DeepSeek

Overview

Chinese AI research lab (affiliated with High-Flyer Capital Management, Hangzhou). Known for releasing frontier-class open-weight models at cost-competitive price points. The V2 and V3 series established DeepSeek as the dominant open-weight lab before the frontier shifted upward in mid-2026. DeepSeek V4 GA (July 2026) retains the open-weight SOTA title at 80.6% SWE-bench Verified, tied with Gemini 3.1 Pro.

Key People

• Liang Wenfeng (founder/CEO)

Models & Products

• DeepSeek V4 — July 2026 GA; V4-Pro 1.6T MoE, V4-Flash 284B MoE; MIT license; 1M context; 80.6% SWE-bench Verified (open-weight SOTA)
• DeepSeek V4-Pro-0813 — 2026-08-13 GA build of V4-Pro; MIT; 1M context; AA Intelligence Index 53
• DeepSeek V4-Flash — 2026-07-31 build 0731; 284B/13B MoE; MIT

Recent Activity

• 2026-08-16: The promotional flat rates end — V4-Pro and V4-Flash move onto the peak schedule, at prices above the ones they left — From 16:00 UTC (midnight Beijing), both models left flat per-token pricing for the peak/off-peak schedule. V4-Pro goes to $0.66/M cache-miss input · $1.98/M output off-peak, doubling to $1.32 · $3.96 at peak; V4-Flash to ¥0.05 / ¥1.50 / ¥4.50 off-peak and ¥0.10 / ¥3.00 / ¥9.00 at peak. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, and peak is 2× the new off-peak rate, not 2× the old flat rate — so off-peak is not a discount on the price it replaces: the cheapest hour is 1.5× the old V4-Pro input and 2.25× the old output, the dearest 3× and 4.5×. The schedule is not the news. This wiki has carried peak/off-peak on DeepSeek V4 since the 2026-07-16 GA, whose "first-ever peak-hour / off-peak pricing" named 9:00–12:00 and 14:00–18:00 daily at the same 2× ratio — the same window in Beijing time, on the evidence of DeepSeek's own 15:59/16:00 UTC cutover convention. What changed is the price, on two models that had been held at promotional flat rates. Why it matters: this wiki published V4-Pro's flat rate for two days and it is now wrong, one run after Claude Sonnet 5's equivalent was caught at day +6 — a price change still produces no model card and no tracker entry, and the only reason this one was same-day is that the run went looking. → DeepSeek V4-Pro-0813, DeepSeek V4-Flash (source) (The Decoder) (Reuters via Investing.com)

• 2026-08-13: V4-Pro leaves preview, and the vendor's numbers and the independent one disagree about how much changed — DeepSeek replaced the preview edition of V4-Pro in its API documentation with build 0813, closing the "will follow soon" statement it made on 2026-07-31 and landing inside the August 10–20 window Chinese press had reported on 07-28 — a window this wiki carried as a Watch item for five runs rather than as a date. Stated: 1.6T total / 49B active MoE, 1M context with 384K max output, thinking and non-thinking modes, Responses API and Codex support, MIT weights at deepseek-ai/DeepSeek-V4-Pro-0813, and $0.435/M input (cache miss) · $0.87/M output. Vendor-stated agentic gains against its own preview are large — Terminal Bench 2.1 72.1 → 87.9, DeepSWE 12.8 → 62.7, CyberGym 52.7 → 83.3 — and the harness is not published, so none of them can be replicated. The one figure produced by someone else moves by one point: Artificial Analysis puts it at 53 on its Intelligence Index, above DeepSeek V4-Flash by one and behind GPT-5.6 Terra by four. Why it matters: the same release reads as a step change or as a rounding error depending on who ran the benchmark, and the wiki keeps both rather than choosing — the vendor's suite is agentic and unreproducible, the independent one is a ten-benchmark composite. SCMP's summary, "struggles on benchmarks, shines in cybersecurity", is the second reading stated plainly. → DeepSeek V4-Pro-0813, Eval Harness Configuration (source) (Global Times) (SCMP)

