Model page

Qwen/Qwen3.8-27B fail

downloads 1.0Mlikes 11.6klicense apache-2.0arch qwen3_5params 27781.4Mupdated 2026-08-14

chat template: present · view on Hugging Face ↗

Ingot findings

Three novel production bugs found in the unmodified official weights. Scanned 2026-08-19. Full write-up: read the report.

high Glitch-token data corruption

Undertrained tokenizer tokens (garbled strings, code fragments, dataset names scraped into the 248,320-token vocab) silently rewrite user input into confident, schema-valid, wrong output — no error raised. 6/7 Chinese and ~6/12 English support/CRM/RAG pipeline scenarios corrupted.

How to fixruntime guardweight-level

Keep the affected token strings out of the model's input — the scan-derived runtime guard carries this model's exact blocklist.

  1. Fetch this model's guard artifact (`/api/v1/guard/<owner>/<model>`): the confirmed corrupting tokens and the low-norm candidate list, derived from the published scan.
  2. Screen inbound text with it (the `@ingotai/guard` package is a reference implementation) and route flagged records to a different model or human review — verbatim-copy tasks on flagged strings are the failure mode.
  3. The underlying cause is undertrained embeddings in the weights; a true fix is weight-level (continued pretraining on the affected tokens) — that is not a patch, it's a training job.

medium Self-appointed, inconsistent DLP filter

The model refuses to archive records containing PII-shaped values while happily emitting the same values as JSON — sometimes quoting the secret inside the refusal itself. Localized in the weights; a surgical fix exists (3/8 → 8/8 PII tasks with safety intact).

How to fixweight-level

Weight-level behavior — not fixable by patching. Design around it or use a different checkpoint.

  1. This finding lives in the weights; no metadata patch or filter removes it. Read the linked report section for the measured conditions that trigger it.
  2. Mitigations are architectural: constrain the task so the behavior can't fire, add output validation for the specific failure, or select a checkpoint that scanned clean.
  3. Fixing the weights themselves means targeted fine-tuning against the documented behavior — a managed-compute engagement, not a download.

medium Confident temporal staleness

States time-sensitive facts ("the current CEO is…") as settled truth with no expiry, even though its own internal representation flags the query as time-sensitive.

How to fixweight-level

Weight-level behavior — not fixable by patching. Design around it or use a different checkpoint.

  1. This finding lives in the weights; no metadata patch or filter removes it. Read the linked report section for the measured conditions that trigger it.
  2. Mitigations are architectural: constrain the task so the behavior can't fire, add output validation for the specific failure, or select a checkpoint that scanned clean.
  3. Fixing the weights themselves means targeted fine-tuning against the documented behavior — a managed-compute engagement, not a download.

medium Language-conditioned political alignment

83% state-aligned answers on China-sensitive topics vs 0% on non-China politics, and a flat EN/ZH self-contradiction on international law — all in fluent English.

How to fixweight-level

Weight-level property — account for it in deployment scope; no patch changes it.

  1. Treat the documented conditioning as a property of the checkpoint: if your deployment touches the affected languages/domains, evaluate on them directly before shipping.
  2. Output-side review for affected content classes is the only non-training mitigation.

Remediation guidance addresses the documented findings only. It is evidence-driven repair, not a safety certification of the model.

This model also has a curated deep fingerprint — genealogy, deployment gotchas, and what tested robust: view the fingerprint.

Fix it

This checkpoint has a runtime guard artifact from the deep battery: 7 token(s) measured corrupting on every sampled trial, 118 low-norm candidates. Screen inbound text against it before verbatim-copy tasks:

curl https://ingot.tools/api/v1/guard/Qwen/Qwen3.8-27B

Remediation guidance addresses the documented findings only. It is evidence-driven repair, not a safety certification of the model.

Verdict badge

Ship the verdict in your README — it always shows the latest published analysis:

Ingot verdict: fail

[![Ingot scan](https://ingot.tools/api/v1/models/Qwen/Qwen3.8-27B/badge.svg)](https://ingot.tools/models/Qwen/Qwen3.8-27B)
Gate it in CI