← All models

MiniMax-M1-40k

Published Publisher MiniMax

Model positioning

MiniMax-M1-40k is a large-scale hybrid-attention reasoning model in the MiniMax M1 series. It uses a mixture-of-experts (MoE) and Lightning Attention architecture and is designed for complex tasks requiring long contexts and deep reasoning. The official model card states that this version was trained with a thinking budget of 40K tokens; M1 natively supports contexts of up to 1 million tokens. (huggingface.co)

Capabilities and uses

  • Can be used for long-text understanding, mathematical reasoning, software engineering, and agentic tool use; the model supports function calling. (huggingface.co)
  • The officially reported score for MiniMax-M1-40k on SWE-bench Verified is 55.6%; this result used the officially specified evaluation procedure and 486 tasks that could run on its infrastructure, and should not be regarded as a performance guarantee in all scenarios. (huggingface.co)
  • The model weights have been officially released; the model card lists the license as Apache 2.0 and provides deployment guidance for vLLM and Transformers. (huggingface.co)

Limitations

40K is the officially described thinking budget and should not be treated directly as a uniform maximum output limit for all deployments or APIs. The model may produce errors, and tool calls, actual context and output limits also depend on deployment and service configuration; important results need to be verified. The official announcement states that the 80k version outperforms the 40k version on most benchmarks. (minimax.io)

Pricing

The following are historical M1 series API prices from the official websites in the announcement published on June 16, 2025. The announcement did not explicitly distinguish API prices for the 40k version; these prices cannot be considered current or taken as a guarantee that this version is currently available. The prices on the two sites correspond to services in different regions and are not currency conversions. Date checked: November 18, 2026.

Domestic site: Chinese yuan (CNY)

Tiered by input length; all prices are per million tokens:

Input length Input Output
0–32K ¥0.8 ¥8
32K–128K ¥1.2 ¥16
128K–1M ¥2.4 ¥24

Source: MiniMax domestic site announcement. (minimax.cn)

International site: US dollars (USD)

Tiered by input length; all prices are per million tokens:

Input length Input Output
0–200K $0.4 $2.2
200K–1M $1.3 $2.2

Source: MiniMax international site announcement. (minimax.io)

As of the date checked, the domestic pricing overview and international pricing overview consulted did not list MiniMax-M1; current pricing and service availability for the 40k version have not been verified. (platform.minimax.io)

Release and official sources