GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
G33-k
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
G33-k
Model Tree
Qwen/Qwen3-VL-4B-Instruct
Input Modalities
Output Modalities
Supported Functionality
A narrow text tool-use controller. The corpus is text-only; the vision tower does no
work and was not loaded or trained (FastModel with text_only=True). It learns the
same six behaviors: input-form mapping, tool order, argument discipline, clarification,
honest status reporting, and expiry/foreign-ID recovery.
Synthetic only. Trained on sft/train.jsonl of corpus revision 2026-09-21
(1,032 conversations, 2,856 assistant decisions, 43 base families), generated by
executing the real tools against deterministic fixture worlds. Zero live-company or
human-reviewed training examples. Fixture success is not real-world accuracy.
| Trainer | train_unsloth.py --family qwen3vl (unsloth 2026.9.7, FastModel text_only=True) |
| LoRA | r=16, alpha=16 (alpha == r, unsloth recommendation), dropout 0 |
| Learning rate | 2e-4, cosine, warmup 5% (unsloth QLoRA starting recommendation) |
| Epochs | 2 (4,170 optimizer steps) |
| Quantization | NF4 (4-bit) base, bf16 compute |
| Tokenization | per-decision, assistant-only masking, native chat template, max 16,384 (longest decision 11,124 tokens) |
| Host | NVIDIA RTX PRO 6000 Blackwell (sm_120), MIG 4g.96gb slice |
| Wall clock | 3h 04m, 1,420 tok/s overall, 5.8 GiB peak |
| Validation loss | 0.0 over 714 decisions (saturation signal) |
Config note: this run uses the unsloth-recommended alpha/LR; the repo's earlier SAPPAN recipe (alpha=32, LR 1e-4) was started, killed, and is not the shipped adapter.
Full 258-episode validation on corpus revision 2026-09-21
(training-runs/reports/nim-validation-2026-09-21.md): 258/258 episodes (scenario macro
1.000, every one of the 43 families 6/6) against the untouched base at 0/258. Serving was the
unmerged adapter over the same NF4 base, greedy (temperature 0), thinking off, on the nim host
(RTX PRO 6000 Blackwell MIG 4g.96gb). The base model failed every episode, most often on
verification arguments (VERIFY_ARGUMENT_expected_company 198,
VERIFY_ARGUMENT_jurisdiction 192) and the final-answer schema (185). Paired difference +1.000,
standard error clustered by family. Synthetic fixture worlds only: fixture success is not
real-world accuracy.
The adapter is a LoRA delta, meant to be attached to the 4-bit base, not merged:
python
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfigfrom peft import PeftModel base = "Qwen/Qwen3-VL-4B-Instruct"quant = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=torch.bfloat16)tok = AutoTokenizer.from_pretrained(base)model = AutoModelForCausalLM.from_pretrained(base, quantization_config=quant, device_map={"": 0})model = PeftModel.from_pretrained(model, "G33-k/qwen3-vl-4b-brand-tools-controller-lora")Call the two tools with the public schemas, one tool call per assistant turn, greedy and with
thinking off. The source repository wraps this into an OpenAI-compatible endpoint
(training-runs/serve_openai.py) and scores it with the fixture harness:
sh
(cd minicpm5-eval-kit && node --experimental-strip-types scripts/evaluate.ts \ --mode model --split validation --base-url http://127.0.0.1:30000/v1 \ --model qwen3vl-adapter --out ../training-runs/reports/qwen3vl-adapter-validation.json)Full instructions, including the decode and argument-coercion rules for this model family, are in
minicpm5-eval-kit/docs/EVALUATION_GUIDE.md section 7 of
github.com/geekhoo/minicpm5-2b-voice.
The adapter is keyed for the text-only decoder (model.layers.N). Load it onto a
text-only load of the base, not AutoModelForImageTextToText without key handling
(unsloth#3560). Qwen-style <tool_call>{json}</tool_call> output must be parsed with
schema-typed coercion. Serve greedy, thinking off, enable_thinking=False at render
time. Evaluate the unmerged adapter on the NF4 base it trained against.
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfigfrom peft import PeftModel base = "Qwen/Qwen3-VL-4B-Instruct"quant = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=torch.bfloat16)tok = AutoTokenizer.from_pretrained(base)model = AutoModelForCausalLM.from_pretrained(base, quantization_config=quant, device_map={"": 0})model = PeftModel.from_pretrained(model, "G33-k/qwen3-vl-4b-brand-tools-controller-lora")(cd minicpm5-eval-kit && node --experimental-strip-types scripts/evaluate.ts \ --mode model --split validation --base-url http://127.0.0.1:30000/v1 \ --model qwen3vl-adapter --out ../training-runs/reports/qwen3vl-adapter-validation.json)