DeepSeek

lab
ai-labopen-weightsmixture-of-expertsfrontier-modelchina

DeepSeek

Type: AI Research Lab (open-weight frontier models)

DeepSeek is the leading open-weight entrant at the 2026 frontier and the entity most associated in this KB with the thesis that the closed-model capability lead has compressed to a thin, hardest-tasks-only margin while price collapses. It appears through two strands: a frontier-scale open-weight MoE model family (V4), and primary architecture research on new axes of sparsity.

In the April-2026 frontier-release wave, DeepSeek V4 (24 Apr 2026, MIT-licensed) landed the same week as GPT-5.5 (Apr 23) and days after Claude Opus 4.7 (Apr 16) — an ~8-day window that compressed the frontier. V4 ships two MoE variants (V4-Pro: 1.6T total / 49B active; V4-Flash: 284B / 13B active), a 1M-token context, and roughly one-sixth the cost of its closed peers, reaching near-state-of-the-art capability (SWE-bench Verified ~80.6%, GPQA ~90–92%) on saturated benchmarks while still trailing the closed models on the hardest agentic-coding splits.

On the research side, DeepSeek (with PKU collaborators) authored "Conditional Memory via Scalable Lookup" (Jan 2026), introducing the Engram module and a U-shaped scaling law for allocating capacity between MoE compute-sparsity and static memory-sparsity — a frontier-architecture contribution beyond just shipping a model.

DSpark — Publishing the Serving Stack (Jul 2026)

DSpark (arXiv:2607.05147, 2026-07-06) is a speculative-decoding framework measured inside DeepSeek's own production system, and it is the most commercially revealing DeepSeek disclosure in this KB. The paper's abstract states it was deployed "within the DeepSeek-V4 serving system under live user traffic"; the author roster includes Wenfeng Liang and Damai Dai, DeepSeek core research names. (Institutional affiliation is not stated verbatim in the fetched abstract — the DeepSeek attribution rests on the serving-system reference and the author roster, both of which are explicit.)

Two mechanisms:

  • A semi-autoregressive architecture — a parallel drafting backbone coupled to a lightweight sequential module — restoring intra-block token dependency to fight the acceptance decay that plagues parallel drafters.
  • Confidence-scheduled verification — verification length is tailored per request from estimated prefix-survival probability and engine-specific throughput profiles, so batch capacity is not wasted verifying tokens likely to be rejected.

The measured result: 60–85% faster per-user generation at matched throughput versus the established production baseline (MTP-1), and — the claim DeepSeek chose to emphasize — prevention of severe throughput degradation under strict interactivity constraints, "shifting the Pareto frontier of our serving system."

Why this is strategically notable: DeepSeek's public identity in this KB is cost disruption (V4 at ~1/6th closed-peer pricing). DSpark shows part of the machinery behind that, and shows it is serving-layer software on unchanged hardware — no new silicon, no new weights. A lab whose competitive position rests on cost publishing its inference-cost optimization is an unusual choice; whether it reads as confidence, recruiting, or ecosystem-building is an open question. Evidence: moderate — preprint, not peer-reviewed, and the production figures are first-party measurements of the authors' own system with no independent reproduction.

The Repricing — Cost Disruption Ends (Aug 2026)

On 2026-08-16 at 16:00 UTC, DeepSeek raised its API prices, and this is the entity in this KB whose entire public identity was that it would not. The lab warned developers on 6 August that a "significant" increase was coming without publishing rates; the rates landed with the change itself. Flat per-token pricing — in place since May 2026 — was replaced by a two-tier peak/off-peak structure, described in DeepSeek's own changelog as "we will adopt peak/off-peak pricing, with off-peak prices set at half of the peak-hour prices." Peak hours are 01:00–04:00 and 06:00–10:00 UTC; all other hours are off-peak.

