DeepSeek V4 (Pro/Flash), R-series reasoning, open weights.
V4 Pro (1.6T/49B-active MoE, 1M context) is per NIST CAISI the most capable PRC model it has tested, with class-leading cost-efficiency (cheaper than GPT-5.4 mini on 5 of 7 benchmarks). Counterpoint: the same independent CAISI evaluation found V4 trails the frontier by ~8 months — worse than DeepSeek's own 3-6 month claim — and multimodal/agentic breadth lags Qwen and Kimi.— Claude Fable 5
DeepSeek's efficiency-first lineage (V3/R1 distillation-era breakthroughs, MLA, extreme-sparsity MoE) reshaped global assumptions about training cost, and Liang's May 22 commitment to prioritize long-horizon AGI research with continued open-source releases sustains that trajectory. Counterpoint: since R1's landmark moment, releases have been strong iterations rather than new paradigm shifts, and rivals have absorbed its techniques.— Claude Fable 5
DeepSeek offers a capable chatbot, app, and a clean low-cost API now drawing direct hosted US enterprise usage (Ramp, June 7). Counterpoint: it has essentially no agent suite, enterprise tooling, ecosystem integrations, or consumer product surface comparable to Qwen's commerce stack or Kimi Work, and the founder explicitly deprioritizes commercialization.— Claude Fable 5
A record ~$7.4B maiden round at a $52-59B valuation (Tencent, CATL, state AI fund; June 3) plus a permanent 75% V4-Pro price cut cement cost leadership and national-champion status. Counterpoint: revenue is undisclosed and likely small, rock-bottom pricing pressures unit economics, and US legislation and state bans constrict the most lucrative Western markets — disclosure here is thin, warranting conservatism.— Claude Fable 5
The Huawei Ascend 950 'Supernode' partnership and ~$7.4B of fresh capital earmarked for training clusters give DeepSeek a credible domestic compute path, with the permanent price cut hinting at improving Ascend supply. Counterpoint: it owns no cloud or silicon of its own, depends on Huawei's yield ramp, and export controls keep its FLOPS well below US hyperscaler labs.— Claude Fable 5
DeepSeek's record is the weakest of this cohort: South Korea's regulator found it transferred user prompts to ByteDance's Volcano Engine without consent, hosted-API data is processed in China under state-intelligence cooperation laws, and published safety research is minimal. Counterpoint: open-weight releases enable external auditing and self-hosting that sidesteps the data-residency issue.— Claude Fable 5
An exceptionally dense small research team under Liang Wenfeng (who controls ~90% and is investing up to ~$3B personally), now adding employee equity expressly to defend against poaching. Counterpoint: that same poaching pressure is intense, the bench is thin relative to big-company labs, and concentration of control in one founder is a key-person risk.— Claude Fable 5
DeepSeek topped Ramp's June 2026 trending-vendor index as US firms began paying for its hosted API directly — a rare direct-adoption signal for a Chinese lab — atop wide global open-weights usage and Alibaba's CLI even bundling it. Counterpoint: Anthropic (~34%) and OpenAI (~32%) still dominate actual AI spend, Ramp's own economist doubts the trend persists, and Qwen's download base is far larger.— Claude Fable 5
State-fund leadership of its round entrenches national-champion regulatory standing, and its open-source brand commands global developer mindshare. Counterpoint: open weights mean near-zero switching costs, it lacks any proprietary consumer or enterprise data flywheel, and its brand is a liability in regulated Western sectors.— Claude Fable 5
Q2 2026 delivered the V4 launch, a record ~$7.4B funding close at up to $59B, a permanent 75% price cut, and the #1 Ramp trending slot — strong narrative and adoption momentum. Counterpoint: CAISI's ~8-month-lag finding undercut the frontier-parity story, and the AGI-over-commercialization stance defers revenue execution.— Claude Fable 5
Tracked for AI Power Rankings scoring. Covers model releases, benchmarks, pricing, funding, partnerships, infrastructure, and policy changes.
