GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
zhangjiaxin2012
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
zhangjiaxin2012
Model Tree
google/gemma-4-26B-A4B-it
Input Modalities
Output Modalities
Supported Functionality
From own canonical causal Stage2: reasoning-token SFT, 1,024 updates, rank32/alpha64 attention-LoRA, FP32 AdamW LR1e-5; native merged export. No subsequent full-language RL updates.
Original Google base: google/gemma-4-26B-A4B-it@4d7ae4984b7db7de8f8457170b3f1a419ee76d52. Native architecture: Gemma4ForConditionalGeneration.
All native tensor bytes match the completed training artifact. Completed public evaluations are bound to that exact artifact identity; pending evaluations are marked below. No dtype conversion or quantization was applied during publication. Stored tensor dtypes are recorded in release.json; measured inference loads BF16.
The same checkpoint supports both policies below. Policy, prompt, option mapping and reasoning budget are part of the reported serving system. Direct evaluation generates zero tokens. Gated thinking uses native thinking only when max(p0)<0.9, up to4,096 tokens, then0.25p0+0.75p1 after JSON analysis reinjection. Post-calibration is identity.
| Policy | Correct /231 | Accuracy | Intelligence | Calibration | Capability | NLL ↓ | Brier ↓ |
|---|---|---|---|---|---|---|---|
| Direct | 193 | 83.55% | 60.58 | 77.20 | 68.89 | 0.440887 | 0.234730 |
| Gated 4K thinking | 219 | 94.81% | 83.36 | 85.20 | 84.28 | 0.233831 | 0.102571 |
Capability=(Intelligence+Calibration)/2; it is different from answer accuracy. Public231 is an adaptive development panel, with different realized reasoning costs across checkpoints. These measurements are not official leaderboard scores.
Full comparison and method: https://github.com/jxzhangjhu/srf-jev.
Use Python3.10+ and install a CUDA build of PyTorch2.6.0 suitable for your machine, then:
bash
python -m pip install huggingface_hub==1.5.0hf download zhangjiaxin2012/SRF-Jev-Gemma4-26B-A4B-Stage3-Reasoning-SFT1024-Think4K-v2 --local-dir checkpointpython -m pip install -r checkpoint/requirements.txtpython checkpoint/predict.py --checkpoint checkpoint --input checkpoint/examples.jsonl --output predictions.jsonl --policy autothink4096 --device cuda:0For exact experiment tracking, add --revision <immutable commit from the GitHub checkpoint table> to the download command. Run with --policy direct for the direct row; add --verify-weights to check every downloaded shard before loading.
Input is JSONL with exactly id and visible per row. The examples show Choice, NoUL and Score schemas. Keep IDs stable across policies; they determine the label-blind option permutation and thinking seed. Keep gold labels in a separate file. Outputs include semantic option IDs, probabilities, the selected option and actual token/cap counts. Existing output files are never overwritten.
The exact native class in config.json can also be loaded through Transformers5.17.0 with dtype=torch.bfloat16 and attn_implementation="sdpa". Generic chat prompts do not reproduce the decision table. The supplied decision runner uses one CUDA device and math-only SDPA, matching the H200 reference setup; multi-GPU/quantized backends are separate evaluation configurations.
model*.safetensors, native config/index/tokenizer files, release.json, evaluation.json, portable-qualification.json, release-files.json, the standalone inference code, examples and Apache-2.0 license/attribution. Original optimizer states, training/evaluation records, labels, internal logs and private runtime probes are not published.
The original native artifact has export/reload qualification. The portable encoder/prompt/probability code is checked on CPU against its saved231-question captures when that evaluation is complete. That replay does not constitute a fresh GPU portability test of this new CLI. Final publication.json records complete Hub file/hash verification and anonymous access.
Reasoning-SFT supervises native thinking continuations; full RL, when present, optimizes generated thinking continuations. The weights support both Direct and gated Thinking-4K evaluation. The name identifies the reported inference recipe; enable_thinking=True alone does not reproduce the gate, decision readout, JSON reinjection and probability mixture.
python -m pip install huggingface_hub==1.5.0hf download zhangjiaxin2012/SRF-Jev-Gemma4-26B-A4B-Stage3-Reasoning-SFT1024-Think4K-v2 --local-dir checkpointpython -m pip install -r checkpoint/requirements.txtpython checkpoint/predict.py --checkpoint checkpoint --input checkpoint/examples.jsonl --output predictions.jsonl --policy autothink4096 --device cuda:0