Qwen3.5-0.8B Text — YVEX source-faithful GGUF
A text-only, source-faithful GGUF built by YVEX from
Qwen/Qwen3.5-0.8B at 2fc06364715b967f1860aea9cf38778875588b17.
Upstream weights and tokenizer are by Qwen. The included LICENSE retains Apache-2.0 terms.
YVEX changes the container, tensor names and execution metadata, and selects the text projection: 320 tensors (284 BF16 and 36 F32), preserving the precision and values of each selected source tensor. The 153 vision and 15 MTP tensors are excluded. The output head reuses the tied token embedding; no duplicate head is stored. This is not a low-bit quantization or the complete upstream multimodal package.
Exact artifact
- File:
qwen3.5-0.8b-source-faithful.gguf - Size: 1,528,566,432 bytes
- Full-file SHA-256:
0c5776eb6b1f2abb3a35f2324aabc4d8b7693856650b799e88161f7167feded6 - Container: GGUF v3, 45 metadata fields, 320 tensors.
- Exact tensor descriptors:
tensor-manifest.json. - Source, producer, artifact identity and bounded evidence:
manifest.json. - Producer: YVEX
17e5e0a14fdb934cdb19a4cc720fd798d13f5283, treebfa57cf065c44c82ac78f9ff3743d57eadc89935.
Earned execution scope
Tested on macOS arm64, Apple M5 Pro with 24 GiB unified memory, using the explicitly selected YVEX CPU backend and a compiled runtime binding for this exact artifact. Two raw-text prompts completed eight greedy new tokens at context capacity 32 and prefill chunk size 1. Both token sequences exactly matched an independent upstream Transformers CPU oracle using the same source tensor precision. Native YVEX detokenization and generated-text digests were also checked.
Prompt: The capital of Italy is
Rome.
The capital of France is
Generated token IDs: [21047, 13, 198, 760, 6511, 314, 9338, 369].
Prompt: def add(a, b):
return a + b
def
Generated token IDs: [198, 262, 460, 264, 478, 292, 271, 727].
The normal YVEX CLI host also loaded and unloaded the exact model and stopped cleanly in an isolated runtime namespace. Software regression lanes passed at the exact clean producer revision: macOS CI 120, macOS native 14, Linux CPU CI 120 and Linux native 14.
These are bounded execution and numerical observations, not a model-quality benchmark. The tested prompt policy is verbatim raw text. Chat templates/conversation mode are not qualified by these results. No full-model logits tolerance, long-context, vision, MTP, throughput or production readiness claim is made.
Use with YVEX
This artifact carries YVEX exact-target and physical-plan metadata. Use the producer revision or a compatible YVEX revision and its normal artifact verification, compilation/binding and deployment flow. Model bytes alone do not replace the compiled runtime binding required by YVEX. The qualified execution command is:
yvex bench transformer generate \
--target qwen3.5-0.8b \
--artifact /path/to/qwen3.5-0.8b-source-faithful.gguf \
--runtime-binding /path/to/exact.yvex-runtime-binding \
--backend cpu --generation-mode target-only --strategy greedy \
--text 'The capital of Italy is' \
--context-capacity 32 --max-new-tokens 8 --prefill-chunk-tokens 1 --output json
See the YVEX macOS qualification report for source acquisition, compilation, native qualification and limitations.
YVEX separately has a qualified Metal primitive foundation on the tested Mac. This model was not executed end to end on Metal. The model card does not imply compatibility with llama.cpp or other GGUF consumers merely because they read GGUF. An independent official GGUF reader accepted the emitted container structure.
Redistribution
Redistribution and derivative use follow the included Apache-2.0 LICENSE. Retain the license and applicable notices, credit the upstream Qwen source, and disclose changes. No Qwen or hardware-vendor endorsement is claimed.