GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
simaai
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
simaai
Model Tree
google/gemma-4-E4B-it
Input Modalities
Output Modalities
Supported Functionality
| Component | Method | Weight format | Effective targets |
|---|---|---|---|
Decoder Linear layers and lm_head | GPTQ | symmetric INT4, group size 256, static act-order | 344 |
| Vision encoder Linear layers | RTN | symmetric INT8, per-channel | 113 |
| Audio tower | none | source BF16 | 134 explicit exceptions |
| Vision/audio embedding projectors | none | source BF16 | 2 explicit exceptions |
The lm_head follows the default GPTQ INT4/G256 policy; no INT8 escalation was
made. Text GPTQ calibration used lmms-lab/flickr30k test[:512],
deterministic dataset order, 512 image-text samples, sequence length 2048,
batch size 1, no concatenation, and no padding to maximum length. Vision RTN
does not use Hessian calibration. Exact targets and exceptions are saved in
recipe.yaml.
Full MMStar used all 1,500 examples, VLMEvalKit commit
7055d3010c38ccb5dcae1bc9535ca19c7fe5d79f, deterministic generation, a
matched letter-only system prompt, maximum 16 new tokens, SDPA, and
exact_matching without an API judge.
| Checkpoint | Overall accuracy |
|---|---|
| Source | 47.1333% |
| Existing text GPTQ + vision GPTQ checkpoint | 40.4667% |
| This text GPTQ + vision RTN checkpoint | 43.4667% |
| RTN change versus vision GPTQ | +3.0000 percentage points |
| RTN relative change versus vision GPTQ | +7.4135% |
Evaluation date: 2026-07-17. Slurm jobs: source 4708, vision-GPTQ baseline
4709, and this vision-RTN candidate 4742. The comparison is recorded in
vlmeval_results/mmstar_results_vision_rtn.csv; raw predictions and status
files remain under vlmeval_results/mmstar_full_vision_rtn/.
Validation: MMStar inference passed (1,500/1,500).
Full matched WikiText-2 word perplexity (EleutherAI/wikitext_document_level, wikitext-2-raw-v1, no example limit; 2026-07-18): source 256.5985; this checkpoint 388.6980; absolute degradation +132.0995; relative degradation +51.48%. Independent native-logit scoring confirms the direction of degradation, but untemplated raw WikiText likelihood is not a suitable Gemma 4 instruction-model release gate. Investigation and raw evidence: releases/vlm-full-wikitext-perplexity/GEMMA4_PPL_INVESTIGATION.md and perplexity_results/vlm_full_wikitext/.
This directory contains the exact quantize.py, recipe.yaml, and
versions.txt used for the artifact:
bash
python quantize.py \--model-path /path/to/source-model \--output-dir /path/to/new-quantized-model
text
Python: 3.13.2torch: 2.11.0+cu128CUDA: 12.8transformers: 5.10.1llm-compressor: 0.12.0auto-round: 0.13.0compressed-tensors: 0.17.1
This is the pre-LLiMa quantized Hugging Face artifact. LLiMa compiler output must remain separate unless its format and redistribution are approved. No Hugging Face upload is authorized for this batch.
Quantization quality can vary by language, visual domain, prompt format, context length, and downstream runtime. MMStar does not replace application-specific validation. Audio weights and modality projectors remain in source precision. Users must validate compiler/runtime compatibility on the target deployment.