Skip to content

Run frontier MoE models on hardware you already own β€” pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦

C

docs: adopt dev integration branch β€” PRs target dev, batched into stable main (ZacharyZcR's suggestion)

JustVuggf38e74efe7b1
diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
index a85b8602..0755056b 100644
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -2,6 +2,17 @@
Keep changes focused and preserve Colibri's dependency-free default CPU path.
+## Branches
+
+- **`main`** is the stable branch. It's what users clone, and it stays known-good
+ (engine always passes the token-exact oracle: `SNAP=./glm_tiny TF=1 ./glm 64 16 16`).
+- **`dev`** is the integration branch. **Open your PR against `dev`.** Reviewed PRs
+ land there first; once a batch is tested and stable, the maintainer fast-forwards
+ it into `main`. This keeps `main` clean instead of taking every PR one at a time.
+
+Every PR β€” on either branch β€” is reviewed for a clean build (0 warnings), the oracle
+(32/32 TF + 20/20 greedy), and its own targeted validation before merge.
+
## Local checks
Run the lightweight checks locally: