Qwen3.5 dense hybrid — mlx4-g32: a tiny toolchain fixture

A mlx4-g32 storage/reader fixture, built from a tiny random reference checkpoint. Use it to inspect packing, reconstruction, scope accounting, and own-head fidelity comparisons—not to choose a production quantizer.

These weights are untrained, not an assistant, and not a language-quality benchmark. Generated text has no useful semantic quality. No upstream trained weights or training data are implied by an architecture name.

Weight lineage

This is a quantization of a shared native checkpoint, not an independent random initialization and not a fine-tune. HF base_model / base_model_relation: quantized metadata names the real source model.

Exact construction source: malaiwah/qwen3-5-tiny-random-bf16@a430e41d5814ba5e7bfa5ee29d86935aac56d95d.

All 8 derivatives in this family were measured against the same complete source checkpoint identity:

f65faf236cd335fdb5d5e9bd0dc3635b82a47c9e5eb3158a0349d5269b55eebb

That identity binds config and weight bytes; equal architecture names or random seeds are not used as a substitute. Quantized storage cases use round-to-nearest packing and do not claim GPTQ/AWQ/ModelOpt optimization or calibration.

Browse the Matched-Weight Quantization Families collection.

At a glance

Property	Observed value
Artifact role	Child storage fixture; captured role `quant`; classification `lossy`
Serialized top-level weight files	379,316 bytes (0.362 MiB); metadata, tokenizer and evidence excluded
Source-model generated parameters (before packing)	245,378
Fixture vocabulary	272 tokens; independent byte tokenizer, not upstream vocabulary
CPU stack	Python 3.12; Torch 2.11.0+cpu; Transformers 5.16.1; two Torch threads in the recorded capture workflow
Remote model code	Not required: native Transformers class behind QFS

Size is serialized artifact size, not runtime RAM or a capacity/performance guarantee. Packed-array element counts are not model parameter counts.

What is actually included

Complete native Qwen3_5ForConditionalGeneration wrapper: four dense text layers (three GatedDeltaNet, one gated full attention), a real small vision tower/merger, and an untied vocabulary head. Text-only proof; no image/video execution, unequal-axis mRoPE test, or processor.

Storage intervention: mlx4-g32; RTN packing, optimizer-not-run.

Dense Qwen3.5 hybrid; all 12 text MLP matrices, other weights native; not every architecture/export
Output-head policy: native. The comparison replays each side through its own sealed head, including any head error; the reference head is not substituted.
Native BF16 forward; no activation quantization or serving kernels exercised

Only the 12 dense text-MLP matrices are packed. Attention, embeddings, norms, vision and head remain native. GPTQ/AWQ names identify storage conventions here, not a run of their optimizers. MLX layout does not claim execution in the Apple MLX runtime.

The repository file inventory includes config.json, generation_config.json, tokenizer.json, tokenizer_config.json, scope.json. Exact build/runtime/license inventories are linked below. A file may describe historical provenance without being an executable entry point.

Good community uses—and boundaries

Learn how to download pinned artifacts, seal a small synthetic token panel, capture hidden states and replay a full vocabulary head.
Debug model-family adapters, strict tensor loading, storage decoders, or reproducibility tooling without downloading a production-sized checkpoint.
Reproduce a narrowly scoped result, report an adapter/reader regression, and retain the source, panel and runtime identities needed to explain it.

Not established: trained-model accuracy; useful instruction following; quantizer optimization quality; original production-weight compatibility; GPU/NPU/serving-kernel parity; cross-hardware determinism; long-context behavior outside the recorded panel; throughput or paid-compute admission.

RTN is round-to-nearest packing, not optimization. These cases do not run GPTQ, AWQ or AutoRound optimization, calibrated ModelOpt/CT/QAT training, or activation-aware tuning. CPU reconstruction does not execute the original packed GEMM or serving activation arithmetic.

Model family, root dataset and evidence

Exact construction source: malaiwah/qwen3-5-tiny-random-bf16@a430e41d5814ba5e7bfa5ee29d86935aac56d95d. This is the actual source pin; later documentation commits do not rewrite that history. Family fidelity-root dataset repository and storage evidence bundle. The root model is separate; this child must not be registered or represented as the base model. Canonical roots and their evidence are published; registry links distinguish native bases from child/control artifacts.

