MiniMax M3.1-Flash: the 428B 'Space Bunny' Everyone Switched To in October
Status board (checked 2026-11-04)
| Who made it | MiniMax (Shanghai); community nickname “Space Bunny 🐰🚀” |
| License | MiniMax Community License (model card) |
| Export-control status | ⚠️ Chinese lab with overseas export registration done; catalogue still in comment period — watch this box |
The numbers
| Total / active | 428B / 23B |
| Open weights | 2026-10-23; >100 trillion tokens processed by real users in the first 72h (their stat) |
| Tricks | full-model sparse attention + quantization-aware training (QAT) → ~70% serving cost cut vs M3 |
| Home packs | Maya-XS ≈ 93 GB (whole model in a 128 GB box’s memory) · Maya-S ≈ 122 GB (closer to the original) · EXL3 quants ~95–200 GB |
| Measured speeds | 48.6 tok/s on one RTX PRO 6000 96 GB at 256K · 38.2 on a 3090 + 8×5060 Ti · 21–23 on dual 3090s — top-token match with the original: 87.6% |
| Native multimodal | text+image; engines route vision to CPU or GPU by card size |
Getting started (3 steps)
- Pick it in Project Maya’s setup (
--model minimax-maya-xs) or grab the safetensors/GGUF/EXL3 off HuggingFace. - Let the plan tell you what fits; 23B active means decode speed follows your memory bandwidth, not your core count.
- Bench it on your box (
--bench) before trusting any viral benchmark clip — thinking levels (Off/Low/High/Max + a token budget) are built into the local APIs.
Best use
Agentic work + cybersecurity-flavored tasks (MiniMax pushes these), with more raw capacity than GLM-5.3-Flash at the price of needing a bigger, faster rig.
Honest weaknesses
- Newest of the trio: ecosystem tooling (templates, drafter support) is still weeks behind.
- 428B doesn’t care about your feelings — on 48–64 GB rigs it’s an SSD-crawl experience; this model wants 128 GB of memory.
FAQ
M3.1-Flash vs GLM-5.3-Flash?
Closer call than the old days: 428B/23B vs 320B/18B — M3.1 brings more total capacity, GLM’s attention architecture brings cheaper long context and smaller packs. Bench your own workload.
vs DeepSeek-V4.1-Flash (748B)?
DeepSeek stays the “biggest headline” tier; M3.1 sits between it and the 320B class in capacity and in hardware appetite.
Why “Flash” in the name?
It’s the served-cheap-and-fast tier of each lab’s flagship line — the local packs let you serve it yourself instead, which is the entire point of this site’s Local AI section.
📌 Summarised from official docs and hands-on samples. Product features and pricing change often — official site wins.
← Newer: How a 100B+ Model Fits in a Gaming PC — what Strata and Project Maya changed Older: Project Maya: the Scheduler That Puts 748B on Your Desk →Related: Strata: the Engine That Made 125B a Gaming-PC Hobby
Qwen3.8-Flash-Next: the 125B That Made 100B-at-Home a Hobby
Project Maya: the Scheduler That Puts 748B on Your Desk