Model page

Qwen/Qwen3-8B warn

Glitch tokens that can silently corrupt ordinary input. Plus 1 minor note.

downloads 10.7Mlikes 2.1klicense apache-2.0arch qwen3params 8190.7Mupdated 2025-07-26

claims base: Qwen/Qwen3-8B-Base · chat template: present · view on Hugging Face ↗

Scan coverageStatic battery2026-08-27Weights battery2026-08-20Behavioral batterycompletedetails
BatteryLooks atStatus
Static batteryMetadata & packagingcomplete 2026-08-27
Weights batteryWeights forensics: no GPU, no downloadcomplete 2026-08-20151,936-token embedding scanned · 2999 undertrained · lineage consistent
Behavioral batteryLive-inference differentialscompletefull differential battery (curated)

Ingot runs three batteries against a model. What each one checks →

Findings

Scanned 2026-08-27 · published from a community scan. Full write-up: read the report.

medium Undertrained (glitch) token surface in vocabulary

Embedding-norm scan flagged 2999 undertrained tokens (norm < 0.3× the vocabulary median of 1.451), including 102 plain-ASCII strings that can appear in ordinary input as identifiers — e.g. "$PostalCodesNL", "ForCanBeConvertedToF", "PostalCodesNL", "ForCanBeConverted", "useRalative", "thuisontvangst", "useRal", "sexkontakte". In models where this class was tested behaviorally, such tokens silently rewrote user input into confident, schema-valid, wrong output. These are candidates from the weights alone; behavioral confirmation requires the behavioral battery.

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.

info Weights consistent with claimed parent Qwen/Qwen3-8B-Base

Mean cosine similarity of 64 sampled token-embedding rows against Qwen/Qwen3-8B-Base is 0.998 — the weights plausibly descend from the declared base (relation: unspecified).

low Partial glitch-token echo degradation

Echo failures on 5/16 undertrained tokens vs 0/8 controls — a differential exists but below the confirmation bar (≥50% glitch failures with clean controls).

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

Put this result in your workflow

Check every checkpoint before it ships

Use the web app, API, CLI, or CI gate to scan candidate checkpoints and catch model drift before deployment. Public-model scans publish to the open database; paid plans add the volume needed for continuous checks.

Fingerprint

The durable profile of this model: measured weights-and-metadata facts, rebuilt on every scan and battery run. Updated 2026-08-20.

architectureqwen3 · 36 layers · 4096-dim
parameters8190.7M
vocabulary151,936 tokens
licenseapache-2.0
serializationsafetensors
chat templatepresent · sha256:a55ee1b1660128b7
claimed lineageQwen/Qwen3-8B-Base
lineage verifiedconsistent vs Qwen/Qwen3-8B-Base — embedding-row cosine 0.998
glitch-token surface2,999 undertrained candidates, 102 plain-ASCII
Full measured fingerprint
architecturesQwen3ForCausalLM
librarytransformers
pipelinetext-generation
repo files15
revisionb968826d9c46
HF snapshot15.7M downloads · 1.3k likes · updated 2025-07-26 · captured 2026-08-20
embedding tensormodel.embed_tokens.weight · BF16 · 151,936×4096
embedding normsmedian 1.4512 · mean 1.3758
lineage checkconsistent — cosine 0.9979 over 64 sampled rows vs Qwen/Qwen3-8B-Base
glitch-token samples"$PostalCodesNL", "ForCanBeConvertedToF", "PostalCodesNL", "ForCanBeConverted", "useRalative", "thuisontvangst", "useRal", "sexkontakte", "NdrFc", "webElementX", "sextreffen", "wannonce"
Battery runs (5)the run trace behind the findings — what each job measured
batterystatusqueueddurationattempts
gpucomplete2026-08-27 04:3159s1
gpucomplete2026-08-25 21:4787s1
gpucomplete2026-08-21 08:0261s1
weightscomplete2026-08-20 18:0258s1
weightscomplete2026-08-20 08:1049s1
gpu run 2026-08-27 — measurements
probes runglitch
gpu run 2026-08-25 — measurements
probes runglitch
gpu run 2026-08-21 — measurements
probes runglitch

Verdict badge

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

Ingot verdict: warn

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