Recorded own-head result

Measurement	Recorded value
Mean KL(reference ∥ candidate)	2.044495190547015e-05 nats
Scored positions	252; synthetic, not benchmark data
Comparability	`advisory`; reconstruction/activation omissions remain in force
Top-1 agreement	0.9603174603174603 on this panel only

Full comparison receipt includes per-context/tokenwise evidence, comparison gates and disclosures. Receipt self-digest: 2ff0eb9978a31e984d2a184eb0ed8f1074e1bb04493dc0b99841775b2d15b014. No threshold or usable determinism floor is invented here.

Material disclosures from the observed receipt:

HEAD-1d: each side replayed through its own sealed head (reference 35f07c780255, candidate 35f07c780255); head error is inside the measurement, as under HEAD-2, and nothing is substituted. The heads are content-identical.
candidate was captured from a WEIGHTS-ONLY RECONSTRUCTION (affine-weight-reconstruction, output bfloat16). The stored weights were decoded before the model forward; native quantized GEMM and serving activation arithmetic are not measured. Decoder provenance is in the sealed runtime receipt. The comparison is advisory.

Runtime requirements and safe local reproduction

Replay the published evidence without loading a model

This uses QFS's existing NumPy FP64 comparator in a Torch-free environment. It reads the stored hidden states and each side's own head. No model forward, remote model code, GPU, upload or registry mutation is involved. Runtime receipts name the actual backend; last-bit differences from another FP64 implementation are not a new quality claim. Synthetic exact controls remain zero.

Set QFS to a reviewed Quant Fidelity Suite checkout and use Bash:

: "${QFS:?Set QFS to your reviewed QFS checkout}"
WORK=$(mktemp -d)
export OMP_NUM_THREADS=2 MKL_NUM_THREADS=2 OPENBLAS_NUM_THREADS=2
python3.12 -m venv "$WORK/replay-env"
"$WORK/replay-env/bin/pip" install 'numpy==2.5.3' 'huggingface-hub==1.30.0'
"$WORK/replay-env/bin/hf" download malaiwah/qfs-affine-tiny-cpu-format-v1 --repo-type dataset \
  --revision 2a1f5fb39eedbe038eb66e851ef3d19b0daa4493 --include 'raw/candidates/mlx4-g32/evidence/**' --local-dir "$WORK/evidence"
"$WORK/replay-env/bin/python" "$QFS/bin/fidelity_dataset.py" verify \
  "$WORK/evidence/raw/candidates/mlx4-g32/evidence/root" --verify-tensors
"$WORK/replay-env/bin/python" "$QFS/bin/fidelity_dataset.py" verify \
  "$WORK/evidence/raw/candidates/mlx4-g32/evidence/candidate" --verify-tensors
"$WORK/replay-env/bin/python" "$QFS/bin/fidelity_dataset.py" compare \
  --reference "$WORK/evidence/raw/candidates/mlx4-g32/evidence/root" \
  --candidate "$WORK/evidence/raw/candidates/mlx4-g32/evidence/candidate" \
  --out "$WORK/replayed" --device cpu --replay-device numpy --replay-dtype float32 \
  --vocab-chunk 8192 --verify-tensors --own-heads --force-compute

Reconstructed format comparisons are intentionally advisory and normally return exit code 2 while writing a valid receipt. Inspect that receipt; do not silence refusals or interpret an advisory result as a production-quality ranking.

Capture the actual checkpoint

Capture uses a separate pinned Torch CPU environment. The input below is the original token-panel format, not a sealed capture's internal panel/ folder. The native source, tokenizer and model revisions remain explicit. Set AUTHOR to your own HF handle; the dataset repository argument is attribution only and nothing is uploaded by these commands.