V4-Pro, per MtokFlat, to 2026-08-16Off-peak, from 2026-08-16Peak, from 2026-08-16
Input (cache miss)$0.435$0.66 (+52%)$1.32 (+203%)
Input (cache hit)$0.003625$0.022 (6.1×)$0.044 (12.1×)
Output$0.87$1.98 (+128%)$3.96 (+355%)
V4-Flash, per MtokFlat, to 2026-08-16Off-peakPeak
Input (cache miss)$0.14$0.22 (+57%)$0.44 (+214%)
Input (cache hit)$0.0028$0.007 (2.5×)$0.014 (5×)
Output$0.28$0.66 (+136%)$1.32 (+371%)

The cached-input line is the one to read. Every other item roughly doubles to quadruples; cached input rises 6× off-peak and 12× at peak. Cached input is what agentic, multi-turn, long-context workloads consume most of — the harness layer's characteristic cost. DeepSeek raised the price of the thing its own newly-published agent harness generates, three days after publishing it.

Why the increase is more interesting than its size. DeepSeek is the KB's standing example of the argument that near-frontier capability is available at a fraction of closed-model cost, and the DSpark disclosure above (same page) showed the serving-side machinery that made it possible. A lab that publishes how it drives serving cost down and then raises price is saying something about demand it cannot serve, not about cost it cannot control — which is the capacity-rationing mechanism described in price-rung persistence, observed for the first time as an outright increase rather than a hold.

Evidence: strong — read 2026-08-18 from DeepSeek's own pricing page and changelog; prior rates from the Token Price Index rows observed 2026-08-03. The 6 August warning is secondary (wire coverage); the 16 August effective date and the structure change are primary.

V4-Pro reached general availability on 2026-08-13 (DeepSeek-V4-Pro-0813), rolled out to app, web and API — three days before the repricing. The GA and the increase are one commercial move read in sequence, not two unrelated events.

Key Contributions

  • DeepSeek V4 (open-weight, MIT): 1.6T/49B-active (Pro) + 284B/13B-active (Flash) MoE, 1M context, ~$0.435/$0.87 per-Mtok at the April 2026 launch — near-frontier capability at ~1/6th the cost of Opus 4.7 / GPT-5.5 (DeepSeek V4). That price no longer holds: from 2026-08-16 the rate is $0.66/$1.98 off-peak and $1.32/$3.96 at peak (see the repricing section above). The launch figure is kept as the record of what was claimed, not as a current price.
  • Cost-efficiency engineering: 1M-token inference at ~27% of V3.2 FLOPs (Pro) and ~10% KV-cache, the basis for the pricing disruption (DeepSeek V4)
  • Conditional memory / Engram research: A new sparsity axis orthogonal to MoE, with a U-shaped allocation scaling law (scaled to 27B params) (Engram)
  • DSpark speculative decoding: semi-autoregressive drafting + confidence-scheduled verification; 60–85% faster per-user generation at matched throughput vs. the MTP-1 production baseline, measured under live traffic in the DeepSeek-V4 serving system (DSpark, arXiv:2607.05147)

Mentioned In

Related Entities

  • Google DeepMind — closed frontier peer (Gemini 3)
  • OpenAI — closed frontier peer (GPT-5.5)
  • Anthropic — closed frontier peer (Claude Opus 4.7)
  • Meta FAIR — other open-weight-leaning lab (Muse Spark / V-JEPA)
  • Moonshot AI — the other open-weight frontier lab; competes on capability-at-price rather than cost

Changelog

  • 2026-08-18 — Recorded the 2026-08-16 repricing and the 2026-08-13 V4-Pro GA. This is the first increase by this entity in the KB's record and it inverts the page's framing: DeepSeek is catalogued here as the cost-disruption case, and cost disruption is what stopped. Added the full before/after rate tables for V4-Pro and V4-Flash (primary, DeepSeek's own pricing page and changelog, read 2026-08-18), flagged cached input as the outlier at 6–12×, and amended — not deleted — the Key Contributions line that still quoted the April launch price as current. Also linked the newly published deepseek-harness, recorded on the harnesses rung.
  • 2026-07-22 — Compiled new source (dspark-confidence-scheduled-speculative-decoding): added the DSpark serving-stack section (semi-autoregressive drafting, confidence-scheduled verification, 60–85% production speedup vs. MTP-1), with the attribution basis and preprint/first-party evidence caveats stated explicitly.