MiniMax M3: First Open-Weights Model to Combine Three Frontier Capabilities

MiniMax (official) (@MiniMax_AI): "Introducing MiniMax M3: The First Open-Weights Model to Combine Three Frontier Capabilities

• Coding & Agentic Frontier: 59.0% SWE-Bench Pro, 66.0% Terminal Bench 2.1, 34.8% SWE-fficiency, 28.8% KernelBench Hard, 74.2% MCP Atlas
• MiniMax Sparse Attention scales context to 1M
• Natively Multimodal from Step Zero

API: http://platform.minimax.io Token Plan: https://platform.minimax.io/subscribe/token-plan 🚀New! MiniMax Code: http://code.minimax.io

Weights & Tech Report in ~10 Days" | XCancel

@MiniMax_AI

5h

Introducing MiniMax M3: The First Open-Weights Model to Combine Three Frontier Capabilities

• Coding & Agentic Frontier: 59.0% SWE-Bench Pro, 66.0% Terminal Bench 2.1, 34.8% SWE-fficiency, 28.8% KernelBench Hard, 74.2% MCP Atlas
• MiniMax Sparse Attention scales context to 1M
• Natively Multimodal from Step Zero

API: platform.minimax.io Token Plan: platform.minimax.io/subscrib… 🚀New! MiniMax Code: code.minimax.io

Weights & Tech Report in ~10 Days

Jun 1, 2026 · 1:59 AM UTC

278

610

4,283

844,229

@MiniMax_AI

5h

🔥API Pricing & Promotion

• 50% off standard usage (≤512K context) during the first 7 days
• Priority access available through: [email protected]
• Self-serve access for all users coming in the next few days

11

5

264

25,346

@caleb_shepherd

4m

Replying to @MiniMax_AI

I care less about the benchmarks and more about the sparse attention breakthrough. That’s the real headliner here. Well done.

61

@Brice_Gboyou

1m

Replying to @MiniMax_AI

Congrats! I’m getting this error in MiniMax Code Desktop.

1

@nijokestu

35m

Replying to @MiniMax_AI

This scale has some potential.

81

43m

Replying to @MiniMax_AI

Pls pls pls improve on creative writing and EQ, currently its horrendous

119

@techedgedaily

1h

Replying to @MiniMax_AI

MiniMax dropping an open-weights model that competes with frontier closed models is the kind of release that makes every $200/month subscription look optional.

2

786

@harsette

35m

Replying to @MiniMax_AI

'natively multimodal' does this include vision for image and video or just the former? @grok

1

888

1h

Replying to @MiniMax_AI

:3 do u guys give out token packages for OSS creators

886

@AgiWhen

1h

Replying to @MiniMax_AI

why did you guys stopped the $10 plan? it was so good for light workflows, now for $20, choosing between minimax, claude and GPT is a done deal. GPT > Claude > anyother.

you guys should have continued with the $10 plan.

1

1,305

@enpassantguy

2h

Replying to @MiniMax_AI

This is actually insane. I may ditch codex and Claude for personal use if this model provides a good developer experience.

2,597

@shree_code

3h

Replying to @MiniMax_AI

Impressive benchmarks 👏 The real test is consistency across messy, real-world repos and edge cases. I'd love to see how it handles long-horizon debugging and spec ambiguity, that's where trust is earned. 🚀

1,979

2h

Replying to @MiniMax_AI

Will it have a standard open license for weights?

533

3h

Replying to @MiniMax_AI

You're destroying the world, what about terrorists using cracked open source model to create bio weapons to eliminate all humanity?

573

@greatscottdev

4h

Replying to @MiniMax_AI

How many parameters?

2,026

@the_vc_intern

3h

Replying to @MiniMax_AI

Big leap. The real test comes when the weights drop and people run it on real agent workflows, benchmarks are one thing, messy production codebases another

4

1,811

@mustafaergisi

3h

Replying to @MiniMax_AI

59% SWE-Bench Pro at $0.30 in is the price point I have been waiting for. My overnight refactor pipeline burns a meaningful chunk of budget on frontier coding calls every week.

