Breaking

MiniMax M3 Open Weights Land on Hugging Face, NVIDIA Adds Free Endpoint

June 12, 2026 at 12:18 EDT

Foundation Models
Open Source
AI Agents

China's MiniMax has released the open weights for MiniMax-M3 on Hugging Face, while NVIDIA simultaneously launched a free inference endpoint for the model on its own platform. Developers can now test the model at no cost through an OpenAI-compatible API.

June 2026 · MiniMax

MiniMax M3 goes open-weight — 1M-token, multimodal, at a fraction of the cost

The Chinese frontier model publishes its weights on Hugging Face while NVIDIA opens a free testing endpoint — claiming closed-model performance at roughly 5–10% of the price.

428B

total params (~23B active, MoE)

max context tokens

100T+

pretraining tokens (multimedia)

5–10%

of closed-model cost (claimed)

MiniMax Sparse Attention speedups at 1M tokens

Relative to a 1× baseline — the taller the column, the faster.

1×

baseline

9.7×

faster prefill

15.6×

faster decode

Vendor-reported benchmark claims

59.0%

SWE-Bench Pro

66.0%

Terminal-Bench 2.1

83.5

BrowseComp (vs Opus 4.7: 79.3)

Many figures are vendor-reported and can vary under independent testing.

What excites developers

1M context in an open-weight model — rare and significant
Competitive with Claude Opus 4.8 & GPT-5.5 on code & refactoring
Open weights enable local runs, fine-tuning & self-hosting
Promo pricing: $0.30 in / $1.20 out per 1M tokens

Where caution is warranted

Benchmarks are vendor-reported, vary independently
Can lag Opus 4.8 on real-world reliability & responsiveness
Local trouble on engines lacking MSA support
Data-regulation concerns for a Chinese-company model

Continue reading

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