October 2026 Summaries

MiniMax has released open weights for MiniMax M3.1 Flash, a 428B-parameter mixture-of-experts model with 23B active parameters per token designed for coding, agentic workflows, computer use, and security tasks. Previously tested anonymously as “Space Bunny,” the model reportedly served more than 200 trillion tokens before launch and introduces full-model sparse attention, training-time 8-bit quantization, and a DSpark draft head for speculative decoding, which MiniMax says reduce serving costs to roughly one-third of M3’s and increase speed threefold. Artificial Analysis rates it 43 on its Intelligence Index at an estimated $0.22 per task, with a 48% score on Humanity’s Last Exam, while MiniMax reports strong results across automation, software engineering, desktop operation, cybersecurity, and visual reasoning benchmarks. Featherless hosts the model with FP4 quantization, vision support, and a 256K-token context window through an OpenAI-compatible API, while MiniMax’s native context capacity reaches 1 million tokens; Featherless also promotes dedicated GPU deployments for higher-volume use.

StepFun has released open weights for Step 5 Preview, a 600-billion-parameter mixture-of-experts model with 27 billion active parameters per token, designed for long-horizon agent tasks such as software development, frontend generation, and financial analysis. Its “narrow but deep” architecture uses 92 transformer layers and Sparse GQA to support a native one-million-token context window while aiming to reduce long-context processing costs; Featherless currently offers it with a 256K-token FP8 context window and vision support. Artificial Analysis reportedly gave the model an Intelligence Index score of 44, compared with 58 for Claude Opus 5.5, while estimating its evaluation cost at $1.03 per task versus $5.98, positioning it as a lower-cost alternative for advanced agent work. Featherless says it tested the model by porting a game codebase to TypeScript and producing acquisition theses for more than 100 technology companies, and it provides access through an OpenAI-compatible API at rates of $1 per million input tokens, $0.05 per million cached input tokens, and $2.70 per million output tokens.