Qwen3.8-Max: A New Bar for Coding and Cowork
Alibaba shipped the full Qwen3.8-Max release two weeks after previewing it with no benchmarks — the production model runs on a new Qwen3.5 architecture at 2.4 trillion parameters (95 billion active) and is live now through Alibaba's cloud API, with open weights following next week on Hugging Face and ModelScope. Alibaba's own benchmark table puts it well ahead of its predecessor Qwen3.7-Max (Terminal Bench 2.1: 86.6 vs 74.5) and in the range of the current Anthropic and OpenAI flagships on coding tasks, though none of those figures have been independently verified yet.
Sources & depth
MiniMax H3 Day-0 Support in ComfyUI: Open Weights, Native Audio, and 2K Video
MiniMax released H3, its first open-weights video model and third generation overall, with day-zero support in the ComfyUI generation tool. The model takes text, image, video, or audio as input and generates clips up to 2K resolution and 15 seconds long, with native stereo audio produced jointly with the video rather than added afterward. Its most notable engineering detail: quantization and pruning cut the model's memory footprint by two-thirds, from 123.6GB to 42.5GB, letting it run on a consumer GPU like an RTX 3060 with dynamic VRAM offloading.
Sources & depth
How we built a realtime system for responsive voice AI in six months
OpenAI published a detailed engineering account of GPT-Live, the voice system it launched in July. The core change was removing the turn-detection step altogether: a full-duplex model now listens and speaks at the same time, handing deeper reasoning off to a frontier model asynchronously so it never blocks the audio path. The team also rewrote the inference and media frontend in Go, pushing the new system's p95 latency down to match the old system's p50, and introduced WARP, a proposed IETF protocol that cuts session start from six network round trips to one.
Sources & depth
Kimi K3 + DeepSeek-V4 Flash 0731 + Deep Research + Parallel Chat
Unsloth's latest release adds local fine-tuning and inference support for two of the week's biggest open-weight drops — Moonshot's Kimi K3 (2.8 trillion parameters, 104 billion active) and DeepSeek's V4-Flash-0731 — via Dynamic GGUF quantizations with automatic multi-GPU detection. The release also ships two new local-agent features: Deep Research, which turns a local model into a research agent that plans, searches the web, and writes a cited report, and Parallel Chat, for running multiple conversations at once. AMD and Intel GPU support was expanded alongside it.
Sources & depth