DeepSeek models
Each model has its own URL. Benchmark cells that had no source were removed or replaced with the figure from the linked model card.
- DeepSeek-V4.1-Flash — Ultra-Fast Asymmetric Causal MoE with 1M Context & Engram Memory
- DeepSeek-V4-Pro-0813 — Massive 1.7T Sparse MoE for Autonomous Software Engineering & Heavy Reasoning
- DeepSeek-R1 — Pioneering Open Reasoning Model Trained via Large-Scale RL & GRPO
- DeepSeek-V3 — The Efficiency Landmark: 671B MoE with Multi-Head Latent Attention
- Janus-Pro — Decoupled Vision-Language Foundation Model for Unified Understanding & Generation
Models & Architecture Lineage
Compare flagship models across DeepSeek-V4.1-Flash (Asymmetric MoE + Engram), V4-Pro (1.7T Sparse), R1 GRPO reasoning, V3 foundation baseline, and Janus-Pro multimodal vision.
Architectural Synthesis Note: DeepSeek-R2 Integration
Hardware & Architecture SpecTODO: not verified. The claim that a standalone R2 run was absorbed into V4/V4.1, including the Ascend 910C detail and a --effort low|medium|high CLI flag, is not in the current API pricing page or the V4.1-Flash model card. Do not treat this note as a specification.
DeepSeek-V4.1-Flash
V4.1-Flash (Sep 2026)September 2026Ultra-Fast Asymmetric Causal MoE with 1M Context & Engram Memory
DeepSeek-V4.1-Flash is a multimodal Mixture-of-Experts model with a 552B-parameter backbone, a Causal Encoder-Decoder that activates 8B parameters per token during prefill and 16B during decode, a 1,048,576-token context, Engram conditional memory (196B parameters), and a global KV cache of 890 bytes per token using FP4 (E2M1) caching. API model name: deepseek-flash.
Core Architecture & Hardware Specifications
Verified Performance & Research Benchmarks
Key Innovations & Capabilities
- Asymmetric causal routing: 8B active parameters during prefill, 16B active during generation across 6 of 384 fine-grained experts.
- Engram conditional memory: 196B parameters in constant-time O(1) hashed memory for persistent factual and codebase recall.
- FP4 E2M1 ultra-compact KV cache: Compresses key-value state to only 890 bytes per token, making 1M token contexts computationally negligible.
- The API supports thinking mode (on by default) and reasoning_effort values low, high, and max. The model card also describes a continuous effort setting from 1 to 100.
- A single tokens-per-second throughput figure is not stated on the model card or the API pricing page, so none is listed here.
DeepSeek-V4-Pro-0813
V4-Pro (Aug 2026)August 2026Massive 1.7T Sparse MoE for Autonomous Software Engineering & Heavy Reasoning
DeepSeek-V4-Pro is the 1.6T-parameter Mixture-of-Experts model (49B activated) with a 1,000,000-token context, published on Hugging Face as DeepSeek-V4-Pro. The API serves checkpoint DeepSeek-V4-Pro-0813 as model name deepseek-v4-pro. The V4.1 model card's comparison table reports this checkpoint's benchmark scores; a separate V4.1-Pro had not launched when the 10 September 2026 API note was published.
Core Architecture & Hardware Specifications
Verified Performance & Research Benchmarks
Key Innovations & Capabilities
- 1.6 trillion total parameters and 49 billion activated parameters, from the DeepSeek-V4-Pro model card.
- API context length 1M tokens and maximum output 384K tokens, from the Models & Pricing page.
- Vision is not supported on deepseek-v4-pro. V4.1-Flash is the API model that lists vision.
- Off-peak API rates are half of peak rates. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday–Friday, excluding Chinese public holidays.
- An SWE-bench Verified score of 80.6% was removed: it is not in the V4.1-Flash model card comparison table or the V4-Pro model card excerpt used for this page.
DeepSeek-R1
R1 (Jan 2025 Reference)January 2025Pioneering Open Reasoning Model Trained via Large-Scale RL & GRPO
DeepSeek-V3
V3 (Dec 2024 Baseline)December 2024The Efficiency Landmark: 671B MoE with Multi-Head Latent Attention
Janus-Pro
Janus-Pro (Multimodal)February 2025Decoupled Vision-Language Foundation Model for Unified Understanding & Generation
Comprehensive Frontier Benchmark Matrix
| Model Architecture | AIME Olympiad | SWE-bench Verified | Codeforces Elo | Terminal-Bench 2.1 | Key Milestone |
|---|---|---|---|---|---|
DeepSeek-V4.1-Flash (Sep 2026)Speed & Elo #1 | Not on the V4.1 model card (AIME 2026 96.7% removed) | Not published on the V4.1 model card | 3,471 (model card) | 90.6% Pass@1 (Harness Minimal) | GPQA Diamond 90.9% · HLE with tools 63.9% |
DeepSeek-V4-Pro-0813 (Aug 2026)SWE 80.6% | Not in the sourced comparison table | Removed (80.6% was unsourced) | 3,348 (V4.1 card, V4-Pro column) | 87.9% Pass@1 | 1.6T / 49B active · GPQA Diamond 92.4% |
DeepSeek-R1 (Jan 2025) | 79.8% AIME 2024 (R1 README) | Removed (49.2% was unsourced) | 2,029 rating / 96.3 percentile (R1 README) | Removed (68.5% was unsourced) | MATH-500 97.3% (R1 README) |
DeepSeek-V3 (Dec 2024) | Removed (39.2% was unsourced on this page) | Removed (42.0% was unsourced) | Removed (1,740 was unsourced) | Removed (54.1% was unsourced) | MMLU 88.5 · HumanEval-Mul 82.6 (V3 README) |
Janus-Pro (Feb 2025) | N/A (Vision) | N/A | N/A | N/A | MMBench: 85.2% · GenEval: 0.81 |