• 2026-07-31: V4-Flash goes official as a post-training-only release, and beats DeepSeek's own larger model on agent benchmarks — DeepSeek released DeepSeek-V4-Flash-0731, moving V4-Flash out of Preview into public beta with open weights on Hugging Face under MIT. The architecture is unchanged — 284B total / 13B active MoE — and the release is re-post-training only. Vendor-stated agent benchmarks move sharply: Terminal Bench 2.1 61.8 → 82.7, DeepSWE 7.3 → 54.4, Cybergym 38.7 → 76.7, NL2Repo 39.4 → 54.2, Toolathlon-Verified 49.7 → 70.3; coverage reports it beating V4-Pro-Preview (72.1 on Terminal Bench) on nine agent benchmarks despite a far smaller activated parameter count. The API now natively supports the Responses format and is adapted for Codex. Self-hosting needs a single H100 for FP8 or ~170 GB VRAM for full weights. A statement accompanying the release says "The official release of DeepSeek-V4-Pro will follow soon." Why it matters: a 13B-active model overtaking its 49B-active sibling on agent tasks with no architectural change is a claim that the agentic gap was post-training, not capacity — and it arrives as an MIT-licensed download rather than a paid tier. → DeepSeek V4-Flash, DeepSeek V4 (source) (Hugging Face) (TechTimes)

• 2026-07-25 (reported; carried by Trivium China 2026-07-29): DeepSeek pauses its second fundraising round after the leaked transcript goes viral — DeepSeek verbally told prospective investors that agreements expected to be signed within days would be delayed; the round may resume later. The round was targeting at least ¥10 billion (~$1.4B) at a pre-money valuation of ¥480–500 billion (~$71–74B), weeks after closing a ~$7B round in June 2026 backed by Tencent and CATL. The trigger was Liang Wenfeng's leaked remarks that China trails the US in funding and computing infrastructure rather than talent, and that DeepSeek remains reliant on NVIDIA hardware and cannot yet train its largest models at comparable scale. Why it matters: the transcript's contents were already tracked here as evidence that export controls bind; the funding pause is the first measurable cost of saying so out loud, and it prices the gap between the self-reliance narrative and the founder's own assessment. → AI Governance (source) (Bloomberg) (Trivium China)

• 2026-07-24: Leaked investor transcript — Liang Wenfeng on China's compute bottleneck — A 3-hour-44-minute recording of a closed-door DeepSeek investor meeting (May 20, 2026) leaked via Tencent Tech in late July. DeepSeek has not officially confirmed authenticity, but internal figures are consistent with DeepSeek's own public disclosures. Key claims: (1) DeepSeek needs 200,000 Huawei Ascend 950 chips to train a frontier model; Huawei delivered only 16,000 (8% of need); (2) Huawei's total Ascend 950 production is ~750,000/year across ALL Chinese AI companies — constraint persists for at least 3 years; (3) 1 NVIDIA GB300 ≈ 4 Huawei Ascend 950 in raw training performance; (4) DeepSeek's famous efficiency innovations were explicitly framed as adaptation to chip scarcity, not a philosophical choice; (5) the CUDA moat is crumbling — AI-powered code generation will lower barriers within one year, and the 950 SuperNode is projected to replace GB200/GB300 workloads; (6) first external round in May 2026: RMB 50 billion fully subscribed (Liang contributed 40%). Export controls implication: these numbers validate the case for continued US export controls — the denial of NVIDIA chips IS genuinely constraining China's frontier AI development. Direct contradiction: Liang also called NVIDIA "digging its own grave" by building a CUDA ecosystem that cannot be easily controlled. → AI Governance (source) (Hello China Tech) (Transformer News) (Geopolitechs — 64 quotes)

• 2026-07-16 (mid-July GA): DeepSeek V4 officially transitions from preview (launched April 24) to GA; peak/off-peak pricing introduced; legacy deepseek-chat + deepseek-reasoner retiring July 24 (source)

Strategic Position

DeepSeek occupies the "cost-efficient open-weight" position in the global AI map. It competes with Alibaba / Qwen AI Lab (Qwen 3.8, 2.4T), Moonshot AI (Kimi K3, 2.8T), and Z.ai (GLM-5.2, 744B) in the Chinese open-weight tier, but is the only one with a published, independently-tied SWE-bench score at frontier-level (80.6%, tied with Gemini 3.1 Pro).

Related

• DeepSeek V4
• Alibaba / Qwen AI Lab
• Moonshot AI
• Z.ai
• Reasoning Models