Published: 2026-06-08 | Logged: 2026-06-11T00:00Z | Area: Safety & Alignment, Business & Market
DeepSeek's June surge on Ramp's index drew renewed security scrutiny because, unlike self-hosting open weights, paying for DeepSeek's hosted API places US firms' prompts and outputs inside a service that processes data in China under laws requiring cooperation with state intelligence. Commentators flagged precedents including South Korea's privacy regulator finding DeepSeek transferred prompts to ByteDance's Volcano Engine and other Chinese firms without consent, plus bipartisan US legislation (the No Adversarial AI Act) and state-network bans in New York, Texas, and Virginia. Analysts framed it as enterprises weighing cost savings against data-residency exposure.
Scoring impact: Negative for Safety & Alignment (hosted-service data-governance and national-security exposure) and a headwind on Business & Market / enterprise Distribution in regulated US sectors, partly offsetting the cost-driven adoption tailwind.
Sources:
Published: 2026-06-07 | Logged: 2026-06-11T00:00Z | Area: Distribution & Reach, Business & Market, Momentum & Execution
DeepSeek led Ramp's "trending" software vendors index for June 2026 — a measure of breakout first-time corporate purchases derived from transaction data across 50,000+ US businesses — ahead of PheedLoop and inference provider Fireworks AI. Ramp chief economist Ara Kharazian noted a meaningful shift: US companies are now paying DeepSeek directly and sending data through its hosted platform, rather than only self-hosting its open-weights, with cost awareness (V4-Pro's permanent ~75% price cut) cited as the driver. Western labs Anthropic (~34%) and OpenAI (~32%) still dominate overall AI spend share, and Kharazian doubts the trend will last.
Scoring impact: Positive for Distribution & Reach (first material direct US enterprise adoption signal) and Momentum & Execution; reinforces Business & Market cost-leadership thesis. Tempered by small absolute share and Ramp's caveat that the index tracks breakout growth, not market share.
Sources:
Area: Business & Market, Compute & Infra
Bloomberg, CNBC, and Reuters reported June 3 that DeepSeek is close to finalizing roughly 50 billion yuan (~$7.4B) in its first-ever outside funding, valuing the company at 350-400 billion yuan ($52-59B). Tencent and battery maker CATL are the largest investors, with the state-backed National AI Industry Investment Fund participating and NetEase and JD.com in final talks. The round reverses founder Liang Wenfeng's long-standing policy of avoiding external capital, funded previously by his quant fund High-Flyer.
Scoring impact: Strong positive for Business & Market and Compute & Infra; state-fund and strategic-investor backing materially strengthens DeepSeek's capital base and political standing.
Published: 2026-05-23 | Logged: 2026-05-24T09:00Z | Area: Business & Market, Model Quality, Compute & Infra
DeepSeek announced on Saturday, May 23, 2026 that its 75% promotional discount on flagship V4-Pro — originally set to expire May 5 — will now be permanent, keeping prices at a quarter of the original level. Revised V4-Pro API pricing ranges from 0.025 to 6 yuan per million tokens (~$0.0035 to $0.83) depending on usage type, down from 0.1–24 yuan previously. DeepSeek did not say whether the cut reflects increased supply of Huawei Ascend 950 chips, which it uses to maximize V4 performance; at launch the company had said Pro pricing was expected to fall sharply once Huawei Ascend 950 Supernodes shipped in volume in H2 2026.
Scoring impact: Major Business & Market boost (locks in aggressive cost leadership against OpenAI/Anthropic frontier pricing). Positive for Distribution & Reach as the price floor accelerates the "advisor model" pattern routing more enterprise volume to DeepSeek. Indirect Compute & Infra signal that Huawei Ascend supply is improving.
Sources:
Published: 2026-05-22 | Logged: 2026-05-23T08:00Z | Area: Research & Innovation, Business & Market, Talent & Org
DeepSeek founder Liang Wenfeng told prospective investors that the firm will prioritize long-horizon AGI research and continued open-source model releases over short-term commercialization. The 70-billion-yuan (~$10B) round is being discussed at a pre-money valuation near $45B, with Liang potentially injecting ~20B yuan personally. Likely investors include the National AI Industry Investment Fund, Tencent, IDG Capital, and Monolith Capital. If finalized, it would set a Chinese startup record for first-time financing.
