MiniMax’s newest model, MiniMax M3.1 Flash, is live in Kilo. And it’s totally free for a limited time.
This is MiniMax’s fourth limited-time free release in Kilo. MiniMax M2 started the trend, and for a while it was one of the top three models used by Kilo Coders. M2.1 followed as a New Year’s gift. Then in June we covered MiniMax M3, the first open-weight model to combine a 1M-token context window, frontier coding and agentic performance, and native multimodality. M3.1 Flash takes that M3 foundation and makes it faster and cheaper.
MiniMax released the model after a stealth period where many agentic engineers around the world tried it in their daily workflows, with the model known only as Space Bunny. Well, the space bunny is here to stay, and even after the free period we expect this new flash model to be a daily driver for a lot of AI builders.
Here’s what the new flash model is, how it scores and where it’s strongest.
What is MiniMax M3.1 Flash?
M3.1 Flash is a multimodal mixture-of-experts model. MiniMax’s product page lists 428B total parameters with about 23B active, along with sparse attention and a native visual encoder. That sparse attention is the MiniMax Sparse Attention (MSA) design we covered at the M3 launch. MSA replaces the quadratic cost of full attention with KV-block selection, which brings per-token compute at 1M context down to roughly 1/20th of the previous generation.
It may seem familiar. OpenRouter said Space Bunny, a stealth model, left its platform on October 5th and returned on October 19th, as M3.1 Flash. That stealth build processed over 50 trillion tokens, so plenty of developers have already used it without knowing what it was.
MiniMax is pitching a big efficiency gain. The company says agentic tasks cost 70% less overall than with its predecessor, and describes M3.1 Flash as delivering frontier-level intelligence across cowork, coding and visual-reasoning tasks at one-third the cost of M3 and three times the speed.
Benchmarks
Keep one caveat in mind before reading any numbers: every score MiniMax published is self-reported, and no independent replication exists yet. That said, the new flash model is impressive at a wide range of agentic work.
The strongest results are in agentic office work. On Automation-Bench v1.0.6, M3.1 Flash scores 62.2, ahead of Opus 5 at 50.3 and GPT-5.6 Sol at 45.8. It ties Opus 5 on BankerToolBench at 76.2 and tops LOCA-Bench 256K at 60.3.
Coding results are mixed. On DeepSWE v1.1 it scores 73.7, just above GPT-5.6 Sol at 73.0, while DeepSeek-V4.1-Flash-1017 scores higher at 77.0. Terminal Bench 4.0 is its weakest result at 38.1, and SWE Refactor Bench comes in at 23.0 versus 31.0 for Opus 5.
Vision is strong. It reaches 91.0 on BabyVision with tools and 82.5 on CharXiv RQ, crushing the benchmark on visual reasoning, alongside polished frontend and presentation creation tasks. Vision capabilities are an increasingly important way to differentiate, as we also saw with the release of NVIDIA Nemotron 3.5 Super VL just last week.
Independent scores are starting to come in. Artificial Analysis gives it an Intelligence Index of 42.7, with 48.0% on HLE, 79.7% on AA-LCR, and 53.4% on SciCode. One gap is worth noting. Artificial Analysis measured 25.8% on Terminal-Bench 4.0, well below MiniMax’s own 38.1. Harness choices can account for a gap like that, but it supports treating vendor numbers carefully. For comparison on our own benchmark, MiniMax M3 scored 48% on KiloBench, so watch the leaderboard to see where the Flash variant lands.
How it compares to the latest flash models
Last month, Gemini 3.8 Flash set the standard for the Flash tier. Flash models are no longer just fast, they’re increasingly capable daily drivers—just don’t rely on them for planning. The two models target the same work in different ways.
Gemini leads on intelligence. Gemini 3.8 Flash scored 59 on the Artificial Analysis Intelligence Index at high effort and 75.3% on KiloBench, and Artificial Analysis measures it at 352 output tokens per second. M3.1 Flash comes in at 42.7 on the same index, and OpenRouter currently clocks it at about 82 tokens per second.
On price, MiniMax is far cheaper. Gemini 3.8 Flash costs $0.75 input and $3.75 output per million tokens at introductory pricing, rising to $1.50 / $7.50 after December 31. At $0.15 / $0.60, M3.1 Flash costs about one-fifth as much per input token and one-sixth as much per output token. Plus it’s free in Kilo right now!
But Google isn’t the only player in the game. Another popular flash model release is MiMo V2.6 Flash from Xioami. M3.1 Flash holds an edge in raw reasoning power and active parameter capacity, while Xiaomi MiMo-V2.6-Flash offers lower output costs and broader native modality (including audio).
MiniMax M3.1 Flash activates around 23B parameters per forward pass (out of 428B), giving it a heavier capacity for deep logic and complex codebases. MiMo-V2.6-Flash uses a lighter footprint of around 15B active parameters (out of 309B) combined with Multi-Token Prediction (MTP). While M3.1 leads in vision, MiMo leads in audio and speech-to-action agent loops.
It’s really an embarrassment of riches. So many flash models to choose from, so little time. Gemini flash models are getting pretty good at planning—they aren’t just for long-horizon agentic tasks—so you could easily start your agentic engineering project with Gemini and then move to subagents on MiMo and MiniMax.
The future is now
This release fits the trend we described in The Age of the Flash Model: the flash tier is becoming the default tier for always-on agentic engineering, with competition shifting to cost per task, throughput, and whether a model stays on track through long tool-calling loops. Gemini 3.8 Flash is the better pick when you need speed. M3.1 Flash is the better pick when you need volume at the lowest cost.
At a recent panel about open-weight models I was asked which model I would use if I were starting a large agentic project from scratch. My response was simply: “pick your favorite flash model and dive in.” Looks like MiniMax is now in the running for that spot.
But these are still all just data points. Give MiniMax M3.1 Flash a spin in Kilo today and let us know what you think. Try it while the free period lasts!
It's free in openrouter, helps you to hook Minimax M3 in other agent harness too