Model page

Qwen/Qwen3.8-2.4T-A95B warn

Glitch tokens that can silently corrupt ordinary input.

downloads 20.6klikes 1.2klicense otherarch qwen3_5_moe_textparams 2446182.7Mupdated 2026-08-12

chat template: present · view on Hugging Face ↗

Scan coverageStatic battery2026-08-25Weights battery2026-08-25Behavioral batterynot rundetails
BatteryLooks atStatus
Static batteryMetadata & packagingcomplete 2026-08-25
Weights batteryWeights forensics — no GPU, no downloadcomplete 2026-08-25248,320-token embedding scanned · 2367 undertrained
Behavioral batteryLive-inference differentialsnot run

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

Findings

Scanned 2026-08-25 · published from a community scan.

medium Undertrained (glitch) token surface in vocabulary

Embedding-norm scan flagged 2367 undertrained tokens (norm < 0.3× the vocabulary median of 0.925), including 660 plain-ASCII strings that can appear in ordinary input as identifiers — e.g. "JernihBer", "priveontvangst", "szexf", "tedothi", "$PostalCodesNL", "ForCanBeConvertedToF", "PostalCodesNL", "ForCanBeConverted". 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.

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

Fingerprint

The durable profile of this model: measured weights-and-metadata facts, rebuilt on every scan and battery run. Updated 2026-08-25.

architectureqwen3_5_moe_text · 92 layers · 8192-dim
parameters2446182.7M
vocabulary248,320 tokens
licenseother
serializationsafetensors
chat templatepresent (chat_template.jinja) · sha256:40ce34a5bcbc0231
glitch-token surface2,367 undertrained candidates, 660 plain-ASCII
Full measured fingerprint
architecturesQwen3_5MoeForCausalLM
librarytransformers
pipelinetext-generation
repo files224
revision207bd685a7e3
HF snapshot20.6k downloads · 1.2k likes · updated 2026-08-12 · captured 2026-08-25
embedding tensormodel.embed_tokens.weight · BF16 · 248,320×8192
embedding normsmedian 0.9251 · mean 0.9281
lineage checkno claimed base model
glitch-token samples"JernihBer", "priveontvangst", "szexf", "tedothi", "$PostalCodesNL", "ForCanBeConvertedToF", "PostalCodesNL", "ForCanBeConverted", "thuisontvangst", "useRalative", "sihteeriopisto", "Kinhted"
Battery runs (1)the run trace behind the findings — what each job measured
batterystatusqueueddurationattempts
weightscomplete2026-08-25 19:002m2
weights run 2026-08-25 measurements
probes runglitch-norm-scan, zero-template-token-scan, lineage-norm-correlation
embedding tensormodel.embed_tokens.weight · BF16 · 248,320×8192
glitch surface2,367 undertrained, 660 plain-ASCII
lineage checknot checked (no claimed base model)

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/Qwen3.8-2.4T-A95B/badge.svg)](https://ingot.tools/models/Qwen/Qwen3.8-2.4T-A95B)
Gate it in CI