🚀 RELEASES & LAUNCHES
StepFun releases Step 5 Preview: 600B MoE agentic model with 1M-token context
StepFun released Step 5 Preview, a 600B-parameter (27B active) sparse MoE flagship model for agentic work with a 1M-token context window.- Targets developers and agents doing software engineering, long-document processing, deep research, and finance work
- Sparse Mixture-of-Experts architecture with 27B active / 600B total parameters; 1M-token context, up to 64k output tokens; accepts text, image, and video input
- Priced at $1.00/M input and $2.70/M output tokens ($0.05/M cache read) on OpenRouter, served by DeepInfra and StepFun; released October 8, 2026
- Scores 44 on the Artificial Analysis Intelligence Index (rank 41/216) with 89 output tokens/sec, above average intelligence and speed but very verbose
- Supports tool calling, JSON mode, reasoning effort levels (low/medium/high), and prompt caching; weights available on Hugging Face as Step-5-Preview-BF16
Microsoft AI releases MAI Code 1.1 Flash to run locally with 137B/6.8B MoE model
Microsoft AI released a local version of MAI Code 1.1 Flash, a coding-tuned MoE model, for on-device use in GitHub Copilot.- Aimed at developers using GitHub Copilot, VS Code, and the Copilot CLI on high-RAM Windows hardware like Surface Laptop Ultra.
- MAI Code 1.1 Flash is a mixture-of-experts model with 137B total and 6.8B active parameters; the quantized on-device version is ~53GB with a full 256K context window.
- It scores 70.8% on SWE-Bench Verified and 66.3% on Terminal-Bench 2.1 locally vs 72.6/62.9% full precision, with peak memory of 75.5GB at 256K context.
- Copilot adds Auto orchestration routing tasks between local and cloud models, plus explicit local-model selection, rolling out by end of the month.
- Local calls carry zero inference charges, and 1.1 Flash itself is faster (25% more token-efficient) at a quarter of the price of the June 1.0 model.
Fish Audio releases Drama 3, open-weight 35B directable TTS model
Fish Audio released Drama 3, an open-weight TTS model controllable via plain-language acting directions.- Aimed at game dialogue, audiobooks, localization, and conversational agents needing multi-character voice generation with voice cloning across 80+ languages.
- Drama 3 is a 35B-parameter MoE (3B active per token) with natural-language stage directions, mid-sentence voice shifts, word-level audio inpainting, and up to 16 speaker tags per script; open weights on Hugging Face under the Fish Audio Community License.
- Hosted API (drama-3-preview) costs $15 per 1M UTF-8 bytes; SGLang-Omni supports day-one self-hosting, with int4 checkpoint cutting weight memory from 37.0 GB to 20.8 GB and first audio in under 100 ms on one H200.
- Launch follows Fish Audio's $52M seed round; integrations include ComfyUI, Runway, Pika Labs, OpenRouter, Vercel, and OpenArt.
- It moves expressive TTS controls like pacing, emotion, and per-word edits from SSML-style tooling into the script itself for open-weight adoption.
IFM releases Jais 3, an open Arabic-focused model family from 0.9B to 375B parameters
IFM released Jais 3, a new open-weight Arabic-focused LLM family spanning 0.9B to 375B parameters, with weights on Hugging Face.- Aimed at Arabic-language developers and local-LLM users; the family is focused on Arabic and reportedly performs well on agentic tasks.
- Model sizes range from 0.9B to 375B parameters, with weights published in the Jais 3 collection on Hugging Face.
- Jais 3 continues the Jais line originally built by Inception/G42 with MBZUAI and Cerebras, trained on the Condor Galaxy 1 supercomputer.
- The original Jais 13B was trained on 395B tokens (116B Arabic, 279B English/code) under an Apache 2.0 license.
- The release gives builders a strong open option for Arabic NLP, a language underrepresented in mainstream AI models.
🔬 RESEARCH
- 📄 Source for Sipeed MaixStation product page is now live — sipeed.com