MiniMax M3.1 Flash goes open-weight: 428B MoE, 1M context, 252 GB
Free to self-host under $20M revenue, with vLLM day-one support — the same day NVIDIA ships its $4,999 DGX Spark 64GB and a 4GB Nemotron edge model.
Copy markdown
A 428B open-weight MoE you can actually run
MiniMax posted M3.1 Flash to Hugging Face today — 428B total, 23B active, a 1M-token context, and 252 GB across 48 shards plus a speculative-decoding drafter. vLLM has day-zero containers for NVIDIA and AMD, the hosted API runs $0.15/$0.60 per million in/out tokens, and it tops Automation-Bench v1 at 62.2% over six rivals.
Read the license before you ship on it
It's the MiniMax Community License, not MIT or Apache. Companies at or below $20M annual revenue just send MiniMax a one-time email to use it commercially; above that you need written sign-off first, and attribution is required either way. The benchmark wins are MiniMax's own runs, so test on your tasks before you commit.
NVIDIA's $4,999 DGX Spark 64GB ships today
The 64GB desktop box lands today at $4,999, and NVIDIA Sync links two of them into 128GB of coherent memory over a single cable — or buy the 128GB unit for $6,950. That's enough local headroom to fine-tune and serve mid-size open models without touching the cloud.
Nemotron Bolt pushes agents onto 4GB edge devices
NVIDIA's Bolt line is built for local agents: the 74B-A9B model is natively 2-bit quantized into 24GB, and a 4GB E4B variant is slated to drop today for RTX, Jetson and DGX Spark — aimed at on-device voice assistants and multi-turn agentic flows with no cloud round-trip.