Apple's LensVLM-9B: open VLM that reads docs as compressed images
The weights are on Hugging Face with GGUF and MLX quants — plus persistent disks for Vercel sandboxes, Prime's 365K-env RL fleet, and an MCP hard-fork.
Copy markdown
Read a page as a thumbnail, zoom in only where it matters
Apple pushed LensVLM-9B's weights to Hugging Face — a Qwen3.5-9B model that ingests text rendered as compressed images and uses learned tools to expand only the regions it actually needs to full resolution. It holds accuracy at ~4.3x effective compression and still beats baselines out to ~10x, so long-document QA, contracts, and whole codebases burn a fraction of the vision tokens; community GGUF and MLX builds already run it locally, Apple Silicon included.
Vercel sandboxes finally keep their disk
Drives for Vercel Sandbox hit public beta: persistent volumes you mount into a box and reuse across runs, from a 1 TiB default up to 16 TiB, four per sandbox, with read-only point-in-time snapshots. At $0.05/GB-month (reads $0.0015/GB, writes $0.004/GB) you can cache models, datasets, and dependency trees — or keep an agent's on-disk memory — instead of rebuilding them on every cold start.
Prime Sandboxes: 365K RL environments at ~4 cents an hour
Prime Intellect opened Prime Sandboxes — hardware-virtualized Linux microVMs with their own guest kernel and Docker Compose, built for agentic RL and wired to a registry of 365,000+ prebuilt environments. Launch pricing through Dec 22 is $0.02/vCPU-hr and $0.0125/GiB-RAM-hr — a 1-vCPU/1-GiB/32-GiB box lands near $0.039/hour — with 1,024 concurrent sandboxes on new accounts.
MCP gets a community hard-fork
OpenMCP (omcp) hit Show HN — a code-first, community-run hard fork of the Model Context Protocol that stays wire-compatible with existing MCP servers and clients but ships changes on passing tests rather than committee sign-off. If the spec has felt slow, it's a drop-in you can point your current tools at without a rewrite.