Independent evaluations will determine whether the abliterated Qwen3.8-27B FP8 checkpoint reduces harmful-request refusals to near zero while preserving general benchmark capability within roughly 1.3 points of the base model.

state: expiredheat: lowuncertainty: highknownscott: lowopen-models model-safety local-inferenceAlibaba Qwen

What is this?

KridgeDookie has published a 27B-parameter multimodal derivative of Qwen/Qwen3.8-27B-FP8 that uses “abliteration” to suppress refusal behavior. Its deployment page claims zero refusals across two internal prompt sets, 23 of 24 coherence checks passed, and retained vision support; the supplied research snippet frames capability preservation as the key tradeoff in evaluating abliteration. The material does not provide an independent evaluation of the derivative or directly substantiate the claimed 0–6% refusal range and sub-1.3-point benchmark delta, so those remain model-card claims awaiting replication.

Why it matters to Scott

Scott’s Evaluation-Driven Development and independent-verification pages already require repeatable external evidence before accepting changed model behaviour; this case currently adds only unreplicated model-card claims. It is relevant to the radar’s established open-model evaluation and safety-capability tradeoff territory, but without independent results it would not yet change what Scott builds or argues.

ip:concept.evaluation-driven-developmentip:concept.capability-auditip:concept.mechanically-different-verifiersradar:concept.open-modelsradar:concept.model-evaluationradar:concept.ai-safetyradar:compressed-llm-fidelity-safety-gap

queries asked of Scott's wikis

abliteration refusal vectors capability preservation
open-weight safety controls after weight release
uncensored local models safety asymmetry
independent evaluation of modified checkpoints
local inference model governance and sovereignty
benchmarking refusal reduction versus capability loss

Measured heat

no measured readings yet — the hourly heat pass fills this in

How the heat travelled

no chain yet — the hourly chain pass fills this in

Evidence (5) — ⭐ canonical anchor

source	object	author	score	comments
🟠 reddit	Qwen3.8-27B abliterated FP8: refusal 64–99% → 0–6%, and MMLU/GSM8K move less than 1.3 points LocalLLaMA	niacolhealth	23	12
🟧 echo.other ⭐	The model card is the source of the Reddit post’s figures. It reports refusal falling from 64–99% for the official FP8 base to 0–6% after ab	OrcaRouter	—	—
🟠 reddit	Huihui-ai Qwen 3.8 Ablit Available LocalLLaMA	Frizzy-MacDrizzle	28	36
🟠 reddit	Eccomerce frontend and a simple game one shooted by Qwen 3.8 27B LocalLLaMA	Tiny-Assumption4263	9	7
🟠 reddit	Long Review: Qwen 3.8 27B is VERY good at tapping into it's real-world knowledge. It's "overthinking" brings it to Sonnet level performance with the potential for Opus level results. LocalLLaMA	maxwell321	420	94

