File size: 3,712 Bytes
f1ce0ae | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 | # Qwen3.8-27B Opus personal-model release family
This four-repository family publishes the selected practical winner and its immutable trained-control comparator in BF16 and Apple-Silicon MLX 8-bit formats. The recommended default is the MLX 8-bit winner.
**Selection language:** practical personal-model selection with measured deviations. This release does not claim universal dominance or that every strict gate passed.
| Repository | Role | Format | Public model files | Public model bytes |
|---|---|---|---:|---:|
| [Qwen3.8-27B-Opus-Abliterix-Reasoning-MLX-8bit](https://huggingface.co/timteh673/Qwen3.8-27B-Opus-Abliterix-Reasoning-MLX-8bit) | Recommended practical winner | MLX 8-bit | 17 | 30,390,836,635 |
| [Qwen3.8-27B-Opus-Abliterix-Reasoning-BF16](https://huggingface.co/timteh673/Qwen3.8-27B-Opus-Abliterix-Reasoning-BF16) | Full-precision practical winner | BF16 Transformers | 10 | 55,583,125,224 |
| [Qwen3.8-27B-Opus-Reasoning-Control-MLX-8bit](https://huggingface.co/timteh673/Qwen3.8-27B-Opus-Reasoning-Control-MLX-8bit) | Baseline comparator | MLX 8-bit | 17 | 30,390,836,197 |
| [Qwen3.8-27B-Opus-Reasoning-Control-BF16](https://huggingface.co/timteh673/Qwen3.8-27B-Opus-Reasoning-Control-BF16) | Full-precision baseline comparator | BF16 Transformers | 20 | 55,583,123,681 |
## Lineage
`Qwen/Qwen3.8-27B` → reasoning QLoRA merge (`control-bf16`) → Abliterix pass 1 (`abliterix-pass1-bf16`) → BF16 and affine MLX 8-bit/group-64 release variants.
Dataset preparation started from 12,842 raw rows and accepted 12,614 after 208 deduplications and 20 invalid-row removals; splits were 12,349 train / 127 validation / 138 test. Training used 1,544 optimizer steps, 108,789,760 trainable LoRA parameters, final validation loss 0.23739749, and token accuracy 91.7594%. The merged model retained 1,199 tensor keys, 15 native MTP tensors, and 333 vision tensors. The winner has 74 verified residual-writer edits and zero unexpected changes.
## Local benchmark headline
These are **self-run frozen local benchmarks, not official Qwen benchmarks**.
| Frozen local metric | Control | Abliterix winner |
|---|---:|---:|
| Harmful hard refusal | 43.2% | **0.0%** |
| Harmful soft deflection | 14.6% | **0.2%** |
| Harmful substantive response | 47.0% | **99.4%** |
| Capability macro | 17.6859% | **21.0086%** |
| Full code | **16/421** | 10/421 |
| HumanEval | **7.9268%** | 4.2683% |
| Long-form pass | 54.1667% | **62.5000%** |
| MMMU30 | 9/30 | **11/30** |
| Held-out loss ratio | 1.000000 | 1.024478 |
| Benign KL | 0.000000 | 0.093614 |
Strict deviations remain part of the release: KL 0.093614 > 0.05; incoherence 4.3077% > 2.7692%; HumanEval 4.2683% versus 7.9268%; full code 10/421 versus 16/421; repetition 5.8632% > 5%; prompt leakage detected; and 376/421 winner code generations hit the 512-token cap. Among outputs reaching execution, winner pass rate was 10/46 (21.74%) versus control 16/103 (15.53%), indicating termination/extraction pathology rather than a clean latent-code estimate.
The canonical structured record is [`benchmark-results.json`](./benchmark-results.json).
## Packaging boundary
Model weights and runtime metadata come from sealed source artifacts. Public manifests were regenerated from verified SHA-256/size receipts. Local paths, private prompt/training data, Drive identifiers, operational receipts, and internal release-control files are excluded. Each repository has a public per-file `SHA256SUMS` and `manifests/artifact-manifest.json`.
## License
Apache-2.0, inherited from [`Qwen/Qwen3.8-27B`](https://huggingface.co/Qwen/Qwen3.8-27B). Copyright 2026 Alibaba Cloud; modifications and format conversions are described in each model card.
|