Key Takeaways
- DeepSeek’s V4 Pro (build 0813) reached general availability on August 12, 2026: a 1.6-trillion parameter Mixture-of-Experts model with a 1 million token context window, self-reporting around 80.6% on SWE-Bench Verified — a figure not yet independently verified.
- The release coincides with DeepSeek resuming an $8 billion funding round at a reported $74 billion valuation, with capital earmarked for data centre expansion in Inner Mongolia, headcount growth and a 2.31% stake in humanoid robotics firm Unitree.
- API prices for V4 Pro and V4 Flash rise from August 17, with increases ranging from 50% to over 1,100% depending on model and usage tier, narrowing the cost gap that had been DeepSeek’s clearest competitive edge. DeepSeek’s V4 Pro scores around 80.6% on SWE-Bench Verified and ranks 12th on the Vals Index, trailing OpenAI’s GPT-5.5, Kimi K3 and Anthropic’s Claude Opus 5, and the company is raising API prices by as much as 1,100% starting August 17, ending the era of near-giveaway inference costs that defined its earlier market entry. The V4-Pro-0813 build became generally available on August 12, 2026, closing a preview period that opened in April.
What V4 Pro Actually Is
At 1.6 trillion parameters, V4 Pro is a Mixture-of-Experts model with a 1 million token context window as standard, supporting extended reasoning, software engineering and long-running agentic workflows. Its Hybrid Attention Architecture combines Compressed Sparse Attention and Heavily Compressed Attention, cutting single-token inference FLOPs to roughly 27% of V3.2’s at a 1 million-token context and KV cache requirements to around 10%. The training stack adds Manifold-Constrained Hyper-Connections for signal stability, a Muon Optimizer for faster convergence, and a dataset exceeding 32 trillion tokens. Mixed precision training uses FP4 for MoE expert parameters and FP8 elsewhere.
Alongside V4 Pro, DeepSeek released V4-Flash on July 31, 2026: a 284-billion parameter variant with the same 1 million token context window, open weights under an MIT license and pricing suited to high-throughput, lower-cost inference. The two-tier lineup gives enterprise buyers a choice between maximum capability and cost-efficient scale.
Benchmark Picture
On standard academic measures, V4-Pro-Base posts 90.1 on MMLU, 73.5 on MMLU-Pro, 92.6 on GSM8K and 76.8 on HumanEval. The V4-Pro-Max build reached a 3206 Codeforces Rating, ahead of Gemini-3.1-Pro High, and leads open-weights models on GDPval-AA with a score of 1554. The SWE-Bench Verified figure of around 80.6% places it roughly level with Gemini-3.1-Pro and a tenth of a point behind Anthropic‘s Claude Opus 4.6’s 80.8%, though recent Artificial Analysis Intelligence Index scoring puts V4-Pro-0813 at 53, on par with Zhipu AI’s GLM-5.2 from June but behind OpenAI‘s GPT-5.6 Terra and Moonshot AI‘s Kimi K3. The Vals Index ranking of 12th, behind GPT-5.5, Kimi K3 and Claude Opus 5, is harder to wave away, V4 Pro is competitive but not the frontier.
One figure that matters for production deployments: V4 Pro recorded a 94% hallucination rate on the AA-Omniscience benchmark in cases where the model lacks knowledge, meaning it nearly always generates a response rather than abstaining. For use cases where confidence calibration is critical, legal, financial, clinical, that number warrants specific mitigation design, not just a general RAG layer. The pattern of frontier models struggling with document-grounded accuracy makes this worth tracking across the V4 series as it matures.
The Funding Round
The model release lands as DeepSeek resumes its second external funding round, targeting $8 billion at a reported $74 billion valuation. The first external round, closed in June, raised approximately $7.4 billion at a $52 billion valuation, a sharp departure for a company that had avoided outside capital entirely until this year. Annualised revenue as of July was reported at between $400 million and $500 million. The gap between that revenue run-rate and the capital being sought reflects how expensive the next stage of competition has become: chips, data centre capacity and specialist engineering talent all require sustained capital investment regardless of model performance.
A substantial share of the new capital is earmarked for additional data centres and compute, with Inner Mongolia named as a priority location. DeepSeek also plans to at least double headcount across data centre and AI agent teams and has been hiring chip-design engineers, pointing toward development of proprietary AI processors to reduce dependence on Nvidia and Huawei. Vertical integration in inference infrastructure has become a recurring theme across the leading Chinese AI labs, and DeepSeek’s hiring pattern fits that direction.
The Unitree Stake
DeepSeek invested 140.8 million yuan (approximately $20.8 million) for a 2.31% stake in humanoid robot maker Unitree through that company’s Shanghai IPO placement. The arrangement is reciprocal: Unitree gains priority access to DeepSeek’s training services; DeepSeek gains priority access to Unitree hardware. The stated integration goal is pairing DeepSeek’s language models with Unitree’s motion-control systems to improve instruction comprehension and real-time task execution in physical robots. It is a small position financially but a directional signal, the lab is moving toward embodied intelligence, not just cloud inference.
Pricing Shift
DeepSeek’s price increases, effective August 17, introduce peak and off-peak rates across V4 Pro and V4 Flash. Depending on model tier and usage pattern, increases range from 50% to over 1,100%. V4 Pro output tokens were priced as low as $3.48 per million, compared to roughly $25 per million for Claude Opus 4.6, a gap that made DeepSeek the default cost argument in enterprise procurement conversations. The new structure narrows that gap considerably. Whether DeepSeek remains the low-cost option at scale depends on where individual organisations land across peak and off-peak usage, and the answer will vary. Teams currently budgeting on pre-August pricing should remodel their inference costs before the 17th.
Originally published at https://autonainews.com/deepseeks-v4-pro-debuts-fuels-8-billion-funding-push/
Top comments (0)