Open models
MiniMax Opened a 428B-Parameter Model That Activates Only 5% of It
MiniMax released the weights of M3.1 Flash on October 23 on Hugging Face and ModelScope. The model has 428 billion total parameters but activates about 23 billion per token, roughly 5%, and supports a 1 million token context with text, image and video input. Its published API price is about $0.15 per million input tokens and $0.60 per million output tokens. The release matters less as a leaderboard entry than as a price and deployment signal: multimodal capacity is now available as downloadable weights and as a very cheap API at the same time.
The architecture and the price
M3.1 Flash is a mixture-of-experts model. The headline parameter count describes total capacity; the activation count describes what each token actually costs to compute. At roughly 5% activation, inference cost tracks the active experts rather than the full 428B, which is how MiniMax can publish a price in the cents-per-million-token range.
The official MiniMax news page lists M3.1 Flash as a current product. Chinese financial media reported the open-weight release and the pricing. The same reports say the model was previously tested anonymously on OpenRouter, where it ranked first in usage for ten consecutive days and processed about 56.7 trillion tokens, and that it passed 100 trillion tokens within 72 hours of release. Those usage figures are media reports, not an independently audited number.
Why open weights plus a low price is the combination to watch
There are two separate markets here. The API price sets the cost of a hosted call. The open weights set the cost of running the model on your own hardware, with data staying inside your network.
The reports say the weights had day-zero support in vLLM and SGLang and were adapted to Ascend, T-Head, Kunlunxin, Moore Threads, NVIDIA and AMD hardware. That adaptation matters more than the download link. A weight release is only useful if inference stacks and accelerators can run it without a long porting project. If the support is as broad as claimed, the model becomes portable across supply chains, which is a different kind of leverage from a cheaper API.
The caveat is selection. A Flash tier is optimized for cost and speed, not for the hardest reasoning tasks. Parameter counts and vendor benchmarks do not tell you whether it will hold up on your prompts. The only useful test is your own evaluation set, including the long-context and tool-calling cases you actually run.
What this changes for mobile AI features
A 428B-parameter model does not run on a phone, and this release does not change that. The mobile impact is economic and architectural.
At $0.15 per million input tokens, multimodal calls stop being a budget line that needs approval for every experiment. Features that were previously limited to short prompts, such as summarizing a photo library, reading a document or tagging video, can be designed around longer inputs and more retries. The 1M context also changes what a single request can contain, although latency and cost still grow with use.
The second effect is placement. Cheap hosted inference makes cloud-first features easier to justify; open weights make a self-hosted or regional deployment possible when data cannot leave a network. Most mobile teams will end up with a split: small models on device for latency and privacy, a cheap hosted model for the heavy lifting. This release makes the second half of that split cheaper, not the first.
Sources
Corrections: if a source changes or a claim is wrong, email the address in the footer and this page will be updated with a note. Subscribe via RSS.