MiniMax-M3 is a native multimodal model with 1M context running ~428B parameters and ~23B activated parameters. It brings native multimodality. enabling deeper semantic fusion across text, image, and video. M3 also introduces MiniMax Sparse Attention (MSA) to improve long context efficiency, achieving frontier-level performance across long-horizon agentic benchmarks, excelling in both coding and cowork.
Fine-tuningDocs | Minimax M3 can be customized with your data to improve responses. Fireworks uses LoRA to efficiently train and deploy your personalized model |
ServerlessDocs | Minimax M3 is available via Fireworks' serverless API, where you pay per token. There are several ways to call the Fireworks API, including Fireworks' Python client, the REST API, or OpenAI's Python client. |
On-demand DeploymentDocs | On-demand deployments allow you to use Minimax M3 on dedicated GPUs with Fireworks' high-performance serving stack with high reliability and no rate limits. |
Run queries immediately, pay only for usage
A natively multimodal Mixture-of-Experts model from MiniMax, released 2026-06-01. It was trained on mixed text, image, and video data from the first step and uses MiniMax Sparse Attention.
Agentic coding and software engineering: MiniMax reports 59.0% on SWE-Bench Pro, 66.0% on Terminal-Bench 2.1, and 74.2% on MCP Atlas. Also long-horizon agent workflows and multimodal understanding of images and video.
1,048,576 tokens, set by max_position_embeddings in the published config. MiniMax documents up to 1M tokens with a guaranteed minimum of 512K.
Fireworks serves a 512K token context window, below the model's native 1,048,576 tokens, on both the serverless endpoint and on-demand deployments.
Fireworks applies no platform-wide default. When temperature is unspecified it uses the model's generation_config.json, which sets temperature to 1.0 and top_p to 0.95.
Fireworks defaults to 2,048 output tokens and supports generation up to the full context window. If prompt plus max_tokens exceeds the context length, max_tokens is lowered to fit.
Yes to both. Streaming is set with stream, and function calling is supported, with tool call arguments streaming incrementally alongside the response.
428 billion total parameters with roughly 23 billion active per token, across 128 routed experts with 4 selected per token plus 1 shared expert.
Yes. Fireworks supports LoRA fine-tuning for Minimax M3.
Fireworks bills input, cached input, and output tokens separately at $0.30, $0.059, and $1.20 per 1M tokens, reporting them in the usage object and the fireworks-prompt-tokens response headers. Image inputs are tokenized by the model's own tokenizer.
Serverless ceilings for this model's size tier default to 43.2M total prompt TPM, 10.8M uncached prompt TPM, and 432k generated TPM, adaptive per account and model. On-demand deployments are not rate limited.
Fireworks announces serverless model deprecations in advance, following its serverless model lifecycle policy. For long-term version stability it recommends on-demand deployments.
Released under the MiniMax Community License: free for non-commercial use. Commercial use requires a Built with MiniMax M3 notice and notifying MiniMax, or written authorization above 20 million US dollars in yearly revenue.
No. Fireworks operates zero data retention by default and does not log prompt or generation data for open models without opt-in. The Response API is the exception, retaining data 30 days when store is true.
