MiniMax M3.1 Flash

★★★★☆4.0 / 5 editorial Freemium Pay-as-you-go API: $0.15 per 1M input / $0.60 per 1M output tokens, $0.015 cached input — check the official site for current pricing.

428B MoE agentic model with 1M context at budget API pricing.

✓ Editor-researched listing — details reviewed by Grab AI Tool editors; pricing re-checked regularly.

Try it Free → ♡ 1,811 saves

#llm#coding#agents#api#moe

What is MiniMax M3.1 Flash?

MiniMax M3.1 Flash, launched October 19, 2026, is a 428-billion-parameter mixture-of-experts model (about 23 billion parameters active) built for agentic work. MiniMax confirms the model previously topped leaderboards anonymously as “Space Bunny,” which processed over 200 trillion tokens before the reveal. The company says output speed reaches 200+ tokens per second, with agentic-task costs 70% lower than its predecessor, and reports 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, and 82% on CyberGym for security tasks.

M3.1 Flash accepts text, image, and video input with a 1-million-token context window, and reasoning is always on with five effort levels from Low to Max. It is available now through the MiniMax API, the M Plan subscription, and the MiniMax Code desktop app, and open weights followed on October 23, 2026 under the MiniMax Community License. Flat pricing of $0.15 per million input tokens and $0.60 per million output tokens positions it squarely as the budget option for teams running long agent workflows.

Who MiniMax M3.1 Flash Is Best For

Best for: MiniMax M3.1 Flash suits teams especially well: Near-frontier agentic scores at flash-tier pricing for long agent workflows; 1M context and cheap cached tokens fit agents that re-read big codebases. The free plan makes it easy to try on real work before paying anything.

Not ideal for: Less suited to users who will hit its limits head-on: reasoning is always on and cannot be disabled, which changes migration behavior for M3 users; the open weights are a ~252 GB download under the MiniMax Community License — commercial use carries attribution and authorization conditions.

Key Features

428B MoE model (~23B active) tuned for coding and agentic office work
1-million-token context window with text, image, and video input
200+ tokens per second output with always-on reasoning at five effort levels
Flat API pricing of $0.15 input / $0.60 output / $0.015 cached input per 1M tokens
Available on MiniMax API, M Plan subscription, and MiniMax Code desktop app

Pros

Near-frontier agentic scores at flash-tier pricing for long agent workflows
1M context and cheap cached tokens fit agents that re-read big codebases
The Space Bunny stealth run gave it public benchmark history before launch

Cons

Reasoning is always on and cannot be disabled, which changes migration behavior for M3 users
The open weights are a ~252 GB download under the MiniMax Community License — commercial use carries attribution and authorization conditions
All headline benchmarks are MiniMax-reported; independent verification is still pending

Pricing

Model: Freemium
Pay-as-you-go API: $0.15 per 1M input / $0.60 per 1M output tokens, $0.015 cached input — check the official site for current pricing.

MiniMax M3.1 Flash Specs: Size, Context and Speed

“How big is MiniMax M3.1 Flash?” is the question most people mean by the model's size, and it has two answers. In parameters, M3.1 Flash is a 428-billion-parameter mixture-of-experts model, but only about 23 billion parameters are active for each token — that split is what keeps inference fast and cheap while the full model's capacity stays on tap. As a download, the open weights released on October 23, 2026 come to about 252 GB across 48 safetensors shards: the checkpoint is stored with quantization-aware training (MXFP8, with NVFP4 on the routed expert path), which is why it is far smaller than the roughly 856 GB the same 428 billion parameters would occupy at full 16-bit precision.

MiniMax Sparse Attention runs in every layer — including the first three layers, which still used full attention in M3 — and a bundled DSpark speculative-decoding draft head helps serving stacks squeeze out more speed. The model card's recommended sampling settings are temperature 1.0 and top-p 0.95.

Total parameters: ~428 billion (mixture-of-experts)
Active parameters: ~23 billion per token
Context window: 1,000,000 tokens
Inputs / output: text, image and video in; text out
Reasoning: always on, five effort levels (low, medium, high, xhigh, max)
Output speed: 200+ tokens per second (MiniMax-reported)
Weight download: ~252 GB in 48 shards, MiniMax Community License
API price: $0.15 input / $0.60 output / $0.015 cached input per 1M tokens

What MiniMax M3.1 Flash Costs to Run

Because pricing is per token, the real cost depends on the shape of your workload. At $0.15 per million input tokens and $0.60 per million output tokens, a coding agent that reads 10 million tokens of context and writes 1 million tokens of code costs about $2.10 in API fees. Cached input drops to $0.015 per million, so repeated system prompts and long-lived contexts cost far less than the headline rate suggests.

1M input tokens: $0.15
1M output tokens: $0.60
1M cached input tokens: $0.015
Example agent run (10M in / 1M out): about $2.10

Adding Up Your AI Bill?

If you pay for MiniMax M3.1 Flash alongside other AI subscriptions, the monthly total adds up quietly. Tick every tool you pay for in our free AI Tool Cost Calculator to see your real monthly and yearly bill — and spot overlaps you can cut. Still choosing? Start with our guide on how to choose the right AI tool.