mudler/timestral-cpp-gguf

Timestral GGUF models for timestral.cpp

This interim release contains fourteen GGUF files for native CPU forecasting with timestral.cpp. Inference requires no Python or libtorch. Conversion uses the pinned upstream checkpoints listed in manifest.json.

Available models

Directory	Available formats	Validation
5M/	F32, F16, Q8_0	All three pass storage, schema, lineage, and byte-encoding audits. F32 passes trained parity; reduced-format forecast quality is not yet evaluated.
28M/	F32, F16, Q8_0	Storage, schema, source-byte association, RoPE, and encoding audits pass. Native inference and forecast quality remain pending.
123M/	F32, F16	Storage, schema, source-byte association, RoPE, and encoding audits pass. Native inference and forecast quality remain pending.
209M/	F32, F16, Q8_0, Q6_K, Q5_K, Q4_K	All six pass cohort/storage admission. Full quality evaluation completes, but the Q8 mean-deviation gate fails.

All four upstream models now have converted artifacts in this repository. Native qualification is incomplete for 28M and 123M. 123M Q8_0 is unavailable: its 1712-wide FFN matrices do not satisfy the unchanged 32-value block requirement. K formats for 5M, 28M, and 123M are incompatible with the unchanged complete weight policy. No padding or selective exclusions are used.

Quality and performance limitations

The qualified optional standalone MKL profile uses runtime 68a5a53 and the documented SDK version. Results do not establish parity or quality for every backend, SDK, or machine. No proprietary SDK libraries are included here.

The 209M quality evaluation covers 79 items, six formats, 474 native calls, and 79 fresh F32 anchor pairs. All formats have complete coverage. Immutable closing checks pass.

209M format	Aggregate mean normalized deviation from F32	Absolute aggregate WQL change
F32	0	0%
F16	0.00015529176950670028	0.00274317535567626%
Q8_0	0.005657737466196431	0.04116458747472307%
Q6_K	0.013512040093425381	0.06180385023158404%
Q5_K	0.02008887882387394	1.3580854256823647%
Q4_K	0.033048053192290705	2.770376555053301%

F16 passes its mean-deviation bound of 0.001 and WQL-change bound of 0.5%. Q8_0 fails its mean-deviation bound of 0.005. Its WQL change passes the 1% bound. WQL changes above are already percentages. Normalized deviation is not raw percentage forecast error. Strict mean-error ordering passes. K formats have no separate loss bound in this contract. MASE degeneracies remain reported, not excluded.

Formal latency acceptance fails at both tested thread counts. Memory bounds pass. Review and optimization remain pending. This interim publication does not waive any acceptance gate.

See the complete quality report, F32 replay, and runtime documentation. Use the native CLI or C ABI documented in the code repository. GGUF storage does not imply compatibility with unrelated GGUF runtimes.

Licence

These weights use Mistral AI's Modified MIT License, not unrestricted MIT. Read the complete original licence, including its attribution, disclaimer, and revenue restriction. The licence restricts rights when your company or employer exceeds USD 20 million in consolidated monthly revenue for the preceding month. The restriction also covers derivatives. Contact Mistral AI for a commercial licence when required. Keep the original notices with copies and derivatives. This repository does not grant additional rights.

Integrity and provenance

manifest.json records exact sizes, SHA-256 hashes, upstream revisions, and runtime/reference identities. SHA256SUMS contains checksums for all fourteen GGUF files. F32/F16/Q8_0 use the pinned converter; 209M K formats use the reviewed native quantizer. The 5M F32 file is byte-identical to the previously accepted trained-parity artifact.