DeepSeek V4.1 Flash 1017 (Max)

What it claims

Artificial Analysis rates DeepSeek V4.1 Flash 1017 (Max) at 45 on its Intelligence Index, against a median of 19. That median comes from open-weights models in the Large class (over 150B parameters). DeepSeek released the model on October 17, 2026. The page calls it fast but very verbose. A model that is fast per token but long-winded can still be slow and costly per task, and the token-use figures below show the size of that tradeoff.

Evidence

All speed, latency and price figures are measured on DeepSeek's own API. Medians are for open-weights models of similar size.

• Output speed: 216.5 tokens per second, against a median of 77.4 t/s.
• Time to first token: 1.11 s, against a median of 1.70 s.
• Token use: the model generated 210M output tokens across the Intelligence Index, against a median of 140M. That is about 50% more than a typical comparable model.
• Cost per task: about $0.26 on average across the Intelligence Index.
• List prices: $0.30 input and $1.20 output per 1M tokens, against medians of $0.43 and $1.68.

The page gives a blended rate of $0.18 per 1M tokens. It assumes a 7:2:1 ratio of cache-hit, input and output tokens. Workloads with fewer cache hits or more output will pay more than that.

The page lists many other benchmarks, including AA-Omniscience (hallucination), long-context reasoning, agentic coding, terminal use and visual reasoning. The fetched text shows no scores for any of them.

How to use it

• Weights: open and downloadable for self-hosting under the MIT licence, which allows commercial use.
• Architecture: Mixture of Experts with 552B total parameters and 16B active per token. Self-hosting means holding all 552B parameters in memory, even though per-token compute scales with the 16B active.
• Hosted access: 10 API providers. Prices vary by provider, so the DeepSeek API figures above may not apply elsewhere.

Limits

• The "(Max)" label suggests a maximum reasoning-effort setting. The page does not define it or report lower-effort variants.
• Verbosity raises output-token spend and wall-clock time per task, partly offsetting the low per-token price and high speed.
• Speed and latency come from a single provider, DeepSeek's API.
• Per-benchmark results do not appear in the page text, so strengths and weaknesses by task type cannot be checked from this source.