DeepSeek

Efficiency-first open frontier models, with sparse and compressed attention for 1M-token context at a fraction of compute, MoE at 1.6T parameters, aggressive pricing and optimization for Chinese chips. A small, low-hierarchy research team with no KPIs, per founder.

Founded: July 2023
Headquarters: Hangzhou, China
Founders: Liang Wenfeng
Pushes against: Spending-driven frontier scaling; closed weights.
Now: V4 matches mid-2026 frontier-minus-one models on company-reported benchmarks. Closed a first ~$7B round (2026-06), then paused a second after leaked founder remarks that China still lags and depends on Nvidia (Bloomberg, 2026-07). IPO prep reported.

Key people

Liang Wenfeng (founder, CEO)

Funding

6 October 2026 reported
Amount
Reported at least 80 billion yuan, about $12 billion, possibly up to 100 billion yuan, about $15 billion
Valuation
About 500 billion yuan, about $74 to 75 billion, possibly higher
Lead
CATL and Tencent with the largest commitments, plus Geely Auto, Monolith Management and Loyal Valley Capital
Source americanbazaaronline.com, livemint.com, rswebsols.com
July 2026 reported
Amount
Second round, paused
Valuation
~$71B pre-money (reported)
Source finance.yahoo.com
June 2026 reported
Amount
~50B yuan (~$7.4B), first outside round
Valuation
$52B-$59B post-money (reported)
Lead
Founder Liang (20B yuan), Tencent, CATL; national AI fund and others (reported)
Source cnbc.com

Every funding round in the atlas

Latest launches

DeepEP-Ascend and DeepGEMM-Ascend 29 September 2026
DeepSeek ports its MoE communication and FP8 GEMM libraries to Huawei Ascend NPUs, targeting Ascend 950 with CANN 9.2.0.
DeepSeek Elastic Compute (DSec) sandbox platform 19 September 2026
DeepSeek describes the sandbox platform behind its agentic RL: about 3 million sandboxes a day, 380,000+ concurrent, 5,000+ created per second.
DeepSelect TopK kernels and DeepJIT runtime 10 September 2026
DeepSeek opens its sparse-attention TopK kernel (2-20x faster than torch.topk) and a lightweight CUDA and Ascend JIT runtime.
DeepSeek-V4.1-Flash 10 September 2026
Native-multimodal 552B-backbone model with a causal encoder-decoder (8B active in prefill, 16B in decode) and a KV cache of 890 bytes per token.
DeepSeek-V4-Flash-Vision-Exp 21 August 2026
First multimodal V4 model, an experimental vision add-on to V4-Flash that matches its text ability; ships with a new Files API.
DeepSeek Harness (dsh) 13 August 2026
DeepSeek's own open-source agent harness ('Everything is a Plugin') with web, desktop and SSH UI, plugins, subagents and MCP. It is a developer preview under MIT.
DeepSeek-V4-Pro-0813 (general availability) 13 August 2026
V4-Pro goes GA with large agentic gains (Terminal Bench 2.1: 87.9, DeepSWE 62.7), low/high/max effort, Responses API and 50% off-peak pricing.
DeepSeek-V4-Flash-0731 (official) 31 July 2026
Official V4-Flash, with agent benchmarks up sharply versus the preview (DeepSWE 12.8 to 54.4, Terminal Bench 2.1 72.1 to 82.7) and a native Responses API for Codex.

Launches by year

67 launches and papers since 1 November 2023, oldest first within each year.

2026 16: DeepSeek-R1 paper v2 (expanded on arXiv), Engram: Conditional Memory via Scalable Lookup, Engram: conditional memory via scalable lookup, DeepSeek-OCR 2 (Visual Causal Flow), TileKernels (TileLang kernel library), DeepSeek-V4 (CSA/HCA attention, mHC, Muon), DeepSeek-V4-Pro and V4-Flash (preview), DSpark speculative decoding and DeepSpec, DeepSeek-V4-Flash-0731 (official), DeepSeek-V4-Pro-0813 (general availability), DeepSeek Harness (dsh), DeepSeek-V4-Flash-Vision-Exp, DeepSeek-V4.1-Flash, DeepSelect TopK kernels and DeepJIT runtime, DeepSeek Elastic Compute (DSec) sandbox platform, DeepEP-Ascend and DeepGEMM-Ascend
2025 28: DeepSeek-R1 and R1-Zero, DeepSeek-R1: Incentivizing Reasoning via RL (arXiv), Janus-Pro (1B, 7B), DeepSeek chat app tops US App Store, Native Sparse Attention (NSA), Native Sparse Attention (NSA), Open Source Week (FlashMLA, DeepEP, DeepGEMM, DualPipe, 3FS), DeepSeek-V3/R1 inference system overview ('one more thing'), DeepSeek-V3-0324, DeepSeek-GRM / SPCT (inference-time scaling for reward models), DeepSeek-Prover-V2 (7B, 671B), Insights into DeepSeek-V3: hardware co-design paper, DeepSeek-R1-0528, DeepSeek-R1-0528-Qwen3-8B, DeepSeek-V3.1, DeepSeek-R1 in Nature (peer-reviewed), DeepSeek-R1 published in Nature, DeepSeek-V3.1-Terminus, DeepSeek Sparse Attention (V3.2-Exp), DeepSeek-V3.2-Exp (DeepSeek Sparse Attention), DeepSeek-OCR (Contexts Optical Compression), DeepSeek-OCR (contexts optical compression), LPLB: linear-programming expert-parallel load balancer, DeepSeekMath-V2, DeepSeek-V3.2 and V3.2-Speciale, DeepSeek-V3.2 (DSA, scaled RL, agentic synthesis), mHC: Manifold-Constrained Hyper-Connections, mHC: Manifold-Constrained Hyper-Connections
2024 21: DeepSeekMoE, DeepSeekMoE, DeepSeekMath (introduces GRPO), DeepSeekMath 7B (introduces GRPO), DeepSeek-VL (1.3B, 7B), DeepSeek-V2, DeepSeek-V2 (Multi-head Latent Attention), DeepSeek-Prover (V1), DeepSeek-Coder-V2, ESFT: Expert-Specialized Fine-Tuning for MoE, API context caching on disk, DeepSeek-Prover-V1.5, Fire-Flyer AI-HPC: cost-effective cluster co-design, Auxiliary-loss-free load balancing for MoE, DeepSeek-V2.5, Janus (and JanusFlow), DeepSeek-R1-Lite-Preview, DeepSeek-V2.5-1210, DeepSeek-VL2, DeepSeek-V3, DeepSeek-V3 Technical Report
2023 2: DeepSeek Coder (1.3B-33B), DeepSeek LLM 7B / 67B

People moves

Of the moves the atlas tracks, 0 joined and 1 left.

Luo Fuli 12 November 2025
DeepSeek to Xiaomi
Core developer, DeepSeek

Sources

This profile was checked against its sources on 6 October 2026. How we check