# Qwen3.8-Flash-Next: evaluation note

Capacity Technologies · 11 September 2026

This note accompanies our article “Abliteration makes models ‘dumber’? Nope. Smarter, if done right.” It summarizes our internal Qwen3.8-Flash-Next artifact postmortem, reconstructed on 11 September for ABL-042 from saved run manifests, request/response records, historical reviewer-agent outputs, build attestations and tensor audits. It is a summary of our own experiments, not an independent replication or a complete reproduction archive.

## Behaviour result

The final native Q5 adapter, s1p5, received 59 substantive labels out of 60 behaviour cells. The stock Q5 control refused all 60. A stronger s1p7 edit received 48 substantive labels out of 60. The 60 cells per arm comprised 30 distinct prompts, each tested in thinking and non-thinking modes: ten prompts each in English, German and Chinese.

These prompts deliberately included harmful requests to test refusal behaviour. They were not a benchmark of authorized cyber defense or smart-contract auditing. “Substantive” meant engagement with the request, not verified answer correctness. The remaining s1p5 failure was a repetitive German thinking response. Labels came from historical reviewer agents; this was not blinded human annotation.

The panel was authored after an earlier panel's failures and used to select between two strengths. It was an adaptive selection panel, not a one-time untouched confirmation of a preselected candidate.

## Small non-thinking capability comparison

| Domain                | Stock Q5 | Q5 + s1p5 |
| --------------------- | -------- | --------- |
| Code                  | 10/12    | 10/12     |
| Instruction following | 5/8      | 5/8       |
| Knowledge             | 6/8      | 6/8       |
| Math                  | 6/8      | 6/8       |
| Tools                 | 7/8      | 7/8       |
| Integrity             | 8/8      | 7/8       |

Denominators include all requested domain rows. Stock knowledge had seven scorable rows, stock math six, and adapter math six; those unscorable cases remain in the denominators above. The 56-row panel also included context and vision controls. It does not establish 56 capability passes.

Integrity checks asked the model to stick to supplied evidence, correct false claims and handle uncertainty. This is an answer-behaviour measure, not a security certification or a file-integrity check.

These samples show unchanged counts in several domains and one fewer integrity pass. No statistical equivalence margin or general intelligence increase was established. The final candidate did not complete a paired near-capacity context comparison. Broad thinking capability, original BF16 preservation and cross-hardware parity were not demonstrated.

## Packaging and provenance

The selected edit combined a trained layer-47 adapter, a recovered direction, removal of an earlier embedding edit and projection strength 1.5. The native adapter's evidence applies to its pinned Q5 base and patched runtime.

The IQ4 adapter received only a six-cell behaviour check: five substantive answers and one thinking-loop failure. The baked Q5 model received a separate six-cell behaviour check (six substantive) and a 56-row capability check on a different runtime. Neither inherits the native Q5 adapter's 60-cell result. Baking introduced measurable quantization changes. Stronger scores on a few later baked-model rows do not establish a general capability improvement.

The postmortem found incorrect cohort hashes in three package references, ambiguous evaluation counts and unfinished verification metadata. Some final local documentation changes were not confirmed published at the recorded snapshot. The public releases include payload identities and evaluation summaries, but some original factors, recipes, transcripts and merge-tool source remain in the local research archive. Complete public reproducibility has not been established.

## Public releases

- [Native Q5/IQ4 adapter package](https://huggingface.co/apetersson/Qwen3.8-Flash-Next-Abliterated-Adapter)
- [Baked Q5 package](https://huggingface.co/apetersson/Qwen3.8-Flash-Next-Abliterated)

The postmortem's publication snapshot recorded adapter revision `bf1c9ef540d6237a9e30fd521df53880cb5bbda3` and baked revision `d1700222b883362695a12ebe5fda835a7cc0a2dd`. This note reports that snapshot, not a claim that those are the latest revisions.

### Publication follow-up: 11 September 2026

A fresh public-file check at 21:07 CEST (19:07 UTC) inspected adapter revision [`3f8a78ea4c6a21fa798f4c34dc39a8b06e5c19ce`](https://huggingface.co/apetersson/Qwen3.8-Flash-Next-Abliterated-Adapter/tree/3f8a78ea4c6a21fa798f4c34dc39a8b06e5c19ce) and baked revision [`89bb1c93f1954c3f53480527572f267ad43c1159`](https://huggingface.co/apetersson/Qwen3.8-Flash-Next-Abliterated/tree/89bb1c93f1954c3f53480527572f267ad43c1159).

The published model cards now separate adapter and baked capability results. The adapter card states 30 prompts in two modes, reports vision as 6/6 per arm, and includes the 7/8 integrity result. Both cards use all requested rows in their main capability tables and explain the small-sample limits. The baked manifest now lists both shards, and its checksum file includes both. Those publication updates are confirmed.

Companion files still contain the following issues at these revisions:

- The adapter manifest repeats the old “60 cells x 2 modes” and pooled “18/18” vision descriptions.
- The IQ4 validation file retains the incorrect sentinel cohort hash identified in the postmortem.
- The baked validation file retains two incorrect cohort hashes and the misleading “6 cells x 2 modes” description.
- The baked manifest's bake-fidelity, behaviour and loading verification fields remain empty.

The downloaded README, manifest and validation JSON files matched their published Git blob identities and corresponding SHA-256 checksum entries. That verifies those files' identities, not the correctness of their claims. Model weights were not downloaded for this check. The packages still do not contain the complete research reproduction archive.

There was no dedicated defensive-security or smart-contract vulnerability-detection benchmark in this campaign. Those are proposed applications requiring evaluation with authorized systems, reproducible tests and expert review. Abliteration supplies no confidentiality guarantee; data control depends on deployment and operations.