DeepSeek V4 Pro / V4 Flash

DeepSeek V4 Pro and V4 Flash are an open-weights mixture-of-experts model family released in preview on April 24, 2026, with both checkpoints open-sourced alongside a technical report. V4 Pro has 1.6T total parameters (49B active) and V4 Flash has 284B total (13B active); both carry a 1M-token context window. SemiAnalysis assessed V4 Pro as roughly 3–6 months behind US frontier models while being the cheapest model in its tier. (Source: bloomberg.com; simonwillison.net; newsletter.semianalysis.com)

The 1M-token context window is a step up from V3's 128K. Both versions are reported as the cheapest in their respective classes.

A July 19, 2026 report said DeepSeek was preparing to launch the full V4 in Pro and Flash versions as early as July 20, 2026, with some users already gray-scale-tested onto V4; early tests reportedly showed lower pricing, off-peak billing, and performance near Claude Opus level. The report is single-source and pre-launch (Source: aidisruption.ai).

V4-Flash-0731 public beta

DeepSeek moved the V4-Flash API to public beta on July 31, 2026, releasing the checkpoint as DeepSeek-V4-Flash-0731. The company states that the checkpoint keeps the same architecture and size as V4-Flash-Preview and was only re-post-trained. The release natively supports the Responses API format and is adapted for Codex; the V4-Pro API and the app and web models were unchanged, with DeepSeek saying an official V4-Pro release would follow (Source: api-docs.deepseek.com; bloomberg.com).

The checkpoint was published on Hugging Face with MIT-licensed, ungated weights. The repository reports 304B parameters against the 284B of the V4 technical report, because the checkpoint ships with the DSpark speculative-decoding draft module attached (Source: marktechpost.com; huggingface.co).

DeepSeek reported agent-benchmark scores it says exceed those of its larger V4-Pro-Preview, produced using its own DeepSeek Harness in minimal mode at max effort with topp=0.95 and temperature=1.0. That harness has not been released, so the figures cannot be independently reproduced under the same conditions (Source: marktechpost.com):

(Source: api-docs.deepseek.com)

Independent measurement of the 0731 checkpoint

Artificial Analysis scored DeepSeek V4 Flash 0731 at 50 on its Intelligence Index on July 31, 2026 — ten points above the April 2026 V4 Flash and six above V4 Pro, one point behind GPT-5.6 Luna (max, 51) and seven behind Kimi K3 (max, 57). Its GDPval-AA v2 Elo rose to 1559 from 1189. Its AA-Omniscience hallucination rate fell to 84%, a decline the article states as 12 points in text and 11 points in a chart caption, with accuracy unchanged at 37%. Artificial Analysis measured Terminal-Bench 2.1 at 79%, against the 82.7 DeepSeek reported using its own harness — the gap between vendor-harness and independent measurement being the reason both figures are recorded here.

Pricing was unchanged at $0.14/$0.28 per 1M input/output tokens, with a cache-hit price of $0.0028 per 1M tokens, a 98% discount that Artificial Analysis identifies as the main reason the model's cost per task runs roughly 60% below GPT-5.6 Luna (max) even after that model's 80% price cut. Weights were expected "in the coming weeks" (Source: artificialanalysis.ai).

V4-Pro-0813 release

DeepSeek released the flagship checkpoint as V4-Pro-0813 on August 12, 2026 across the DeepSeek API and DeepSeek Chat, priced at $0.435 per 1 million input tokens and $0.87 per 1 million output tokens — a quarter of the $1.74/$3.48 charged for the April preview. A leaderboard compiled by the benchmarking startup Vals AI ranked V4-Pro second among all open-source models, behind Moonshot AI's Kimi K3, which charges $3 per million input and $15 per million output tokens. Engineers testing the model and posting on Chinese social media reported that it sometimes lost continuity in its reasoning, performed weakly on image tasks, and on certain coding tasks trailed DeepSeek's smaller V4-Flash. DeepSeek was second only to Anthropic in total tokens consumed in July 2026 (Source: wccftech.com; theinformation.com). Later reporting dates the official version of V4-Pro-0813 to August 13, 2026, against the August 12 date above; both are recorded (Source: venturebeat.com).

The 0813 launch pricing did not hold. From 16:00 UTC on August 16, 2026 DeepSeek replaced flat API rates with peak and off-peak pricing, moving V4-Pro to $0.66 per million cache-miss input tokens and $1.98 per million output tokens off-peak, and $1.32 and $3.96 at peak; Reuters reported the increases range from 50% to more than 1,100% across models, token categories and times of use (Source: venturebeat.com).

DeepSeek Harness v0.1

Alongside the 0813 release, DeepSeek published DeepSeek Harness v0.1 on August 13, 2026 as an MIT-licensed agent harness in developer preview, built on the Cordis framework. Its stated premise is that models, tools, skills, sessions, sandboxes, filesystems, loops, orchestration and user interfaces are all replaceable plugins (Source: venturebeat.com). The release bears on the reproducibility caveat recorded above: the DeepSeek Harness used to produce the V4-Flash-0731 agent-benchmark figures was unreleased at that time, and whether v0.1 is the same harness in the configuration those runs used is not stated. See Agentic harnesses and capability elicitation and AI Coding Agents.

