DeepSeek

DeepSeek is a Chinese AI company whose open-weight R1 release is used in ^[src-061] as the reference point for a major shift in frontier-model competition.

Key facts

• The "DeepSeek moment" refers to DeepSeek R1's January 2025 release, which Fridman describes as an open-weight Chinese model that surprised the field with near state-of-the-art performance and allegedly much lower compute cost ^[src-061].
• Raschka says DeepSeek won goodwill among open-weight-model practitioners because it shared strong models openly ^[src-061].
• Lambert argues that DeepSeek kicked off a broader Chinese open-weight movement, with other labs such as Z.ai, MiniMax, and Kimi/Moonshot becoming more visible and sometimes leapfrogging DeepSeek's crown ^[src-061].
• The episode treats DeepSeek less as a one-off company story and more as evidence that open weights can reshape influence, research diffusion, and business-model pressure across the global AI market ^[src-061].
• DeepSeek-V3.2 is also discussed as an example of sparse attention and indexing for more efficient long-context processing ^[src-061].

Related entities

• Lex Fridman
• Sebastian Raschka
• Nathan Lambert
• Openai
• Anthropic
• Gemini

Related concepts

• Open Weight Model Strategy
• Model Lab Differentiation
• Agentic Context Management
• Training Inference Compute Balance

Source references

• ^[src-061] Lex Fridman – "State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490" (2026-01-31)

2026-06-29 DeepSpec, DSpark, and token-economics update

• DeepSeek's DeepSpec/DSpark release adds reusable inference-optimisation machinery to its open-model strategy: semi-autoregressive draft generation plus confidence-scheduled verification for speculative decoding ^[src-179].
• The DeepSeek-V4-Pro-DSpark Hugging Face artifact is explicitly not a new base model; it is the same checkpoint with an attached speculative-decoding module, which makes the systems contribution clearer ^[src-179].
• VentureBeat frames DeepSeek V4 pricing and cache-read economics as a threat to Western frontier labs' high-margin token business, especially for cache-heavy autonomous-agent workloads ^[src-180].

2026-07-10 AI Watch batch update

• New source evidence added: ^[src-214], ^[src-223], ^[src-225], ^[src-231]. See the paired source summaries for verified claims and evidence caveats.

Evidence checked on 14 September 2026

The model card describes a multimodal MoE with 552B backbone parameters, 8B/16B activation for prefill/decode and a one-million-token context. It includes additional Engram memory and compressed KV state. ^[src-469] 552B is backbone size, not the complete repository parameter total. Benchmark results depend on reasoning effort and harness; not every metric beats V4 Pro.

• DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression

A Hacker News post reproduces a notice about routing Pro requests to V4.1 Flash and changing Flash prices around 10 September. ^[src-470] The official live routing and price schedule were not independently verified. Retain as a dated community report, not an operational pricing reference.

• DeepSeek V4.1 Flash routing and pricing discussion