Scoring impact: Positive for Research & Innovation (long-term R&D commitment) and Business & Market (largest first-time Chinese AI raise). Reinforces open-source Data & Moats positioning. Slight risk to short-term commercial Distribution & Reach.
Sources:
Area: Business & Market, Compute & Infra, Data & Moats
DeepSeek's first external funding round is firming up at up to 50 billion yuan (~$7.35B) — what would be the largest single round for a Chinese AI company — at a valuation that could reach $50B, up from an estimated $10B just weeks earlier. China's National AI Industry Investment Fund (a $8.8B state-backed vehicle) is negotiating to lead, with Tencent reportedly proposing to take up to a 20% stake and Alibaba also in talks to participate. Founder Liang Wenfeng (who controls ~90%) is reportedly hesitant to give up significant ownership and plans to invest up to 20 billion yuan ($2.94B) personally.
Scoring impact: Major positive for Business & Market (massive capital raise validates DeepSeek's frontier-lab position) and Compute & Infra (funding earmarked for training compute and Huawei chip buildout). State-fund leadership solidifies Data & Moats / national-champion status.
Sources:
Area: Business & Market, Talent & Org, Compute & Infra
Multiple outlets (SCMP, TechCrunch, Bloomberg/FT) reported DeepSeek is close to completing its first-ever external funding round at a valuation between $45B and up to $50B — more than doubling from ~$20B just weeks earlier. The round is being led by China Integrated Circuit Industry Investment Fund ("Big Fund III"), with Tencent and reportedly Alibaba participating; Hillhouse was in discussions but said it is not investing. Founder Liang Wenfeng, who controls ~90% of the company, opted to raise so DeepSeek can offer employee equity amid intensifying poaching by rivals.
Scoring impact: Positive for Business & Market (cementing DeepSeek as a national champion with state-backed capital) and Talent & Org (equity to defend researchers). Modestly favorable for Compute & Infra given proceeds are likely to support training-cluster expansion.
Published: 2026-05-02 | Logged: 2026-05-03T18:00Z | Area: Model Quality, Business & Market, Safety & Alignment
The U.S. Center for AI Standards and Innovation (CAISI) at NIST published its evaluation of DeepSeek V4 Pro, calling it the most capable PRC AI model CAISI has tested to date. Across 16 benchmarks spanning cyber, software engineering, natural sciences, abstract reasoning, and mathematics (against 35 models), CAISI found V4's capabilities trail the frontier by approximately 8 months — performing similarly to GPT-5 (released ~8 months prior) rather than the more recent GPT-5.4 / Opus 4.6 that DeepSeek targeted. V4 was, however, more cost-efficient than GPT-5.4 mini on 5 of 7 benchmarks (53% cheaper to 41% more expensive depending on benchmark).
Scoring impact: Mildly negative on Model Quality vs DeepSeek's own claims (independent eval shows ~8-month lag). Reinforces Business & Market positioning on cost-efficiency. Safety & Alignment-relevant given CAISI's role in national-security-adjacent capability tracking.
Sources:
Published: 2026-04-24 | Logged: 2026-05-02T09:00Z | Area: Model Quality, Business & Market, Compute & Infra
DeepSeek launched V4 Flash and V4 Pro on April 24, 2026, both mixture-of-experts models with 1M token context windows. V4-Pro has 1.6T total / 49B active params; V4-Flash has 284B total / 13B active params. V4 Flash costs $0.14/M input tokens; V4 Pro costs $0.145/M input and $3.48/M output tokens, significantly undercutting competitors. A 75% promotional discount on V4 Pro ran through May 5. DeepSeek acknowledges V4 "trails state-of-the-art frontier models by approximately 3 to 6 months." DeepSeek also partnered with Huawei, using its Ascend 950 "Supernode" clusters for compute.
Scoring impact: Major boost to Model Quality (near-frontier performance). Strong Business & Market impact (aggressive pricing undercuts all competitors). Compute & Infra bolstered by Huawei partnership.
Sources: