MiniMax M3.1 Flash went live on October 19 at $0.15 input and $0.60 output per million tokens, half of what M3 costs, with a 73.7% DeepSWE v1.1 score that sits half a point under DeepSeek V4.1 Flash's September checkpoint. If you came here to find out whether you can run it locally, the answer today is no. The announcement says so in one line: "Model weights will be released later."
So for now it's an API model with an open-weights promise attached. Below I check the 70% savings claim, price it against DeepSeek V4.1 Flash, and work out what hardware you'd need if the weights do ship.
What MiniMax M3.1 Flash actually ships with
It's available now through the MiniMax API (model ID MiniMax-M3.1-Flash), the M Plan subscription and the MiniMax Code app. These are the numbers MiniMax publishes:
- Pricing per 1M tokens: $0.15 input, $0.60 output, $0.015 cached input
- Speed: "nearly three times faster" online inference, 200+ TPS output in everyday use, P50 of 170+ TPS
- Cost: agentic-task usage cost 70% lower than the predecessor
- DeepSWE v1.1: 73.7%
- CyberGym: 82%
- OSWorld V2: 61.7 (MiniMax says first among Chinese models)
- BankerToolBench: 76.2
- Agents' Last Exam: 29.6
The blog leaves out parameter count, context length and license, and it doesn't put a single named competitor in a chart. You get "on par with top models" and nothing to check that against.
One detail buried in the post: the anonymous "Space Bunny" model, which processed more than 200 trillion tokens before anyone knew whose it was, was an early version of M3.1 Flash. If you used Space Bunny, you've already tried this model.
The 70% cheaper claim, checked against M3
MiniMax M3 lists at $0.30 input and $1.20 output on MiniMax's API. M3.1 Flash is $0.15 and $0.60. Per token, that's 50% off, not 70%.The 70% figure is about "agentic-task usage cost," so the remaining 20 points have to come from the model finishing tasks in fewer tokens. Maybe it does, but you can't verify that from a blog post. One commenter on Hacker News (sinuhe69) noted that MiniMax models spend a lot of time in chain-of-thought, which makes effective output slow on OpenRouter. And chriddyp, posting results from Plotly's data-analytics bench, reported that it tied Opus 5.5 at 41/43 for about a third of the price, but used "A LOT of tokens" and was the slowest model they'd tested. If it burns more reasoning tokens on your workload, your bill drops by less than half, not more.
MiniMax M3.1 Flash vs DeepSeek V4.1 Flash
DeepSeek V4.1 Flash is the obvious comparison: same "Flash" positioning, same headline price, and it shipped a month earlier.
| | MiniMax M3.1 Flash | DeepSeek V4.1 Flash |
|---|---|---|
| Input / 1M | $0.15 | $0.15 off-peak, $0.30 peak |
| Output / 1M | $0.60 | $0.60 off-peak, $1.20 peak |
| Cached input / 1M | $0.015 | $0.003 off-peak, $0.006 peak |
| DeepSWE v1.1 | 73.7% | 74.2 (Sept), 77.0 (Flash-1017 update, Oct 17) |
| CyberGym | 82% | 88.1 |
| Parameters | not disclosed | 552B total, 8B/16B active |
| License | not disclosed | MIT |
| Weights | "later" | on Hugging Face (~510 GB FP8) |
DeepSeek figures come from its launch materials as compiled by Yotta Labs and gradually.ai. DeepSeek's own off-peak rates are quoted in RMB (1 / 4 / 0.02), with peak hours at double. Every benchmark in that table on both sides is vendor-reported.
Cached input is where the two split, and agent loops live on cache hits. Take a session with 10M input tokens, 80% of them cached, plus 1M output:
MiniMax M3.1 Flash: 2M × $0.15 + 8M × $0.015 + 1M × $0.60 = $1.02
DeepSeek V4.1 (off-peak): 2M × $0.15 + 8M × $0.003 + 1M × $0.60 = $0.92
DeepSeek V4.1 (peak): 2M × $0.30 + 8M × $0.006 + 1M × $1.20 = $1.85If your agents run during DeepSeek's peak windows (weekdays 01:00–04:00 and 06:00–10:00 UTC), MiniMax's flat pricing wins. Batch jobs that run off-peak are a wash, and DeepSeek has the weights you can actually download.
On quality, I'm not going to pick a winner off two vendor charts. The HN thread is split. dom96 says it performs on par with DeepSeek V4 Pro on their adversarial programming-language benchmark. eskibars, who runs cybersecurity evals, says it "far underperforms" GLM 5.3, which is an awkward thing to read next to an 82% CyberGym headline.
Can you run MiniMax M3.1 Flash locally?
Not yet. When the weights do land, the hardware question depends on a number MiniMax hasn't published.
Everything in this section is an estimate. One HN commenter said MiniMax's FAQ lists M3.1 Flash at 428B total / 23B active, the same as M3, and a couple of model-tracker sites repeat that figure. The announcement doesn't state it, and nothing official I can find does either. Another commenter guessed ~150B because of the "Flash" name and the 3x speed claim. Here's what each scenario implies, based on real M3 checkpoints:
| Precision | If ~428B (M3-sized) | If ~150B (guess) |
|---|---|---|
| BF16 | ~856 GB | ~300 GB |
| FP8 / MXFP8 | ~444 GB (M3 MXFP8 source) | ~150 GB |
| 4-bit | ~237 GB (community M3 MXFP4) | ~80 GB |
Those numbers cover weights only. KV cache comes on top, and at long context it adds a lot.
For reference, the vLLM recipe for M3 targets 8× H200 or H20 for BF16, and the official MXFP8 variant runs from TP=4 on AMD MI350X/MI355X (MI300X still uses TP=8). The community MXFP4 quant is 237 GB and fits a single ≥256 GB GPU like a B300. Someone reported it running on 8× 48 GB 4090s, and the card says it won't fit on 2× 128 GB DGX Sparks.
If M3.1 Flash really is M3-sized with a faster serving stack, "local" means a multi-GPU workstation or a 256 GB+ unified-memory box running a 4-bit quant. Your gaming PC is out; this is homelab-with-a-budget territory. If it's closer to 150B, a 4-bit quant on a 128 GB machine becomes plausible. We don't know which, and MiniMax could settle it with one line in the blog post.
What to watch on Hugging Face
When something drops under MiniMaxAI, check these before you start a 400 GB download:
config.json: total experts, num_experts_per_tok, max_position_embeddings. That tells you the real size and context.--tool-call-parser minimax_m3 --reasoning-parser minimax_m3 and --block-size 128.You can poll the org's newest repos without refreshing the page:
curl -s "https://huggingface.co/api/models?author=MiniMaxAI&sort=createdAt&direction=-1&limit=5" | jq -r '.[].id'Until that list shows something with "M3.1" in it, a local MiniMax M3.1 Flash setup doesn't exist. If you need open weights at this price point right now, DeepSeek V4.1 Flash is on Hugging Face under MIT today, and M3 is there if you have the hardware.
My take
The pricing is real and the speed claims are easy to test through the API, so test them on your own agent traces before you believe the 70%. The "open model" part stays a promise until there's a repo. Labs have said "weights later" before and then slipped by months, changed the license, or released a different size. I'll count M3.1 Flash as an open model the day config.json is public, and not before.
Do you trust "weights coming later" from a lab, or does a model not count as open until it's on Hugging Face? And if you ran Space Bunny on real work, did it hold up against DeepSeek V4.1 Flash?