DeepSeek models
Each model has its own URL. DeepSeek-V4.1-Flash is served as DeepSeek-V4.1-Flash-1017 on the API name deepseek-flash. Benchmark figures link to official sources where this directory publishes them. Model name migration guide and off-peak pricing guide and LLM API pricing comparison. For third-party benchmarks against Claude, Gemini, GPT and Kimi, see DeepSeek vs frontier models.
Model comparison (context, parameters, API pricing)
API prices apply only to models on the current DeepSeek price list. Other rows link to official model cards.
- 1M
- 552B total · 16B active
- Asymmetric Causal MoE (6 of 384 experts)
- $0.30 / $1.20
- 1M
- 1.6T total · 49B active
- 1.6T Sparse MoE (49B activated) with hybrid CSA
- $1.32 / $3.96
| Model | Context | Params (active) | Architecture | API price (in/out per 1M) | Source |
|---|---|---|---|---|---|
| DeepSeek-V4.1-Flash | 1M | 552B total · 16B active | Asymmetric Causal MoE (6 of 384 experts) | $0.30 / $1.20 | Docs |
| DeepSeek-V4-Pro-0813 | 1M | 1.6T total · 49B active | 1.6T Sparse MoE (49B activated) with hybrid CSA | $1.32 / $3.96 | Docs |
| DeepSeek-R1 | 128K | 671B total · 37B active | DeepSeekMoE | — | Model card |
| DeepSeek-V3 | 128K | 671B total · 37B active | DeepSeekMoE (256 routed | — | Model card |
| Janus-Pro | — | 1.5B total · dense | SigLIP-L understanding | — | Model card |
V4.1 Flash vs V4 Pro
Figures below come from this directory's model data, which mirrors the official Models & Pricing table and model cards. API model names.
| Topic | deepseek-flash (V4.1-Flash-1017) | deepseek-v4-pro (V4-Pro-0813) |
|---|---|---|
| Peak price (cache-miss in / out per 1M) | $0.30 / $1.20 | $1.32 / $3.96 |
| Context / max output | 1M / 384K max | 1M / 384K max |
| Backbone (active params) | 552B (8B prefill / 16B decode) | 1.6T (49B active) |
| Vision on API | Supported | Not supported |
| Best fit (from model pages) | Fast agentic coding, multimodal, lower list price | Heavier reasoning, largest Pro checkpoint |
API list prices & model specs (official)
Peak and off-peak USD per 1M tokens. Off-peak is half of peak; peak hours are listed on the pricing page.
Models & architecture lineage
Compare DeepSeek-V4.1-Flash, V4-Pro (1.6T), R1, V3, and Janus-Pro. Figures link to official sources where available.
DeepSeek-V4.1-Flash
Ultra-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. On 17 October 2026 DeepSeek upgraded the served weights to DeepSeek-V4.1-Flash-1017; the API model name is still deepseek-flash and token list prices are unchanged.
Core architecture & hardware specifications
Peak cache miss $0.30 / off-peak $0.15 per 1M input. Peak cache hit $0.006 / off-peak $0.003.
Peak $1.20 / off-peak $0.60 per 1M output.
Price sourceSourced 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
1.6T 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
Peak cache miss $1.32 / off-peak $0.66 per 1M input. Peak cache hit $0.044 / off-peak $0.022.
Peak $3.96 / off-peak $1.98 per 1M output.
Price sourceSourced 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.
- SWE-bench Verified is not listed in the V4.1-Flash model card comparison table used for sourced figures on this page.
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
Unified multimodal understanding and generation (decoupled vision encoding)
Comprehensive frontier benchmark matrix
- —
- —
- 3,471
- 90.6%
- GPQA Diamond 90.9% · HLE with tools 63.9%
- Not in the sourced comparison table
- —
- 3,348
- 87.9%
- 1.6T / 49B active · GPQA Diamond 92.4%
- 79.8%
- —
- 2,029
- —
- MATH-500 97.3% (R1 README)
- —
- —
- —
- —
- MMLU 88.5 · HumanEval-Mul 82.6 · GSM8K 89.3 (V3 README)
- —
- —
- —
- —
- See Hugging Face model card (architecture only)
| Model | AIME | SWE-bench | Codeforces | Term.-bench | Key |
|---|---|---|---|---|---|
| V4.1-Flash-1017 | — | — | 3,471 | 90.6% | GPQA Diamond 90.9% · HLE with tools 63.9% |
| V4-Pro-0813 | Not in the sourced comparison table | — | 3,348 | 87.9% | 1.6T / 49B active · GPQA Diamond 92.4% |
| R1 | 79.8% | — | 2,029 | — | MATH-500 97.3% (R1 README) |
| V3 | — | — | — | — | MMLU 88.5 · HumanEval-Mul 82.6 · GSM8K 89.3 (V3 README) |
| Janus-Pro | — | — | — | — | See Hugging Face model card (architecture only) |