DeepSeek

Chinese AI lab and model family. Originally a research arm of High-Flyer Quant; spun out as a standalone lab around 2023. Sits in the small cluster of labs (alongside Mistral, Qwen, Yi) that ship competitive open-weights models, and has the strongest published reasoning lineage of the Chinese labs through the R1 / V3 / V4 series.

The lab matters in the LLM landscape for a few reasons distinct from raw capability:

• Open weights for frontier-class reasoning. R1 (Jan 2025) was the first open-weights model to land near the top of reasoning benchmarks, with weights and technical report public. V4 continues that pattern.
• Training-cost reporting that flipped the discourse. The published training compute and budget numbers for V3/R1 were small enough relative to GPT-4-class assumptions that they reframed the cost discussion across the industry.
• MoE architecture choices that pushed the routing-and-expert-load-balancing literature forward, picked up downstream by other labs.

Model line as relevant to this vault

• V3 — Mid-2024 dense+MoE base model. The published training report (sub-$6M figure for the final pretraining run) is the source most often cited in cost-collapse arguments.
• R1 — Jan 2025 reasoning model derived from V3. Open weights. Long visible <think> traces by default, distillation-friendly.
• V4 — Current generation as of 2026. Has reasoning variants (deepseek-v4-flash, deepseek-v4-pro) plus an "Expert Mode" surface in the official app/web. See deepseek-v4-roleplay-instruct for one documented quirk of V4's training-injected style markers and controllable-thinking-style for what that implies about how its CoT data was built.

Did R1 distil o1?

Whether R1 was built on OpenAI o1's outputs was argued from the week it shipped, and the answer has moved. In April 2025 Nathan Lambert wrote confidently that DeepSeek did not distil o1. By September 2026 he had walked that back in open-source-ai-reading-list: once stealing-reasoning-traces-from-proprietary-llm-apis showed that frontier APIs leaked hidden reasoning traces, and anthropic-threat-report-september-2026 reported Chinese labs using the technique, he called it more possible than he had allowed that DeepSeek used some o1 traces to make R1 easier to train. There is still no direct evidence that it did, and he does not think it erases the work. llm-distillation covers the wider fight.

The same list uses V4 Flash as the example of open models sitting on the cost frontier without holding the performance frontier. deepseek-v4-flash-0731-artificial-analysis has the 0731 checkpoint (284B total, 13B active) scoring 50 on the Artificial Analysis Intelligence Index at about 60% less per task than GPT-5.6 Luna, mostly thanks to DeepSeek's cache-hit discount on its own API. At publication that checkpoint was API-only, with weights "expected in the coming weeks", so it stands in for an open model on the assumption that they shipped. open-closed-model-gap puts it in context; on Håvard Tveit Ihle's measure the gap between open and closed models was at its smallest around R1's release and has widened since.

Two more data points from the same wave. Anthropic's February 2026 accusation attributed about 150K exchanges to DeepSeek, which Lambert called negligible next to Moonshot's and MiniMax's (how-much-does-distillation-matter-for-chinese-llms). Anthropic's September report goes much further, saying DeepSeek routed Claude into its own products without telling users and extracted hidden reasoning with the cross-session replay attack (anthropic-threat-report-september-2026) — claims the report gives no way to verify. And Ai2's olmo-3-technical-report sets its own wall-clock training cost against V3's $5.576M figure, the number the ATOM memo calls misleading (atom-project-american-truly-open-models). R1 is also what persuaded z-ai to open its flagship GLM weights.

Agent ecosystem

DeepSeek maintains awesome-deepseek-agent — a curated list of 16 AI agent and coding-assistant tools (Claude Code, Pi, OpenCode, Crush, Reasonix, AstrBot, OpenClaw, etc.) with per-tool guides for using V4-Pro / V4-Flash. The list is the canonical signal of which tools DeepSeek considers first-class for V4, including both Western terminal coding agents and China-facing chat-platform agents (Feishu, WeChat, Telegram).

Things to track

• Future ingests of DeepSeek technical reports (V4 if/when published)
• Distilled / quantized derivatives of V4 from the open-source community
• Other reverse-engineered training-format quirks similar to the roleplay markers
• Cost-curve comparisons (DeepSeek vs OpenAI / Anthropic per-token cost)
• New entries on awesome-deepseek-agent worth a closer look

Related pages

• deepseek-v4-roleplay-instruct — undocumented <think>-style control markers in V4
• controllable-thinking-style — broader pattern these markers exemplify
• awesome-deepseek-agent — DeepSeek's curated list of 16 agent / coding-assistant integrations
• kimi-k3 — V4's mHC and K3's attention residuals are two different attacks on the residual path: mHC widens it, attention residuals weight cross-layer connections by an attention score