MiniMax M3: A Small Model With a New Attention Trick
The second big open-weight release of the day turned out to be tiny by frontier standards — 428B parameters, only 23B active — but its headline is architectural: a new sparse attention that beats GQA where it hurts most, long contexts.
TL;DR
- 🪶 428B total / 23B active — small next to the trillion-param crowd, which makes it cheaper to serve.
- 🧩 New MiniMax Sparse Attention (MSA) — reported notably more efficient than GQA on large contexts.
- 📜 Open weights, released the same day as another open giant.
- 📈 Sparse attention is the lever for cheap long-context inference — exactly what all-day agents need.
What shipped
MiniMax M3 carries just 428 billion parameters with 23B active per token — compact compared to the trillion-parameter releases it shares the calendar with. The main innovation is another flavor of sparse attention, dubbed MSA (MiniMax Sparse Attention), reported to be markedly more efficient than GQA (grouped-query attention) on large contexts.
| Spec | MiniMax M3 |
|---|---|
| Total parameters | 428B (MoE) |
| Active parameters | 23B / token |
| Attention | MSA — MiniMax Sparse Attention |
| Efficiency claim | Beats GQA on large contexts |
| Weights | Open |
Why a new attention is a big deal
Attention cost is what makes long contexts expensive. GQA was the standard trick for taming it; a sparse variant that's meaningfully cheaper at large context lengths means you can feed bigger documents and longer agent histories without the bill exploding. With only 23B active parameters on top of that, M3 is positioned as an unusually cheap-to-serve open model — the kind of efficiency that matters when an agent is reading long files all day.
Run it on MegaBrain
MegaBrain routes 500+ models behind one OpenAI-compatible endpoint, so trying a new open release is a one-line model swap:
import openai
client = openai.OpenAI(
base_url="https://getmegabrain.com/api/gateway/v1",
api_key="mb-your-key-here",
)
response = client.chat.completions.create(
model="minimax/minimax-m3",
messages=[{"role": "user", "content": "Summarize this 200-page spec and list the risks"}],
)Confirm the exact model ID and live pricing on the models page.
Source
Release spotted via @ai_newz.
MegaBrain Gateway
500+ models. One API. No markup.
Use in Claude Code, Cline, Cursor, or any coding agent.
Newsletter
Stay in the loop
Get the latest model comparisons and guides — no spam, unsubscribe anytime.