1

755

3h

Replying to @MiniMax_AI

You’re destroying the world. What about terrorists using cracked open-source models to create bioweapons wiping out humanity?

416

@Rigario

4h

Replying to @MiniMax_AI

Looks like we might have a new workhorse king. Will definitely be testing a lot in hermes tonight and posting my findings.

1

1,859

3h

Replying to @MiniMax_AI

seriously we need benchmark tool for benchmark tool lolz...

1

1,305

@martinovig

3h

Replying to @MiniMax_AI

Huge drop. Benchmark pretty below to frontier models, but way cheaper.

4

1,212

1h

Replying to @MiniMax_AI

@datacurve we want this model evaluated on DeepSWE ASAP !! Because I only Trust you now.😂

1

217

@witcheer

51m

Replying to @MiniMax_AI

let's gooo

60

@avaisaziz

1h

Replying to @MiniMax_AI

Impressive to see open-weights frontier coding and agentic performance with native multimodality and 1M sparse context. 59 SWE-Bench Pro is strong. Weights in 10 days?

331

@collin_michaelm

4h

Replying to @MiniMax_AI

Banger. @grok how does it stack up to deepseek v4 flash? Research the web

1

3

7,735

4h

Replying to @MiniMax_AI

Already testing it! I have to admit it's frontier finding bugs and really works very structured. It's fast enough even on opencode zen (200k context here still) and it's really usable, very well. I'm hyped a bit.

1

2,331

2h

Replying to @MiniMax_AI

I had a one year subscription. But, the previous generous limits seems to be gone overnight. Or, am I missing something?

1

1

1,768

@marchelfah

4h

Replying to @MiniMax_AI

the 1m context with sparse attention is the part i want to test most. most models claim long context but fall apart past 128k. if this actually holds up at 1m for real codebases that changes the game for multi-file refactors

2

3,412

@LowerEngineerIt

2h

Replying to @MiniMax_AI

59% SWE-Bench Pro on open weights. shipped on opus all month and the upgrade math just changed. anyone benchmarked M3 in actual workflows yet or is it still vibes?

3

1,879

@EnTr0pY_88

2h

Replying to @MiniMax_AI @mr_r0b0t

Pricier than I was hoping for, but let's see how it performs in reality. Maybe the price is dirt cheap for the output. Let's see...

5

2,519

@benjaminrestall

4h

Replying to @MiniMax_AI

What’s the deal with the high speed subscriptions now? Are they gone?

2

1,211

4h

Replying to @MiniMax_AI

Congrats! Minimax M2.7 has been my daily driver for agentic work, and the biggest model worth self-hosting. But to me it was the best possible blend of performance, speed, and price. Super excited to use MiniMax M3 to build!

1

3,958

@ctrlswftdelete

2h

Replying to @MiniMax_AI

@grok , I need your help with understanding all of the replies in this post to find out how good actually minimax m3 is

1

2,273

2h

Replying to @MiniMax_AI

@MiniMax_AI I bought a yearly Ultra HighSpeed subscription based on the advertised limits at purchase. Now the terms seem changed: token usage, weekly limits, unclear legacy benefits, and no clear prior notice. This feels unfair and misleading. Please explain or refund.

3

6

1,396

@cristianvega

5h

Replying to @MiniMax_AI

Great work team! I was a big fan of 2.7. Excited to try this one out over the coming week.

1

2,445

@UrMeer289

2h

Replying to @MiniMax_AI

Open weights, 1M-token context, frontier coding benchmarks, and native multimodality in a single model stack that's a compelling combination.

MiniMax M3 pushing performance across SWE-Bench, Terminal Bench, and agentic workflows while staying open-weight is a big win for developers and researchers. Looking forward to the technical report and seeing how the sparse attention architecture scales in real-world applications.

163

@ObinoWilliam

4h

Replying to @MiniMax_AI

Spinning it up in Opencode, let's see how it goes!

2

2,621

Load more