GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
SZLHOLDINGS
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
SZLHOLDINGS
Model Tree
Qwen/Qwen3.5-0.8B
Input Modalities
Output Modalities
Supported Functionality
Emits strict JSON triage fields (e.g. label, state, evidence). Prefers state="REVIEW" when evidence is thin, labels are tied, or the prompt tries to instruct the classifier instead of describing a problem.
*-study5 / *-laptop-lora / *-nemo-candidate / future *-candidate_enterprise_* as this canonical idcandidate_enterprise_* id)main as a serve contract| Field | Status |
|---|---|
Hub datasets: ids | UNDECLARED on prior card — this wave does not invent Hub dataset ids |
| Local / receipt digests | May exist off-card; not asserted as Hub datasets here |
| BOM | UNAVAILABLE as attested CycloneDX energy/BOM claim on this wave (packaging files if present are not MEASURED joules) |
| Base model pin | Qwen/Qwen3.5-0.8B · relation: adapter · Unsloth-adjacent tooling noted in tags |
| Suite | Result | Class |
|---|---|---|
Retained 66-row behavioral gate (immutable rev 24b5c444…) | 66/66 exact labels · 66/66 exact states · 0 malformed · 23 gold-refusal · 0 false labels on refusals · 0 ungrounded spans (gate_report.json) | MEASURED (bounded gate only) |
Literal PROMOTABLE in that gate report | Belongs to bounded behavioral gate only | MEASURED ≠ release clearance |
| Contamination / leakage review | Did not clear → release BLOCKED | MEASURED policy |
| Third-party public leaderboard | — | UNAVAILABLE |
Current main tensors ≡ gate revision | — | UNKNOWN without pin+diff |
Interpretation rules: (1) metrics are documentation artifacts; (2) they do not override contamination; (3) Hub visibility ≠ approval; (4) promotion stays blocked until a cleared superseding run.
| Field | Status |
|---|---|
| Train energy (kWh / Joules) | UNAVAILABLE (no attested meter on card) |
| Infer energy | UNAVAILABLE |
| Model BOM | UNAVAILABLE as MEASURED energy claim; treat any bom/* packaging as SOURCE_BACKED only |
| DSSE / signed receipts | Prefer real-DSSE-or-honestly-UNSIGNED; never fabricate PASS |
python
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizerbase_id = "Qwen/Qwen3.5-0.8B"base_rev = "<BASE_PIN>"adapter = "SZLHOLDINGS/szl-triage-qwen3.5-0.8b-lora"adapter_rev = "<PINNED_SHA>" # prefer gate rev 24b5c444… for replay; never float main for servetok = AutoTokenizer.from_pretrained(base_id, revision=base_rev)base = AutoModelForCausalLM.from_pretrained(base_id, revision=base_rev, device_map="auto")model = PeftModel.from_pretrained(base, adapter, revision=adapter_rev)deny-by-default: treat outputs as proposals for an external controller. Do not deploy from this card.
SZLHOLDINGS/szl-triage-qwen3.5-0.8b-loraszl-triage-qwen3.5-0.8b-lora-study5 (separate SKU)szl-holdings/szl-typesafe-triage*triage*candidate_enterprise_20261104* at wave2 stamp — upgrade that card when it appears under a distinct idStephen Lutar / SZL Holdings
Thank you NVIDIA NeMo/NIM, Unsloth, Hugging Face Hub, vLLM / PagedAttention, and FlashAttention authors. SZL builds clean-room silhouettes with DSSE receipts and MEASURED-or-null energy (never fabricates joules, signatures, or green evals).
Apache-2.0. See LICENSE in-repo when present.
NVIDIA-style card upgrade · wave2 2026-11-04T18:00Z · clean-room silhouette (DSSE receipts + MEASURED-or-null energy).
Prepared as Massed enterprise candidate card shape; does not authorize train, serve, or autonomy. Does not overwrite sibling SKUs.
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizerbase_id = "Qwen/Qwen3.5-0.8B"base_rev = "<BASE_PIN>"adapter = "SZLHOLDINGS/szl-triage-qwen3.5-0.8b-lora"adapter_rev = "<PINNED_SHA>" # prefer gate rev 24b5c444… for replay; never float main for servetok = AutoTokenizer.from_pretrained(base_id, revision=base_rev)base = AutoModelForCausalLM.from_pretrained(base_id, revision=base_rev, device_map="auto")model = PeftModel.from_pretrained(base, adapter, revision=adapter_rev)