Qwen/Qwen3.8-27B fail
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.
- 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.
- 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.
- 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.
- 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.
- 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.
- 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.
- 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.
- 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.
- 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.
- Treat the documented conditioning as a property of the checkpoint: if your deployment touches the affected languages/domains, evaluate on them directly before shipping.
- 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.
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:
[](https://ingot.tools/models/Qwen/Qwen3.8-27B)