DeepSeek V4-Pro Leaves Preview as API Prices Rise

DeepSeek has taken V4-Pro out of preview. The model card for DeepSeek-V4-Pro-0813 went up on Hugging Face on August 13, 2026, closing a preview period that ran since the V4 family launched in late April. The weights stay under an MIT licence, the architecture is unchanged from the preview, and the headline addition is a speculative decoding module called DSpark. Announced alongside it: a rebuilt API price list that takes effect at 00:00 Beijing time on August 17 and raises some rates by as much as twelve times.

Advanced

What Shipped

V4-Pro-0813 keeps the shape described in the DeepSeek-V4 technical report (arXiv:2606.19348, submitted April 26, 2026): a 1.6-trillion-parameter Mixture-of-Experts model activating 49 billion parameters per token, with a one-million-token context window. The report describes a hybrid attention stack combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), Manifold-Constrained Hyper-Connections (mHC) on the residual path, and the Muon optimizer. DeepSeek reports that combination needs “only 27% of single-token inference FLOPs and 10% of KV cache compared with DeepSeek-V3.2.”

What is new in 0813 is bolted on rather than rebuilt. The model card states the release “is built on the DeepSeek-V4-Pro (Preview) model structure, with a DSpark speculative decoding module attached” — a draft-and-verify head that ships inside the same checkpoint, so there is no separate draft model to load. DSpark postdates the April technical report and is not described there.

Two changes matter for anyone wiring this into an existing stack. First, the release ships no Jinja chat template; DeepSeek replaced it with an encoding folder of Python scripts that convert OpenAI-format messages into input strings and parse the model’s output back. Second, reasoning_effort now takes three levels — low, high, and max — controlling how long the model deliberates. For the two upper levels DeepSeek recommends allowing up to 384K output tokens.

Serving it is a one-flag change on both major engines. On vLLM, DSpark is enabled with:

--speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}'

On SGLang it is --speculative-algorithm DSPARK, with no --speculative-draft-model-path, because target and draft weights come from the same checkpoint. DeepSeek’s reference vLLM command serves the model on a single four-way GB300 node using FP8 KV cache and the deep_gemm_mega_moe expert backend.

Benchmarks

The gains over the preview are concentrated in agent work, and some are very large. Terminal Bench 2.1 moves from 72.1 to 87.9. NL2Repo goes from 38.5 to 61.5. Cybergym goes from 52.7 to 83.3. DeepSWE, where the preview scored a barely-functional 12.8, reaches 62.7 — a result that says more about how unfinished the preview’s agent loop was than about the underlying model. Humanity’s Last Exam moves from 37.7 to 42.7 without tools, and from 48.2 to 60.0 with them.

Against other frontier systems in DeepSeek’s own table, 0813 lands in the pack rather than ahead of it. Kimi K3 leads Terminal Bench 2.1 at 88.3 to DeepSeek’s 87.9, with Fable 5 at 88.0 and Opus-4.8 at 85.0. On DeepSWE, Fable 5 (70.0) and Kimi K3 (67.5) both finish ahead of 62.7, while Opus-4.8 trails at 58.0. Two caveats travel with these numbers: they are vendor-run, and the code-agent rows were measured using DeepSeek’s own agent framework in “minimal mode” at max reasoning effort with temperature = 1.0, top_p = 0.95 — a harness choice that is part of the score. That framework, DeepSeek Harness, is now public under MIT as a plugin-based, developer-preview agent runtime.

Independent numbers exist. Artificial Analysis places V4-Pro-0813 at 53 on its Intelligence Index, third of 106 models evaluated, measuring 80 output tokens per second and 1.85 seconds to first token.

The Price Change

DeepSeek is also introducing time-of-day pricing. From 00:00 Beijing time on August 17, 2026, peak hours are 09:00–12:00 and 14:00–18:00 Beijing time, with every other hour billed at half the peak rate. For deepseek-v4-pro, peak rates are RMB 0.30 per million cache-hit input tokens, RMB 9.00 cache-miss, and RMB 27.00 output; off-peak is RMB 0.15 / 4.50 / 13.50.

Measured against the old flat rates, that is a 3× rise on cache-miss input and 4.5× on output at peak — and 12× on cache-hit input, which had been priced at RMB 0.025 per million. Prompt caching was the cheapest thing DeepSeek sold, and it is the line that moves furthest. Workloads built around a large cached system prompt and short completions will feel this more than the headline output multiplier suggests.

What This Means

The interesting part of 0813 is not the parameter count, which did not change, but that the entire jump came from inference-side and harness-side work on a frozen backbone: a speculative decoding head, a reasoning-effort dial, and a published agent framework. DeepSeek is shipping the scaffolding around the model as a product surface, and the DeepSWE result is the clearest evidence of how much of an “agentic” score lives in that scaffolding rather than in the weights.

The pricing move points the same direction as Kimi K3 and the recent Qwen releases: Chinese labs that built their reputation on undercutting Western API prices are now charging closer to what serving a trillion-parameter MoE at a million tokens of context actually costs. DeepSeek’s peak output rate of RMB 27.00 per million works out near US$3.90 — no longer a rounding error against proprietary competitors. For teams that can self-host, the MIT licence and the published vLLM and SGLang recipes remain the real offer; for everyone else, the arbitrage that made DeepSeek’s API the default cheap option is narrowing.

Related Coverage

• DeepSeek Releases V4: Open-Source 1.6T MoE with 1M Context — the April 2026 launch this release graduates from preview
• DeepSeek Founder Details AGI-First, Compute-Bound Strategy in Investor Meeting — the strategy context from July 2026
• Kimi K3 Open Weights Ship: 2.8T Parameters, 1.4 TB to Run — the model that edges out V4-Pro on Terminal Bench
• Qwen3.8-2.4T-A95B: Alibaba Open-Weights Its Max-Tier Flagship — released one day earlier

This post was drafted with AI assistance and reviewed by RITS staff.

Sources

• DeepSeek-V4-Pro-0813 model card — Hugging Face
• DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence — arXiv:2606.19348
• deepseek-ai/deepseek-harness — GitHub
• DeepSeek V4 Pro 0813 — Artificial Analysis
• DeepSeek Ships V4 Pro as Its Flagship Model Leaves Preview — Unite.AI
• DeepSeek Raises API Prices by Up to 12x With the Official V4 Pro Release
• DeepSeek API changelog — model updated to DeepSeek-V4-Pro-0813