Interpretation history

2026-08-17T15:35:25Z
Repeated comment refreshes have exhausted the near-term signal without producing any checkpoint-specific refusal test or controlled capability comparison. The case can fade from active tracking and be reopened if an independent evaluation appears.
2026-08-17T14:43:28Z
The comment refresh adds only more qualitative impressions of standard Qwen3.8 and no evidence tied to the abliterated FP8 checkpoint. Repeated amplification is not narrowing the uncertainty around refusal suppression or capability retention.
2026-08-17T14:08:55Z
The latest comment refresh is repetitive qualitative discussion of the standard Qwen3.8 model, not an independent test of the abliterated FP8 checkpoint. The claimed refusal reduction and capability retention remain uncorroborated.
2026-08-17T12:46:47Z
The refreshed comments are further qualitative amplification of standard Qwen3.8 capability, not checkpoint-specific evidence about the abliterated FP8 model. Without controlled refusal testing or comparative capability results, the claimed tradeoff remains uncorroborated.
2026-08-17T11:31:35Z
The refreshed comments remain qualitative impressions of the standard Qwen3.8 model and add no controlled test of the abliterated FP8 checkpoint. This is repetitive amplification, not evidence for the claimed refusal reduction or capability retention.
2026-08-17T10:35:13Z
The refreshed discussion remains qualitative commentary about the standard Qwen3.8 model and adds no checkpoint-specific refusal test, controlled capability comparison, or reproducible artifact. This is further repetitive amplification rather than independent corroboration of the claimed abliteration tradeoff.
2026-08-17T09:30:29Z
The refreshed comments remain conflicting, method-free impressions of the standard Qwen3.8 model rather than tests of the abliterated FP8 checkpoint. This is repetitive amplification, leaving the refusal-versus-capability claim uncorroborated.
2026-08-17T08:30:30Z
The new hands-on report strengthens the view that base Qwen3.8-27B is practically capable, but it evaluates a different quant and provides no checkpoint-specific refusal testing or controlled capability comparison. The abliterated model’s claimed safety-capability tradeoff therefore remains uncorroborated.
2026-08-17T08:22:28Z
evidence attached: reddit.post.1vqm51f — An independent hands-on report provides useful baseline evidence about Qwen 3.8 27B capability, though it does not directly test the abliterated checkpoint.
2026-08-17T05:27:06Z
The refreshed discussion remains qualitative commentary on base Qwen3.8 outputs and does not evaluate the abliterated FP8 checkpoint. With no checkpoint-specific refusal testing or controlled capability comparison, the claimed tradeoff remains uncorroborated.
2026-08-16T21:35:18Z
Refreshed comments merely critique and praise the base-model coding demo; they still do not identify the abliterated FP8 checkpoint or provide controlled refusal and capability measurements. The claimed safety-capability tradeoff remains uncorroborated.
2026-08-16T18:33:33Z
The coding demo shows that base Qwen3.8 can produce plausible local coding outputs, but it is not tied to the abliterated FP8 checkpoint and does not measure refusals or comparative benchmark retention. The central safety-capability tradeoff therefore remains an uncorroborated model-card claim.
2026-08-16T18:23:12Z
evidence attached: reddit.post.1vq37kt — The user's Qwen 3.8 coding examples provide limited qualitative context on practical local capability, though they do not test the safety tradeoff directly.
2026-08-16T17:40:08Z
Refreshed discussion adds source confusion and anecdotes about other HuiHui or Qwen variants, not an independent evaluation of this checkpoint. The refusal-versus-capability claim remains uncorroborated and dependent on the model card’s limited classifier.
2026-08-16T15:35:13Z
Refreshed comments add only private, method-free anecdotes that some abliterated variants perform well; they neither evaluate this checkpoint nor test its refusal metric. The case remains an uncorroborated model-card tradeoff claim awaiting reproducible independent results.
2026-08-16T13:26:40Z
The additional checkpoint-availability discussion confirms ecosystem interest but does not independently test the claimed refusal reduction or capability preservation. The case remains an uncorroborated model-card claim awaiting reproducible safety and benchmark evaluations.
2026-08-16T13:22:47Z
evidence attached: reddit.post.1vpw006 — This appears to corroborate the release of the Qwen 3.8 abliterated checkpoint, though it provides no independent capability or safety measurements.
2026-08-16T10:32:35Z
Refreshed discussion adds only anecdotal support that abliterated models can perform well on private coding tasks, without methods, results, or a reproducible artifact. The central refusal-versus-capability claim remains uncorroborated and dependent on the model card’s limited classifier.
2026-08-16T07:35:27Z
No independent evaluation or implementation evidence has arrived; the small engagement increase only amplifies the same self-reported, classifier-limited claims. The case remains a testable but uncorroborated safety-capability tradeoff.
2026-08-16T07:27:55Z
grounded: known/low — Scott’s Evaluation-Driven Development and independent-verification pages already require repeatable external evidence before accepting changed model behaviour;
2026-08-16T07:25:12Z
origin walked (codex/luna, conf 0.98): anchor reddit.post.1vppox6 -> echo.other.e333fb1dc1 by OrcaRouter
2026-08-16T07:23:55Z
case created — The claimed refusal-versus-capability tradeoff is concrete and testable, but currently rests on one secondary report without a linked checkpoint or evaluation artifact.

