MiniMax M3.1 Flash just went free on OpenRouter. Here's what it is.

#ai #llm #coding #opensource

I run DEV·TV, a TV-style browser app for developer content. One of its channels, NEW MODELS, watches OpenRouter and puts fresh releases on screen as they appear.

This evening it flashed this:

MiniMax M3.1 Flash (free), added 9 hours ago, 1M context, $0.

That's worth stopping the channel for. Until now, this model was basically locked inside MiniMax's own tools.

Where it came from

MiniMax shipped M3.1 Flash as a preview on September 27, 2026, with almost no fanfare. It showed up inside MiniMax Code before any announcement, and for weeks the only ways to use it were MiniMax Code or a MiniMax Token Plan subscription. No pay-as-you-go price, no public weights, no independent benchmarks.

A free listing on OpenRouter means anyone with an API key can now try it in a couple of lines.

What it is

A lighter sibling of MiniMax M3. M3 is MiniMax's flagship from June 2026, built on MiniMax Sparse Attention, which makes a 1M-token context far cheaper to run than the previous generation. Flash carries the same 1M window.
Aimed at everyday coding. MiniMax positions it for the routine queue: bug fixes, small features, iterative dev work, rather than the hardest problems.
Multimodal in, text out. Text, images and video go in; text comes out. Tool use and agentic workflows are supported.
MoE. It's reported at roughly 428B total parameters with about 23B active per token, which is how a model this size can be "Flash".

The one thing that's different: you can't turn thinking off

M3 lets you disable reasoning. M3.1 Flash doesn't. Thinking is always on, and you pick the effort instead:

low · medium · high · xhigh · max

The default is max. Try to switch thinking off and you get an error.

My take: for a model meant for daily coding, this is a sensible trade. You get a dial instead of a switch, so quick edits run at low and gnarly refactors at max, without changing models.

How fast is "Flash"?

MiniMax hasn't published speed numbers. Early users report decode in the 90–110 tokens/sec range. Treat that as anecdotal until there's a proper benchmark.

Try it

OpenRouter is OpenAI-compatible, so it's just a base URL swap:

from openai import OpenAI

client = OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key="YOUR_OPENROUTER_KEY",
)

resp = client.chat.completions.create(
    model="minimax/minimax-m3.1-flash:free",
    messages=[{"role": "user", "content": "Find the bug in this function: ..."}],
)
print(resp.choices[0].message.content)

Free tiers on OpenRouter are usually rate-limited, so it's for kicking the tyres, not production.

Why it matters

A 1M-context, multimodal, reasoning coding model at $0 is a lot of capability to play with. If it holds up on real repos, it's a strong free option for agents and coding assistants. If it doesn't, you've lost nothing finding out.

Next on DEV·TV's NEW MODELS: DeepSeek V4.1 Flash. It's a busy week for "Flash" models.

📺 Catch new models the day they drop: shouvik12.github.io/devtv

DEV Community