[brunobergher]

Adds Recursive Flint to the models Lamp can download and run on this Mac (Settings → AI Providers → Native Mac AI, and the welcome's picker).

• Catalog entry (LocalModels.catalog): RecursiveSI/flint-int4-gguf, the official Q4_1 GGUF (26 GB, needs 48 GB of memory per its card). Text only, so the blurb says it can't see screenshots. It sits after Qwen3.6 35B-A3B, which stays the recommended pick on 48 GB+ Macs because look_at_app needs vision.
• llama.cpp b11346 → b11814: Flint's architecture landed in models: Add Flint model (FlintForCausalLM) ggml-org/llama.cpp#30928 (merged yesterday); b11814 contains it. Checksums are the release's published digests, and build.sh verified the arm64 one.
• A context of its own: Lamp's prompt alone is over 20K tokens, so the usual 32K overflowed after a few turns (request (33506 tokens) exceeds the available context size). LocalModelSpec.contextSize gives Flint 131,072 (its cache is about 3.4 GB there: all but 13 of its 49 layers look back only 513 tokens). The size goes to llama-server through a presets file (models.ini, [*] for the rest at 32K, a section per model), since a -c on the command line overrides presets, and to pi's models.json.

Tested on a 64 GB M-series Mac, with the bundled llama-server started using Lamp's own router flags (--models-dir, --jinja, -ngl 999, -c 32768):

• Flint loaded and answered an OpenAI tool-calling request with a correct tool call, at about 105 tokens/s.
• With the presets file, a 51.5K-token request to Flint answered correctly; the server used about 28 GB in total.
• Qwen3.5 4B GGUF still loads and calls tools on the new server (regression check).

Engines: only pi runs local models, so pi is the only engine this touches. I didn't run a full build in Lamp with Flint selected.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
forge-updates	Ready	Preview	Oct 31, 2026 7:27am UTC
lamp	Ready	Preview	Oct 31, 2026 7:27am UTC
lamp-updates	Ready	Preview	Oct 31, 2026 7:27am UTC

[roomote-roomote]

No new code issues found. See session

Reviewed the merge with current main, including conflict resolution in server startup and pi registration. Flint's 128K presets remain intact alongside the API key and worker environment; the b11814 pin feeds the new signed runtime-asset packaging and installer consistently.

CI: current-head macOS Swift build and website checks passed; regression tests were skipped by the PR workflow. Full app packaging and native Flint inference were not exercised in this review.

_{Reviewed 14e6809}