# MiniMax M2.5

> large mixture-of-experts language model optimized for coding, agents, and long-context reasoning

**Wikidata**: [Q138297657](https://www.wikidata.org/wiki/Q138297657)  
**Source**: https://4ort.xyz/entity/minimax-m2-5

## Summary
MiniMax M2.5 is a 229-billion-parameter mixture-of-experts (MoE) language model released by MiniMax on 12 Feb 2026 that activates only 10 billion parameters per token and is specialized for coding, agent workflows, and long-context reasoning. Its sparse MoE design lets it scale compute efficiently while keeping inference costs low.

## Key Facts
- Released on 12 Feb 2026; model card posted 14 Feb 2026
- Developer: MiniMax (https://www.minimax.io/news/minimax-m25)
- Total parameters: 229 billion; activated parameters per token: 10 billion
- Architecture: transformer-based mixture-of-experts model
- Optimization targets: coding, agent tasks, long-context reasoning
- Public weights available on Hugging Face: MiniMaxAI/MiniMax-M2.5

## FAQs
### Q: How many parameters does MiniMax M2.5 actually use per forward pass?
A: Only 10 billion of its 229 billion total parameters are activated for any single token, keeping compute and memory usage low while retaining model capacity.

### Q: Is MiniMax M2.5 open-weights?
A: Yes; the full checkpoint is downloadable from the Hugging Face repo MiniMaxAI/MiniMax-M2.5.

### Q: What tasks is the model tuned for?
A: MiniMax optimized M2.5 specifically for code generation, autonomous agent orchestration, and reasoning over long input contexts.

### Q: How does the mixture-of-experts design help?
A: MoE layers route each token to a small subset of feed-forward “expert” networks, so only a fraction of weights are used at once, cutting cost and latency versus dense models of the same size.

## Why It Matters
Large language models keep growing, but running 200-billion-plus-parameter dense transformers is expensive and slow. MiniMax M2.5 shows how mixture-of-experts sparsity can deliver comparable capability with far less active compute. By open-sourcing a 229 B-parameter checkpoint that only lights up 10 B weights per token, MiniMax gives researchers and product builders a practical route to high-end coding, agent, and long-context performance without the price tag of always-on dense parameters. The release also adds a new, competitive data point to the open-weights landscape, widening options beyond Llama, Qwen, and DeepSeek families.

## Notable For
- One of the first open-weights MoE models to exceed 200 B total parameters while keeping < 5 % active
- Explicitly optimized for three demanding use-cases: code synthesis, agent tool-calling, and long-document reasoning
- Released with full model weights on Hugging Face under permissive terms
- Activated parameter count (10 B) is small enough to fit on a single high-end GPU, lowering hardware barriers

## Body
### Architecture
MiniMax M2.5 follows a standard transformer stack augmented with MoE layers. In each MoE layer:
- 229 billion total parameters are split among many feed-forward “expert” sub-networks
- A learned router selects a tiny subset (exact count not disclosed) of experts for every token
- Only 10 billion parameters are exercised in the forward pass

### Training & Release
- Publication date: 12 Feb 2026
- Model card and weights appeared on Hugging Face two days later
- Training details (data mix, compute budget, duration) have not been published as of 14 Feb 2026

### Intended Use
MiniMax documentation highlights three target workloads:
1. Code completion and generation across common programming languages
2. Agent frameworks that require multi-step tool use and reasoning
3. Long-context tasks such as document Q&A and summarization beyond the 32 k-token range

### Availability
- Checkpoint: MiniMaxAI/MiniMax-M2.5 on Hugging Face
- License: permissive open-weights (exact license text not specified in source)
- No accompanying paper or blog post as of 14 Feb 2026

## References

1. [Source](https://blog.galaxy.ai/model/minimax-m2-5)
2. [Source](https://huggingface.co/MiniMaxAI/MiniMax-M2.5)
3. [Source](https://artificialanalysis.ai/models/minimax-m2-5?)