cstr/Josev-2B-SystemOne-GGUF

Josev-2B SystemOne packed F32 GGUF (CrispEmbed)

Byte-preserving F32 packing of mayflowergmbh/Josev-2B (backbone, joint head and tokenizer in one file) for CrispEmbed's opt-in SystemOne API. No tensor arithmetic or quantization was performed. Upstream license and notices are kept unchanged under upstream/.

• File: josev2b-systemone-f32.gguf, 7659180448 bytes, SHA256 72684e39d0eed526b40c5150c78b2d881d324c39e16af48060b3d741f526ccf2
• Evidence source commit: cd0ec9968ac29b05426931c8d1dccd2860f76927; packing run 40456235976; consumer run 40463407332
• Verified: 227 original stages and four complete typed responses in C/Python/CLI/HTTP/Rust/Dart on Linux x64 CPU.
• Not verified: task quality, other platforms, GPU, quantization. All seven model gates stay default-off.

Modified from the upstream checkpoint by repackaging only (Apache-2.0 section 4).

Experimental quantized variants (owner decision 7 November 2026)

Experimental: not quality-qualified. Use the F32 file above for exact typed responses. The quantized files were produced from the F32 file of revision bffe62522f826b4c33f2e731d5b870ced5150a17 with CrispEmbed's C quantizer (tensor types changed; no weights trained or merged; Apache-2.0 section 4(b)).

Quality vs the frozen original reference (four frozen typed requests, portable hosted CPU, receipt tests/results/josev2b/q4k-sweep-41615751516.json):

File	Precision	Bytes	Head-logit cosine (min)	Final-norm cosine (min)	Option argmax exact	Complete typed JSON exact	Max probability error	Strict 0.002/0.99999 stage budgets	Status
josev2b-systemone-q8_0.gguf	Q8_0 (q8-all; Q4_K/Q8_0/F32 tensors 0/150/292)	3,630,289,376	0.99995 (bar > 0.97)	0.99944	4/4	1/4	0.0059	failed (200 stages)	experimental
josev2b-systemone-q4_k-down8.gguf	Q4_K_MIX (all-q4-down-q8; Q4_K/Q8_0/F32 tensors 126/24/292)	3,095,515,616	0.99739 (bar > 0.94)	0.94911	4/4	0/4	0.0358	failed (203 stages)	experimental

• Option choices match the original on 4/4 cases, but the rounded four-decimal probabilities and scores move, so the complete typed JSON is not exact (Q8_0 1/4, Q4_K_MIX 0/4).

• Not measured: independent task-quality panel, public-client (C/Python/CLI/HTTP/Rust/Dart) runs of the quantized files, other platforms, GPU (the Josev runtime is CPU-only), latency/RSS.

• Runtime: the same seven default-off Josev gates as F32 plus CRISPEMBED_JOSEV_QUANT=1 and CRISPEMBED_QWEN35_QUANT=1.

• Q8_0: 150 projections Q8_0, 292 critical tensors F32; 3,462 MiB. Head-logit cosine min 0.99995, final-norm 0.99944 (four frozen cases, portable hosted CPU).

• Q4_K_MIX: Named policy josev_q4k_down8_v1 (docs/josev2b-q4k-sweep.md): all decoder projections Q4_K except the 24 MLP down_proj at Q8_0; embedding, norms, head and other non-projection tensors F32. 2,952 MiB vs 3,462 MiB Q8_0. Head-logit cosine min 0.99739, final-norm 0.94911; option argmax 4/4, complete typed JSON 0/4 (rounded probabilities move).

• josev2b-systemone-q8_0.gguf: SHA256 fc1663bbd050938ff0bdb1b20ca52bd9ff7579dca63c0f7f9a15b3c3b7e36b85

• josev2b-systemone-q4_k-down8.gguf: SHA256 adc7fa473febfef3fac10eac6fa0bebb6f68932675d1613e24706a3b74d5e582

Built by CrispEmbed commit 84b5fbedd8a1b07dd5e8a6582cc02010be75b11a (workflow josev-publish.yml, protocol docs/josev2b-q4k-publication-v1.json).