MiniMax-M1-80k

MiniMax-M1-80K is the most enhanced version of the base model series. It retains the same 456 billion total parameters , with 45.9 billion activated per token , but has been further trained on 7.5 trillion tokens of diverse tasks, including mathematical reasoning , competitive programming, logical reasoning using the SynLogic framework , and real-world software engineering tasks in sandboxed environments based on SWE-bench. Compared to MiniMax-M1-40K , its reasoning budget has been increased from 40,000 to 80,000 tokens .

The model utilizes an enhanced version of the CISPO algorithm , combined with a progressive window expansion strategy . This approach ensures training stability at each stage and enables the model to adapt to increasing task complexity. An important innovation is the Lightning Attention mechanism , which represents a key scientific breakthrough by MiniMax in the field of large language models. It significantly reduces resource consumption during both training and inference — MiniMax-M1-80K consumes only 25% of the FLOPs required by DeepSeek R1 when generating 100,000 tokens , achieving unprecedented efficiency in long-sequence processing. Full reinforcement learning training was completed in just three weeks using 512 H800 GPUs .

MiniMax-M1-80K demonstrates outstanding performance across multiple benchmarks , achieving 86.0% on AIME 2024 and 56.0% on SWE-bench Verified . Particularly impressive are its capabilities in agent-based systems utilizing tool calling , where it scores 62.0% on TAU-bench airline and 63.5% on TAU-bench retail , even outperforming Gemini 2.5 Pro in certain scenarios.These results confirm its suitability for a wide range of research applications and commercial products . Furthermore, the model is supported by modern deployment frameworks , including vLLM and Transformers , and is available under the Apache-2.0 license , making it fully open and unrestricted for use .

Announce Date: 16.06.2025
Parameters: 456B
Experts: 32
Activated: 45.9B
Context: 10240K
Attention Type: Full Attention
VRAM requirements: 289.9 GB using 4 bits quantization
Developer: MiniMax-AI
Transformers Version: 4.45.2
License: Apache 2.0

Use our pre-built public endpoints to test inference and explore MiniMax-M1-80k capabilities.

Rent your own physically dedicated instance with hourly or long-term monthly billing.

We recommend deploying private instances in the following scenarios:

• maximize endpoint performance,
• enable full context for long sequences,
• ensure top-tier security for data processing in an isolated, dedicated environment,
• use custom weights, such as fine-tuned models or LoRA adapters.

Contact our dedicated neural networks support team at nn@immers.cloud or send your request to the sales department at sale@immers.cloud.