GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
arianraje
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
arianraje
Model Tree
Qwen/Qwen3-4B
Input Modalities
Output Modalities
Supported Functionality
| Model | AIME24 pass@1 | AIME25 pass@1 |
|---|---|---|
| This checkpoint | 19.2% | 17.9% |
| OPD 400M | 33.3% | 31.7% |
AIME evaluation: 30 problems per year, 8 samples per problem, thinking enabled, 32,768 generation-token cap. Scores are empirical/unbiased pass@1. These small evaluations support descriptive rankings, not claims of significance.
OPD reference: pinkskin/qwen3-4b-gdn-wsd-ladder/wsd-flat-ext200-400M,
revision ae6b72a357d167fb459dfeda3bded4941decdf41, actual tokens 400,012,332.
The comparison matches nominal loss-token budget and WSD endpoint; actual counts
differ slightly because complete training steps overshoot the token mark.
It does not match wall-clock time or optimizer step. OPD uses reverse KL on
student samples with a growing horizon, so KL direction, trace source/horizon,
and effective RUG token exposure remain confounded.
The checkpoint uses Qwen3NextForCausalLM with dense MLPs and hybrid GDN/full
attention. The training/evaluation environment used Transformers 4.57.3 and
Flash Linear Attention 0.5.1. Load using Transformers AutoModelForCausalLM and
AutoTokenizer with this repository ID. Project vLLM evaluation requires its
Qwen3-Next registry shim for dense layers and tied embeddings.
Inference weights, tokenizer, snapshot metadata, and six evaluation reports are
included. Training optimizer state and sampled generation dumps are excluded.
provenance.json records the exact source, token counts, and OPD pairing.