GLM-5.3-Flash runs local: a 320B, MIT-licensed vision MoE

18B active, 1M context, self-hostable via GGUF. Also: DeepSeek eyes Huawei Ascend, Claude Sonnet 5.5 ships, Cloudflare's cf CLI is agent-first.

Nowline SEP 30 7:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • GLM-5.3-Flash now runs on llama.cpp

    Z.ai's 320B / 18B-active multimodal MoE — text + vision, 1M context, MIT-licensed — just got GGUF support, so you can self-host it with no API. Budget ~100GB of RAM/VRAM at 1-bit or ~150GB at Q4; the payoff is an offline, near-frontier vision model for document- and chart-reading agents you fully control.

  • DeepSeek aims its stack at Huawei Ascend

    DeepSeek has reportedly open-sourced Ascend-targeted tooling — DeepGEMM matrix kernels plus a TileLang-style CUDA alternative — to run its models on Huawei silicon. If NVIDIA supply or pricing is boxing you out, a maturing non-CUDA path is worth tracking.

  • Claude Sonnet 5.5 shipped, in case you missed it

    Anthropic's Sonnet 5.5 is live at $2/$10 per million tokens (1M input, 128K output), scoring 70.6% on Terminal-Bench 4.0 versus 43% before — roughly 30% faster and using fewer tokens per task. Haiku 5.5 is next, and Sonnet 5 stays at the same price. For Claude Code users it's a cheaper, faster default.

  • Cloudflare's cf CLI is built for agents, not you

    Cloudflare rebuilt Wrangler into cf, one command-line tool exposing ~3,000 platform API operations with natural-language search and JSON-first output. It's designed for an AI agent to drive, so your agent can provision Workers, D1, and DNS end-to-end without hand-written scripts.