This edition covers 7:40 am on 19 October to 7:40 am on 20 October 2026, Sydney time. It records two changes: the formal release of MiniMax M3.1 Flash and AWS's technical announcement of TorchNeuron's native PyTorch backend and access conditions.
MiniMax formally releases M3.1 Flash; weights remain forthcoming
On 19 October, MiniMax formally released M3.1 Flash through its API, M Plan and MiniMax Code. Compared with the earlier Preview channel, the new facts are a formal version announcement and named API documentation. The company confirmed that Space Bunny was an early version; anonymous usage volume is not capability validation. Weights remain scheduled for a later release, so current paid access is not described as an open-weight release.
The documented model identifier is MiniMax-M3.1-Flash, with a one-million-token context window. The Anthropic-compatible interface requires thinking to remain enabled, with adjustable effort; setting it to disabled is not supported. The documentation also excludes the service_tier parameter for this model. Standard prices per million tokens are US$0.15 for input, US$0.60 for output and US$0.015 for cached input. Multi-turn tool calls require the complete response, including thinking and tool content, to be retained.
Nearly threefold online inference speed, a 70 per cent reduction in agentic usage costs, and coding, computer-use and cybersecurity results are developer-reported. Reuters independently reported the release but attributed those gains to the company, without supplying an independent performance reproduction. This review made no paid API calls. Token prices are not equated with the final cost of every task, and the claims are not used to establish a universal capability ranking. Direct references are MiniMax's announcement, API documentation and Reuters.
TorchNeuron details a native backend, with public Beta scheduling and compatibility limits
On 19 October, the AWS team published a PyTorch technical announcement explaining TorchNeuron's use of standard extension points, including PrivateUse1, to integrate Trainium2 and Trainium3 with native PyTorch. It provides eager execution, torch.compile and custom NKI kernel paths, distinct from the existing PyTorch/XLA product. This is not Trainium's first support for PyTorch.
The announcement schedules the Apache 2.0 open-source launch and public Beta for 20 October. At the time of checking, the official torch-neuronx repository's source, licence and Beta installation instructions were publicly readable. Installation was not tested on a Trainium instance, and this is not presented as stable service availability for every customer. Reuters corroborated the announcement and quoted AWS as having more than 30 early customers; that count remains supplier-provided.
The repository specifies static shapes for compiled execution, recompilation for new shapes and initial compilation overhead. Some APIs require CPU fallback or are unsupported, and FlexAttention support is limited. Consequently, the no-rewrite description concerns migration through supported code and interfaces, not compatibility with arbitrary CUDA programs, dynamic models or every operator. The material does not supply a cross-chip performance conclusion independently reproduced in this review. Direct references are the PyTorch announcement, AWS repository and Reuters.
Sources
MiniMax; MiniMax M3.1 Flash: Redefining Agent Performance and Cost; 19 October 2026.
https://www.minimax.io/blog/minimax-m31f
MiniMax; Anthropic SDK; continuously updated API documentation, checked for the formal model identifier on 20 October 2026; no separate publication date displayed.
https://platform.minimax.io/docs/api-reference/text-anthropic-api
Reuters; China's MiniMax launches updated AI model M3.1 Flash, claims major improvements; 19 October 2026.
https://www.reuters.com/world/asia-pacific/chinas-minimax-launches-updated-ai-model-m31-flash-claims-major-improvements-2026-10-19/
AWS team / PyTorch; Introducing TorchNeuron: Native PyTorch backend for AWS Trainium; 19 October 2026.
https://pytorch.org/blog/introducing-torchneuron-native-pytorch-backend-for-aws-trainium/
AWS Neuron; torch-neuronx (Beta) repository, README and licence; checked 20 October 2026; no separate publication date displayed on the repository page.
https://github.com/aws-neuron/torch-neuronx
Reuters; Amazon says its AI chips will work with popular coding language, challenging Nvidia; 19 October 2026.
https://www.reuters.com/business/retail-consumer/amazon-says-its-ai-chips-will-work-with-popular-coding-language-challenging-2026-10-19/
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.