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
| Path | Bytes | Role |
|---|---|---|
adapter_model.safetensors (LFS) / adapter_config.json | 25,587,104 / 1,511 | LoRA adapter (base_model_name_or_path: Qwen/Qwen3.5-0.8B, r=16, alpha=32, task_type: CAUSAL_LM) |
model.safetensors (LFS) + config.json + generation_config.json | 3,009,613,624 / 1,864 / 122 | derived merged artifact (see merge receipt) |
tokenizer.json (LFS) / tokenizer_config.json / chat_template.jinja / processor_config.json | 19,989,343 / 1,230 / 7,908 / 1,251 | tokenizer, chat template, processor config |
training_receipt.laptop-blackwell.json | 1,733 | training receipt (unsigned) |
eval_dev_named_n.json | 5,441 | development named-N receipt (12 cases, per-case outcomes) |
frontier_gpu_summary.json | 463 | GPU summary (ra_v3: 11/12 MEASURED) |
merge_receipt.json | 583 | 2026-09-14 manual merge repair receipt |
provenance.json / status.json / LICENSE | 1,766 / 1,405 / 11,357 | license-candidate records and standalone Apache-2.0 text |
bom/model-bom.cdx.json | 5,039 | CycloneDX BOM |
.gitattributes | 1,570 | platform-managed |
Sizes were read from the Hub listing; no weight file was downloaded or hashed by this pass, and the revision SHA was not captured. The Hub-hosted banner referenced by the szl-forge card (holo-banner.svg) is not on this tree, so it is not embedded here.
The training receipt records
180 training rows and train loss 0.0537. Training loss is not evaluation.
The receipt describes the local laptop Blackwell bf16 run, not the Rosie
supervised 4-bit lane; load_in_4bit=false. Recipe as recorded: bf16 LoRA r=16 α=32, seed 11, 3 epochs, warmup 10, learning rate 1e-4, held_out_in_gradients=false, push_to_hub=false; hardware NVIDIA GeForce RTX 5050 Laptop GPU (7.96 GB), torch 2.10.0+cu128, Unsloth 2026.7.2, Windows.
The published development evaluation records:
| Scope | Recorded result |
|---|---|
| Split and count | DEV, n=12 |
| Total | 11/12 |
| Draft | 4/4 |
| Recovery | 3/4, one unparseable output |
| Refuse | 4/4 |
| Test split opened | opened_test_jsonl=false |
The historical 11/12 failure remains part of the evidence. A separate additive
2026-09-29 owner-run DEV check reports 12/12 on adapter SHA-256
90b82ade0b2c84eca92e90cdeb56137d61d7da4b4863ff4d29fa75e9dc037870. It explicitly does not supersede the older record,
records base_model=UNRECORDED and promotion_effect=NONE, and is not a
held-out test, independent benchmark, release gate, or evaluation of the
derived merged checkpoint. Neither check establishes complete reproducible
inference lineage or deployment readiness.
The merge receipt records the 2026-09-14 repair of
an empty base-copy merge (an audit found all 320 tensors byte-equal to the unsloth/Qwen3.5-0.8B text tower), applying the root adapter to 96/96 modules with maximum
weight delta 0.0037259478121995926 (W' = W + (alpha/r)·B@A, scale 2.0). This repair is artifact construction evidence,
not a fresh behavioral evaluation. No hash of the repaired model.safetensors is recorded in that receipt: the merged bytes are PRESENT_UNRECEIPTED by digest.
| Field | Classification |
|---|---|
| Artifact | Trained LoRA adapter weights for Qwen/Qwen3.5-0.8B, not a from-scratch model; plus a derived merged full model rebuilt 2026-09-14. |
| Value | Proposal-only receipt drafts. Validation, approval, execution, and minting stay outside the weights. |
| Evidence | Train loss MEASURED on the 180-row committed split. Separate owner-run DEV n=12 checks: historical 11/12 and additive adapter-bound 12/12; neither held-out nor a merged-checkpoint qualification. |
| Limits | Small synthetic curriculum. No broad capability, factuality, safety, or third-party benchmark claim. |
| Runtime | Drafts only. Not an autonomous agent. |
Untested in this card pass. Adapter path (the declared artifact):
python
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer base_id = "Qwen/Qwen3.5-0.8B"tok = AutoTokenizer.from_pretrained(base_id)base = AutoModelForCausalLM.from_pretrained(base_id)model = PeftModel.from_pretrained(base, "SZLHOLDINGS/szl-receiptagent-qwen35-0.8b-v3")The root also contains config.json and the merged model.safetensors; callers must verify which path their library selects rather than relying on loader precedence.
Proposal-only drafts. Validation, approval, execution, and receipt minting stay outside the weights. No autonomy, deployment, promotion, or flagship selection is authorized. The published status record retains blocked lineage, consent, privacy, training-suitability, and deployment fields; its publication status concerns its license-candidate record, not model promotion. Lambda uniqueness remains Conjecture 1, advisory and open, never a theorem.
Use the linked immutable Hub receipts at fe7564886100316d0843d147abd38da39beb3a4f and the canonical
v3 source directory (pinned revision; current main; card source: receiptagent-v3/card/README.md; GitHub is canonical, this Hub card is the mirror, re-synced 2026-09-25 with the historical 11/12 receipt retained; current Forge source distinguishes it from the separate additive adapter-bound 12/12 DEV record).
The v2 adapter, this v3 research SKU, the derived merged bytes, and the 1.5B
ReceiptAgent remain separate artifacts with separate evidence scopes.
Apache-2.0. Copyright 2026 SZL Holdings · Stephen P. Lutar Jr.
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer base_id = "Qwen/Qwen3.5-0.8B"tok = AutoTokenizer.from_pretrained(base_id)base = AutoModelForCausalLM.from_pretrained(base_id)model = PeftModel.from_pretrained(base, "SZLHOLDINGS/szl-receiptagent-qwen35-0.8b-v3")