: "${AUTHOR:?Set AUTHOR to your Hugging Face handle}"
"$WORK/replay-env/bin/hf" download malaiwah/qwen3-5-tiny-random-mlx4-g32-rtn-format --revision 7e124e9a5672de3770272cbc754b39b8d9b6d508 --local-dir "$WORK/model"
"$WORK/replay-env/bin/hf" download malaiwah/qwen3-5-tiny-random-bf16 --revision a430e41d5814ba5e7bfa5ee29d86935aac56d95d --local-dir "$WORK/source"
"$WORK/replay-env/bin/hf" download malaiwah/qfs-fixture-root-captures-v1 --repo-type dataset \
  --revision f53b204091c988ce4a2161af81886f3018745559 --include 'requirements-capture.txt' --include 'roots/qwen3_5/input-panel/**' \
  --local-dir "$WORK/inputs"
python3.12 -m venv "$WORK/capture-env"
"$WORK/capture-env/bin/pip" install -r "$WORK/inputs/requirements-capture.txt"
"$WORK/capture-env/bin/python" "$QFS/bin/fidelity_dataset.py" architectures prepare \
  --architecture qwen3_5 --model-dir "$WORK/model" \
  --model-repository malaiwah/qwen3-5-tiny-random-mlx4-g32-rtn-format --model-revision 7e124e9a5672de3770272cbc754b39b8d9b6d508 \
  --panel "$WORK/inputs/roots/qwen3_5/input-panel" --tokenizer-root "$WORK/source" \
  --author "$AUTHOR" --dataset-repository "$AUTHOR/qwen3-5-tiny-random-mlx4-g32-rtn-format-capture" \
  --dataset-id "fidelity--$AUTHOR.qwen3-5-tiny-random-mlx4-g32-rtn-format" --out "$WORK/workflow" \
  --role quant --scope-file "$WORK/model/scope.json" --codec int4 --bits 4.0 \
  --reference hf://malaiwah/qwen3-5-tiny-fidelity-root-v1@7e1a755b7c9ee43d2b95c981de177392241e780d

Inspect workflow.json. This launcher executes its exact capture/verification commands and selects the Torch-free interpreter only for the final comparison:

export OMP_NUM_THREADS=2 MKL_NUM_THREADS=2 OPENBLAS_NUM_THREADS=2
"$WORK/replay-env/bin/python" - "$WORK/workflow/workflow.json" <<'PY'
import json, subprocess, sys
workflow = json.load(open(sys.argv[1]))
for step in workflow["commands"]:
    argv = list(step["argv"])
    if step["step"] == "compare":
        argv[0] = sys.executable
    result = subprocess.run(argv)
    if result.returncode:
        raise SystemExit(result.returncode)
PY

Use the actual root at malaiwah/qwen3-5-tiny-fidelity-root-v1@7e1a755b7c9ee43d2b95c981de177392241e780d. The community collection groups models and captures. Custom code, where required above, is explicitly pinned and executed only after your consent; hash verification is provenance, not a sandbox.

Licensing and detailed provenance

The fixture repository retains its mit license in LICENSE. This covers the fixture only according to that exact text; architecture inspiration is not relicensing of upstream weights, cards, code, or configuration. Dependencies retain their own licenses (native Transformers implementation: Apache-2.0). No upstream trained tensors are claimed to be copied.

The exact source license is retained, unchanged, from malaiwah/qwen3-5-tiny-random-bf16. The storage conversion does not grant new upstream rights.

Immutable provenance, historical cards, source inventories and full caveats

Model/card snapshot used while authoring: malaiwah/qwen3-5-tiny-random-mlx4-g32-rtn-format@7e124e9a5672de3770272cbc754b39b8d9b6d508.
Full preserved earlier model card, SHA-256 ef3f227c17eadde9ff8fc3f52ee18065916c4d2ec88d57b205de4cb4af8d1bc9. This is an immutable historical record, including original build pins, measured resource tables, command logs, limitations and failed-attempt provenance. Historical publisher commands are not part of the local-use recipe above.
scope.json.
Retained reproduction tools/dependencies and source-inventory.json. Source paths in recorded argv are historical execution locations; restore the dataset’s raw trees and map them to your own workspace. Do not substitute moving upstream branches or an unreviewed publisher workflow.

Selected original provenance fields (full tensor/component evidence remains in the linked inventories):

{
  "source": {
    "repository": "malaiwah/qwen3-5-tiny-random-bf16",
    "revision": "a430e41d5814ba5e7bfa5ee29d86935aac56d95d"
  },
  "format": "mlx4-g32; RTN packing, optimizer-not-run",
  "classification": "lossy",
  "architecture_restrictions": [
    "Dense Qwen3.5 hybrid; all 12 text MLP matrices, other weights native; not every architecture/export"
  ],
  "activation_omissions": [
    "Native BF16 forward; no activation quantization or serving kernels exercised"
  ],
  "head_policy": "native"
}