Training and architecture

DeepSeek describes two attention changes in the V4 family: Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA). DeepSeek claims these cut single-token inference FLOPs to 27% and the KV cache to 10% of V3.2 at 1M-token context, following the company's earlier architectural-efficiency work; see Distillation and Mixture of Experts (MoE). DeepSeek attributes API pricing it describes as 80–100× cheaper than US frontier rivals to this hybrid Compressed-Sparse plus Heavily-Compressed Attention architecture. (Source: huggingface.co)

According to ChinaTalk's April 27, 2026 reporting on the V4 technical report (released April 23, 2026), V4 was still trained on Nvidia chips, with Huawei Ascend support confined to post-training and inference. ChinaTalk also reported that DeepSeek suffered a mid-2025 training failure during its Huawei-Ascend migration, and that the company lost core-LLM, agents, OCR, and multimodality talent to ByteDance, Tencent, Xiaomi, and DeepRoute.ai over the prior year. (Source: chinatalk.media)

Capabilities and benchmarks

The DeepSeek V4 Technical Report (April 27, 2026) reports the following for V4-Pro-Max:

(Source: huggingface.co)

In a head-to-head coding test published April 27, 2026 covering TypeScript LRU Cache and CLI engineering tasks, GPT-5.3 Codex High beat DeepSeek V4 Pro on quality, though V4 Pro's price-per-token undercut OpenAI by roughly 8x. (Source: aidisruption.ai)

Independent evaluation

SemiAnalysis, in hands-on testing on April 24, 2026, assessed V4 Pro as lagging US frontier models by roughly 3–6 months and found that Claude Opus 4.7 still beats V4 Pro on difficult Chinese-language writing tasks. SemiAnalysis praised the new attention architectures and described cost-performance as V4's strongest argument. (Source: newsletter.semianalysis.com)

ChinaTalk's April 27, 2026 account noted that DeepSeek's authors themselves conceded V4 is "3 to 6 months behind" frontier closed models, matching SemiAnalysis's earlier independent assessment. ChinaTalk presented this as a counter to a narrative of China efficiency-first decoupling. (Source: chinatalk.media)

MIT Technology Review's Caiwei Chen (April 24, 2026) characterized V4 as more efficient and a win for Chinese chipmakers, framing the release in terms of DeepSeek's compute-efficiency narrative rather than as independent technical evidence. (Source: technologyreview.com)

UK AISI cyber evaluation (July 2026)

The UK AI Security Institute published its first public measurement of the open-weight cyber gap on July 17, 2026, using V4-Pro and GLM-5.2 as the two open-weight models tested. On AISI's narrow cyber tasks — a 70-task subset covering vulnerability research and exploitation, reverse engineering, web exploitation and cryptography across four difficulty levels — V4-Pro performed comparably to Claude Opus 4.5, released five months earlier. On AISI's cyber ranges, which measure end-to-end autonomous attack execution, V4-Pro fell below Claude Sonnet 4.5, a model AISI describes as sub-cyber-frontier and released seven months earlier. AISI aggregated the two open-weight results into a 4-to-7-month lag behind the closed frontier, narrower than the 6-to-10-month gap it measured internally for open-weight models released between January and September 2025 (How Far Behind the Frontier are Leading Open Weight Models on Cyber? (UK AISI, July 2026)).

Two further findings bear on the safeguards and cost arguments the V4 release is cited in. On safeguards, AISI reported that its evaluations of recent open-weight models were largely unimpeded: V4-Pro occasionally refused narrow cyber tasks, mainly in reverse engineering, and AISI circumvented the refusals "simply via a small number of repeat attempts at refused tasks." On cost, at advertised first-party prices a 100-million-token cyber-range run cost an estimated $1.19 for V4-Pro against roughly $85 for Opus 4.5, and across tasks both models solved with full reliability V4-Pro cost $0.28 per task against Opus 4.5's $12.50 — the widest cost-per-capability spread in AISI's comparison. AISI noted it did not use first-party providers for the open-weight models it tested, so real compute costs may vary, and that its setup likely slightly underestimates open-weight maximum capability because it pursued no model-specific elicitation (How Far Behind the Frontier are Leading Open Weight Models on Cyber? (UK AISI, July 2026)).

Hardware support

The Huawei Ascend Supernode, based on the Ascend 950 chip, was announced the same day as supporting V4 inference. (Source: reuters.com)

Policy context

The V4 release was tied by several commentators to the US-China AI race frame. It coincided with the OSTP "Adversarial Distillation of American AI Models" memo (NSTM-4), issued April 23, 2026, and the BIS H200 China block confirmed in the same news cycle (April 22, 2026 Reuters); see Export Controls (AI). On April 24, 2026 the State Department ordered a global diplomatic cable spotlighting alleged DeepSeek IP theft. (Source: reuters.com) SCSP framed the V4 release and the OSTP memo as joint pivots in US-China AI competition. (Source: scsp222.substack.com)

Relationships

• instance-of: Open-Source AI / Open-Weight Models, Mixture of Experts (MoE)
• depends-on: Scaling Laws, Inference Economics and Token Pricing
• supports: AI Race Dynamics, Distillation
• deployed-by: DeepSeek
• regulated-by: UK AI Safety Institute (AI Security Institute) — subject of the July 2026 open-weight cyber evaluation described above.
• related: DeepSeek-V3, DeepSeek-R1, GLM-5.2 — the paired open-weight model in the AISI cyber comparison; Adversarial Distillation, Export Controls (AI), Open-Weight Frontier Models