GPT-6 Luna
MiMo-V2.6-Flash
- input
- $0.075
- cached
- $0.0075
- output
- $0.30
- Context
- 1Mtokens
- Released
- 3w ago
- Size
- 428Bparameters
Cost/1M tokens
MiniMax M3.1 Flash is a mixture-of-experts model built on MiniMax-M3 and positioned as the next generation of MiniMax's Flash line, adding sparse attention, quantization-aware training, and DSpark speculative decoding for efficiency. Its launch announcement headlines agent performance and cost, claiming performance at the level of leading models with online inference nearly three times faster and usage costs 70% lower than its predecessor. It is built for coding, agentic workflows, computer use, and cybersecurity.
- Multimodal model
- Open weights · 428B
- Accepts images, video
- 1M context
- Reasoning
- Tools
- Structured output
How it compares
Score vs. costArena WebDev
Each model at its best score · Dashed: the best score at each cost
- MiniMax
- Anthropic
- Google DeepMind
- OpenAI
- Moonshot AI
- Meta
- xAI
- Z.ai
- Xiaomi
Level
LightweightThe lab's smallest, cheapest and lowest-latency hosted line, built for high volume and simple tasks.
Lightweight models from other labs
Vs. your stack
Add the models you use to see how this one compares.
Sign inAdditional source evidence
- Source model
- minimax/MiniMax-M3.1-Flash
- Source family
- minimax
- Last updated
- 2026-10-19
- Model context
- 1M
- Model output limit
- 512K
- Open weights
- Yes
- File attachments
- Yes
- Reasoning
- Yes
- Tool calling
- Yes
- Temperature control
- Yes
Complete source record (JSON)
{
"id": "minimax/MiniMax-M3.1-Flash",
"name": "MiniMax-M3.1-Flash",
"description": "Frontier multimodal coding model with 1M context window and tunable thinking depth",
"family": "minimax",
"last_updated": "2026-10-19",
"attachment": true,
"reasoning": true,
"tool_call": true,
"temperature": true,
"modalities": {
"input": [
"text",
"image",
"video"
],
"output": [
"text"
]
},
"limit": {
"context": 1000000,
"output": 512000
},
"release_date": "2026-10-19",
"open_weights": true
}Frequently asked questions
Running locally
MiniMax M3.1 Flash is estimated to need 298.9 GB of VRAM at Q4, 523.5 GB at Q8 and 984.9 GB at FP16, based on its 428B parameters. All figures are estimates: weights plus about 15% and 0.5 GB for a short context.
No. MiniMax M3.1 Flash does not fit on a 24 GB, 16 GB, 12 GB or 8 GB GPU at Q4, Q8 or FP16 (estimate).
Yes, but only on a large Mac. At Q4 the model needs an estimated 298.9 GB, and since macOS lets the GPU use about 75% of unified memory, the smallest common Mac memory size that fits Q4 is 512 GB (estimate).
No support is recorded for llama.cpp, vLLM, LM Studio or MLX. The only runtime with recorded support is Ollama, and no Ollama builds have been published.
API and pricing
MiniMax M3.1 Flash costs $0.075 per 1M input tokens and $0.30 per 1M output tokens. Across 11 paid providers, input prices range from $0.06 to $0.15 per 1M tokens.
SambaNova is the cheapest provider by lowest blended price (3:1 input to output). SambaNova charges $0.06 per 1M input tokens and $0.24 per 1M output tokens.
Yes. Cached input costs $0.0075 per 1M tokens versus $0.075 per 1M tokens uncached, a 90% discount.
A typical request with 10,000 input tokens and 1,000 output tokens costs $0.00084 at the cheapest provider, SambaNova.
Of 11 endpoints with uptime data in the last day, 10 were at 99% uptime or better. Best uptime in the last day was 100.0%, by DigitalOcean, InferenceNet, Minimax and 3 more.
Privacy and data
SambaNova, DigitalOcean, Fireworks, InferenceNet, Nebius, Parasail and Together offer zero data retention (ZDR eligible or 0-day retention).
Yes. MiniMax M3.1 Flash has open weights, so it can be self-hosted for full control of your data.
Capabilities and specs
The context window is 1M tokens (1,048,576) and the maximum output is 524K tokens (524,288).
It accepts text, image and video input, so it can read images and video. Audio and PDFs are not among its listed input modalities.
Yes, tool calling (function calling) is supported.
Yes, structured output (JSON schema or JSON mode) is supported.
Yes, reasoning is supported and always on. Effort levels are max, xhigh, high, medium and low, with max as the default. All 11 of 11 tracked providers expose reasoning, and a provider offers the reasoning_effort setting.
Quality and benchmarks
On Arena's own ranks, MiniMax M3.1 Flash is rank 36 on Agent Arena with a score of 0 and rank 41 on Arena WebDev with a score of 1585.
Alternatives matched on model kind, blended price and release date are DeepSeek V4.1 Flash 1017 by DeepSeek at $0.1295 per 1M tokens, MiMo-V2.6-Flash by Xiaomi at $0.175 per 1M tokens and DeepSeek V4.1 Flash by DeepSeek at $0.18 per 1M tokens.
DeepSeek V4.1 Flash 1017MiniMax M3.1 Flash vs DeepSeek V4.1 Flash 1017MiMo-V2.6-FlashMiniMax M3.1 Flash vs MiMo-V2.6-FlashDeepSeek V4.1 FlashMiniMax M3.1 Flash vs DeepSeek V4.1 Flash
Versions and lineage
MiniMax M3.1 Flash has a context of 1M tokens (1,048,576) while MiniMax-M3.1-Flash-Preview has 1M tokens (1,000,000). MiniMax M3.1 Flash costs $0.075 input and $0.30 output per 1M tokens. It was released October 19, 2026, and the preview September 27, 2026.
MiniMax-M3.1-Flash-PreviewMiniMax M3.1 Flash vs MiniMax-M3.1-Flash-Preview
Licensing and access
MiniMax M3.1 Flash has open weights.
The MiniMax M3.1 Flash weights are available in the Hugging Face repository MiniMaxAI/MiniMax-M3.1-Flash.
Basics
MiniMax made MiniMax M3.1 Flash, and it was released on October 19, 2026.
No. It has open weights, but the listed API price is $0.075 input and $0.30 output per 1M tokens.
Generated from Vertical Tabs data, updated November 12, 2026.
Discussion
No comments yet
Be the first to share your thoughts.