All AI models
Open-weightVerified October 23, 2026

DeepSeek

DeepSeek V4

DeepSeek V4 is a high-value option for teams that need a one-million-token context window, tool support, and unusually low published Flash pricing. Since October 17, 2026 the Flash weights are MIT-licensed, a permissive license for a model of this class, but serving it yourself is still a serious infrastructure project.

What changed: Since October 17, 2026, deepseek-flash is served by DeepSeek-V4.1-Flash-1017, which adds vision. The old deepseek-v4-flash names are retired but still accepted. Which model serves deepseek-v4-pro was not clear from DeepSeek's pages on October 23, 2026: the pricing page lists DeepSeek-V4-Pro-0813.
License
MIT (V4.1-Flash-1017 weights)
Context
1M tokens (384K max output)
Modalities
Text · Image (Flash)
Parameters
DeepSeek described V4.1-Flash as a 552B-parameter MoE (8B active for input, 16B for output) on September 10, 2026; the October 17 build's count is not stated in its model card

What the license means

DeepSeek published the DeepSeek-V4.1-Flash-1017 weights on Hugging Face under the MIT License on October 17, 2026. That covers the Flash model that now serves deepseek-flash. It does not tell you the terms for deepseek-v4-pro, so check the exact weight repository before self-hosting any other V4 variant.

Open-weight, open source, and no-cost API access mean different things. Always review the linked terms and your own data, hosting, and compliance requirements before shipping.

API and operating cost

Input / 1M
$0.15 to $0.30
Output / 1M
$0.60 to $1.20

DeepSeek-V4.1-Flash (deepseek-flash) cache-miss input and output per 1M tokens, off-peak to peak. Cache-hit input is $0.003 off-peak and $0.006 peak. deepseek-v4-pro costs $0.66 to $1.32 input and $1.98 to $3.96 output. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday. Prices rechecked on the official pricing page on October 23, 2026.

Check current provider price

Where it fits

  • Cost-sensitive API workflows
  • Long-context text and image agents
  • Tool-call and structured-output integrations

Where it does not fit

  • Assuming the Flash MIT license covers deepseek-v4-pro
  • Assuming Flash pricing applies to Pro or to peak-hour batch jobs

Local deployment reality

The V4.1-Flash-1017 weights are downloadable, but a model DeepSeek described at 552B total parameters in September is a multi-GPU, data-center deployment, not a laptop model. For most teams the hosted API is the practical path; self-host only when data control justifies the infrastructure.

Practical setup notes

  • Choose Flash when the Responses API matters and Pro only after benchmarking the task quality.
  • Model cache-hit and cache-miss input separately in cost estimates.

Capability notes

Reasoning
Thinking (default) and non-thinking modes
Tools and output
Tool calls, JSON output, Anthropic compatibility, and Responses API on Flash

Limitations to plan around

  • Published API pricing doubles during weekday peak hours, so schedule batch work off-peak.
  • DeepSeek's pages did not make clear which model serves deepseek-v4-pro as of October 23, 2026.

Continue comparing

Related model decisions

See all reviewed models
Open source

GLM-5.2

Long-horizon coding agents

Open-weight

MiniMax M3

Agentic coding

Open source

Mistral Small 4

Multimodal document analysis

Editorial note: This is a dated decision page, not a benchmark leaderboard or a substitute for legal, security, or infrastructure review. Model and pricing facts can change after October 23, 2026; use the linked primary sources before committing budget or customer data.Rechecked October 23, 2026: license, context window, output limit, API prices, and peak hours. Tool-support notes date from the August 11, 2026 audit.