Verified end to end on the full released 1.56 TB checkpoint, and made substantially faster, with byte-identical output preserved at every step. The first-run experience, broken on a clean clone, now works.
Faster (output unchanged)
- Fused matmul kernels (fp32, bf16, MXFP4): about 8x less per-token compute, scalar and AVX2 paths bitwise identical.
--preset auto: sizes the trunk and expert-cache budgets from the machine's own free RAM, trunk-first.- Chunk-union prefill: fetches each unique routed expert once per chunk, about half the expert bytes on a prompt, generated token bit-identical.
- Conversation resume (
--save-state/--load-state): a second turn resumes instead of re-reading the prompt, 3.9x faster. - KDA recurrence parallelised over heads;
--spec Nspeculative decode, exact by construction.
Correctness and first run
- All shell scripts are committed executable; the first documented command works on a clean clone.
download-model.shpins an immutable revision with checksum verification and refuses without free space.k3-doctor.shno longer fails a machine that can build and test the engine.- The config-refusal fixtures are gated in
make test, ctest, and CI. - Two silent-corruption paths (the MLA KV overflow and the single-slot trunk reader) now abort or are prevented.
Verification
Token ids byte-identical to the reference from 8 GB to 128 GB+, oracle gates untouched, make test and ctest green.
Full log in CHANGELOG.md; measured speed in docs/data/speed-2026-08.md.