AMD serves Kimi K3.1 Pro day-zero on one 8-GPU MI355X node
Frontier open weights break the Nvidia lock at near-reference scores. Plus an AWS agent that slashes context tokens, and GPT Image 2.6 gains SVG output.

Copy markdown
Frontier open weights, now on AMD silicon
AMD's day-zero guide serves the 2.8T-parameter, 104B-active Kimi K3.1 Pro from its MXFP4 checkpoint on a single 8x MI355X node using stock vLLM or ATOM, hitting OCRBench 0.871 against Moonshot's 0.872 reference and 0.991 tool-call JSON accuracy. If you self-host, the biggest open model no longer needs Nvidia, and AMD's tuning cuts decode latency 33.6% (5.78ms to 3.84ms per output token).
An open-source trick to stop burning agent tokens
AWS open-sourced Deep Insight, a multi-agent data-analysis stack on the Strands Agents SDK and Bedrock AgentCore, built around two moves you can copy today: a single write-and-execute tool that collapses a 105-line script into a ~3-line summary in context (~95% fewer tokens per step), and loading skills on demand from a ~500-token catalog instead of 20,000+ tokens of always-on guides (~97% smaller system prompt). The repo and a hands-on workshop are public.
Update: GPT Image 2.6's API splits into Flare and Sunburst
Beyond the GIFs, the 2.6 image API now exposes two tiers - Flare for fast iteration, Sunburst for detailed compositions - and both can emit real SVG files and do markup-guided edits straight from a prompt, not just raster. Token pricing is unchanged from 2.5: $5 per 1M text input, $8 per 1M image input, $30 per 1M image output, with per-image cost read from the response's usage field.