Breaking

MiniMax Releases M3, a 1M-Context Multimodal Model, Free to Try on NVIDIA

June 12, 2026 at 10:45 EDT

Foundation Models
Open Source
AI Agents

China's MiniMax (Xiyu Technology) released MiniMax M3, a long-context multimodal model that accepts text, image, and video inputs, on June 1, 2026. It supports up to a 1 million token context and can be tried immediately on NVIDIA's free GPU endpoint.

June 2026 · MiniMax M3

An open-weight model claims the frontier trifecta — at a fraction of the price

MiniMax M3 pairs frontier-class coding, a 1-million-token context, and native multimodality in one open model — and NVIDIA is serving it free on GPU-accelerated infrastructure.

token context window
(512K guaranteed)

428B

total params · sparse MoE
~23B active

83.5%

BrowseComp
(vs Opus 4.7's 79.3%)

Free

NVIDIA NIM endpoint
GPU-accelerated, rate-limited

Sparse-attention efficiency gain at 1M tokens

MiniMax Sparse Attention vs the previous generation — taller stack means faster.

1×

Prev gen
baseline

9.7×

Prefill
speedup

15.6×

Decode
speedup

59.0%

SWE-Bench Pro

66.0%

Terminal-Bench 2.1

83.5%

BrowseComp — leads Opus 4.7 (79.3%)

API Pricing · per 1M tokens

$0.30–0.60

input tokens

$1.20–2.40

output tokens

Also via NVIDIA NIM, OpenRouter, SiliconFlow, Together AI, vLLM (Day-0) & open weights on Hugging Face.

Praise

Strong price-to-performance
Self-directed on complex agentic & coding tasks
Practical 1M context — e.g

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.