Decision trace

08-18 01:35expireRepeated comment refreshes have exhausted the near-term signal without producing any checkpoint-specific refusal test or controlled capability comparison. The case can fade from active tracking and be
08-18 01:35alert_silentThe new delta is only repetitive qualitative discussion of the standard Qwen3.8 model, so it adds nothing consequential before the next briefing.
08-18 01:35alert_routeThe new delta is only repetitive qualitative discussion of the standard Qwen3.8 model, so it adds nothing consequential before the next briefing.
08-18 01:21sensor_dirtycomment_update
08-18 00:43repriceThe comment refresh adds only more qualitative impressions of standard Qwen3.8 and no evidence tied to the abliterated FP8 checkpoint. Repeated amplification is not narrowing the uncertainty around re
08-18 00:43alert_silentNo independent checkpoint-specific refusal test, controlled capability comparison, or reproducible result emerged; the refreshed discussion can wait for routine review.
08-18 00:43alert_routeNo independent checkpoint-specific refusal test, controlled capability comparison, or reproducible result emerged; the refreshed discussion can wait for routine review.
08-18 00:21sensor_dirtycomment_update
08-18 00:08repriceThe latest comment refresh is repetitive qualitative discussion of the standard Qwen3.8 model, not an independent test of the abliterated FP8 checkpoint. The claimed refusal reduction and capability r
08-18 00:08alert_silentNo checkpoint-specific refusal evaluation, controlled capability comparison, or reproducible new result emerged; another discussion refresh does not warrant attention before the next briefing.
08-18 00:08alert_routeNo checkpoint-specific refusal evaluation, controlled capability comparison, or reproducible new result emerged; another discussion refresh does not warrant attention before the next briefing.
08-17 23:21sensor_dirtycomment_update
08-17 22:46repriceThe refreshed comments are further qualitative amplification of standard Qwen3.8 capability, not checkpoint-specific evidence about the abliterated FP8 model. Without controlled refusal testing or com
08-17 22:46alert_silentNo consequential new fact emerged; the discussion still lacks an independent evaluation of the abliterated checkpoint and can wait for routine review.
08-17 22:46alert_routeNo consequential new fact emerged; the discussion still lacks an independent evaluation of the abliterated checkpoint and can wait for routine review.
08-17 22:21sensor_dirtycomment_update
08-17 21:31repriceThe refreshed comments remain qualitative impressions of the standard Qwen3.8 model and add no controlled test of the abliterated FP8 checkpoint. This is repetitive amplification, not evidence for the
08-17 21:31alert_silentNo checkpoint-specific evaluation, reproducible artifact, or consequential new fact emerged; the comment refresh can wait for the next briefing.
08-17 21:31alert_routeNo checkpoint-specific evaluation, reproducible artifact, or consequential new fact emerged; the comment refresh can wait for the next briefing.
08-17 21:21sensor_dirtycomment_update
08-17 20:35repriceThe refreshed discussion remains qualitative commentary about the standard Qwen3.8 model and adds no checkpoint-specific refusal test, controlled capability comparison, or reproducible artifact. This
08-17 20:35alert_silentNo consequential new fact has emerged: the comments still do not evaluate the abliterated FP8 checkpoint, so the delta can wait for normal review.
08-17 20:35alert_routeNo consequential new fact has emerged: the comments still do not evaluate the abliterated FP8 checkpoint, so the delta can wait for normal review.
08-17 20:21sensor_dirtycomment_update
08-17 19:30repriceThe refreshed comments remain conflicting, method-free impressions of the standard Qwen3.8 model rather than tests of the abliterated FP8 checkpoint. This is repetitive amplification, leaving the refu
08-17 19:30alert_silentNo checkpoint-specific refusal measurements, controlled capability comparisons, or reproducible evaluation artifacts were added, so the discussion does not change the case enough to precede the next b
08-17 19:30alert_routeNo checkpoint-specific refusal measurements, controlled capability comparisons, or reproducible evaluation artifacts were added, so the discussion does not change the case enough to precede the next b
08-17 19:21sensor_dirtycomment_update
08-17 18:30repriceThe new hands-on report strengthens the view that base Qwen3.8-27B is practically capable, but it evaluates a different quant and provides no checkpoint-specific refusal testing or controlled capabili
08-17 18:30alert_silentThis is qualitative baseline evidence about the standard model rather than an independent evaluation of the abliterated FP8 checkpoint, so it does not materially change the case and can wait for the n
08-17 18:30alert_routeThis is qualitative baseline evidence about the standard model rather than an independent evaluation of the abliterated FP8 checkpoint, so it does not materially change the case and can wait for the n
08-17 18:22alert_silentThe new post is an anecdotal review of a standard Qwen3.8-27B quant, not an independent evaluation of the abliterated FP8 checkpoint. It provides no new refusal measurements or controlled capability c
08-17 18:22alert_routeThe new post is an anecdotal review of a standard Qwen3.8-27B quant, not an independent evaluation of the abliterated FP8 checkpoint. It provides no new refusal measurements or controlled capability c
08-17 18:22attachAn independent hands-on report provides useful baseline evidence about Qwen 3.8 27B capability, though it does not directly test the abliterated checkpoint.
08-17 18:21propose_attachAn independent hands-on report provides useful baseline evidence about Qwen 3.8 27B capability, though it does not directly test the abliterated checkpoint.
08-17 15:27repriceThe refreshed discussion remains qualitative commentary on base Qwen3.8 outputs and does not evaluate the abliterated FP8 checkpoint. With no checkpoint-specific refusal testing or controlled capabili
08-17 15:27alert_silentThe new comments add no independent measurements, reproducible methodology, or checkpoint-specific findings, so nothing consequential would be lost by waiting for the next briefing.
08-17 15:27alert_routeThe new comments add no independent measurements, reproducible methodology, or checkpoint-specific findings, so nothing consequential would be lost by waiting for the next briefing.
08-17 15:21sensor_dirtycomment_update
08-17 11:21sensor_dirtyengagement_update