Qwen3.6-35B-A3B Heretic — selective GPTQ-Pro W4

Public experimental release of a recovered, byte-verified checkpoint. This is not a uniform-group GPTQ model or a claim of production qualification.

Source and allocation

Source: llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved at immutable revision 9599ac17d26daf33daf0fdd8f6c897ff4c6dc89a. Local source README, config and tensor-index hashes were checked against that revision. Apache-2.0 attribution and license are retained.

The 40-layer, 256-expert-per-layer model stores 30,464 routed expert projections as INT4: 19,968 at group size 64 and 10,496 at group size 32. Layer 10's 256 routed down projections remain BF16. Sensitive attention, recurrent-state, shared-expert, router, norm, embedding and output components remain in their exported precision. All 333 vision and 19 MTP tensors are preserved. Tensor payload is approximately 24.08 GiB in 13 safetensors shards. See PROVENANCE.json for exact counts, hashes and allocation.

The original export process failed during serialization. A separate checkpoint was reconstructed without requantization; all 123,077 tensor hashes and all 13 shard hashes matched the independently verified reconstruction. Failure history was preserved rather than relabeled successful.

Measured validation — limited scope

Twelve deterministic checks passed: arithmetic, sorting, counting, logic, sentiment, JSON, English/Spanish translation and Unicode copying. Prefill and decoded-token logits were checked for finite values.

Three short raw-text likelihood comparisons (44 predicted tokens total) used exact original BF16 routed-expert weights as the reference in the same model graph. Weighted mean NLL increase was 0.01036229 nats/token; worst sample increase was 0.02683258. This is a small diagnostic, not an external coding benchmark, a representative perplexity corpus, or proof of general quality equivalence. Thresholds were recorded before scoring. No 250k-context test, vision test, MTP test or production-speed benchmark is claimed.

Runtime caveats

Validation used a local GPTQ-Pro native kernel with BF16 computation retained outside quantized expert operations, which used FP16. The mixed group sizes and BF16 exclusions must not be removed or relabeled to force another loader to accept the files. Stock fused-MoE vLLM compatibility is not qualified.

The test host silently corrupted direct cross-GPU tensor copies: 0/12 direct-copy checks passed, versus 12/12 exact host-staged copies. The original BF16 control also failed with direct copies. Recovered-model inference passed after using a process-local host-staged transfer workaround. No driver, BIOS or IOMMU settings were changed. Validate transfers independently on the deployment host; NCCL settings alone do not necessarily control ordinary PyTorch tensor copies.

Vision and MTP weights being present is not proof that those modalities execute correctly in an arbitrary runtime. This derivative also inherits the upstream ablation behaviour; evaluate suitability and access controls for the intended use.

Only reviewed model assets and these validation/provenance records are published. Calibration data, private logs and local environments are excluded.