameter to three tiers — low, high, and max — controlling how much internal deliberation the model performs before responding. For high and max effort levels, DeepSeek recommends a maximum output length of 384,000 tokens.

Benchmark results

According to DeepSeek's technical report, DeepSeek-V4-Pro-0813 outperforms the preview version and DeepSeek-V4-Flash-0731 across a range of agentic and coding benchmarks:

• HLE (without/with tools): 42.7 / 60.0
• Terminal Bench 2.1: 87.9
• NL2Repo: 61.5
• Cybergym: 83.3
• DeepSWE: 62.7
• Toolathlon-Verified: 74.1
• Agents' Last Exam: 25.7
• AutomationBench (Public): 31.8
• DSBench-FullStack (internal): 71.1
• DSBench-Hard (internal): 67.2

DeepSeek claims these scores put the model "broadly competitive" with proprietary systems including GLM-5.2, Kimi K3, and Anthropic's Opus-4.8, though on several benchmarks — HLE, Terminal Bench, DeepSWE, and DSBench-Hard — Opus-4.8 and the unlisted "Fable-5" comparison model score higher. These figures come from DeepSeek's own report and have not been independently verified.

Deployment

DeepSeek-V4-Pro-0813 does not ship with a standard Jinja chat template. Instead, the company provides Python encoding scripts for converting OpenAI-compatible message formats into model input strings. The model card includes serving instructions for vLLM (tested on a 4×GB300 node with fp8 KV-cache and expert parallelism) and SGLang (with --speculative-algorithm DSPARK). No inference provider had deployed the model as of publication, and DeepSeek has not disclosed API pricing.

What this means

This is an incremental but substantive update: DeepSeek is formalizing a preview model into a production release while layering in a speculative decoding technique (DSpark) aimed squarely at inference cost and latency — a priority for any lab fielding a 1.7-trillion-parameter model. The benchmark gains over the preview are consistent and broad, particularly in agentic and coding-agent tasks (NL2Repo jumped from 38.5 to 61.5, DeepSWE from 12.8 to 62.7), suggesting the preview was genuinely unfinished rather than a marketing snapshot. Whether it holds up against Opus-4.8 and other frontier proprietary models in independent testing remains to be seen, since all comparative numbers here come from DeepSeek's own report. The absence of published pricing and provider deployment means real-world cost and latency — the two metrics DSpark is meant to improve — are still unknown.