colibri v1.2.0

Pure-C MoE engine that streams experts from disk — run a 744B model on a ~25 GB box. Highlights since v1.1.1:

Hardware & correctness

• Fix OOM on integrated / unified-memory GPUs — GB10 / DGX Spark (#653)
• Recognize AMD/ROCm in doctor and resource_plan — a HIP build is no longer reported CPU-only (#662, #663)
• Chat stop-set fix (#633 / #381); async packed-int4 parity (#632); stable KV slot per conversation (#634, #639)

Performance

• AVX2 matmul_e8 — fmt=6 (E8/IQ3) was 92% of decode on the scalar kernel (#654)
• Native SIMD fmt=6 encoder, 15× over numpy, byte-identical (#655)
• AVX-512 matmul_i3 for fmt=5 int3-g64 (#661)

Web & docs

• Web dashboard with the live Expert Atlas (#641); new landing page (#645)
• README / quickstart point to the gs64 model container (#642)

Verify your download against SHA256SUMS.txt.