GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
stanfordasl
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
stanfordasl
Model Tree
Input Modalities
Output Modalities
Supported Functionality
One from-scratch OLMo2-architecture pretraining run from the DIET weight-space project, built for trajectory analysis (same spirit as the jjalora/diet-traj-pythia-410m-* collection, now extended to the OLMo2 architecture on Dolma data).
configs/model/olmo2_medium_dense.yaml in jjalora/diet-lm) --
hidden_size=768, intermediate_size=3072, num_hidden_layers=16, num_attention_heads=12,
num_key_value_heads=12, max_position_embeddings=2048 (~305M params).emozilla/dolma-v1_7-3B (community HF-native mirror of Dolma v1.7, proportionally
sampled across its mixture components), streamed live (not pre-packed).distillation.teacher_schedule=none in
src/run_distillation.py), AdamW lr=3e-5, betas (0.9, 0.95), weight decay 0.1, cosine
schedule, bf16. Only the random seed differs between runs in this collection.stepN for the run's checkpoint cadence (see run.json for the
exact token budget / save_steps used), weights only, bf16 safetensors. main holds the
final weights once the run completes.Provenance (run.json): {"cluster": "makau", "data": "emozilla/dolma-v1_7-3B (streamed)", "git_commit": "unknown", "gpu": "H100", "jobs": [{"job": "9503", "start": "2026-10-29 22:13:56"}], "max_steps": 30517, "model": "olmo2_medium_dense (~305M)", "recipe": "distillation_olmo2_medium_mechanism, teacher_schedule=none", "save_steps": 950, "seed": 10, "token_budget": 2000000000}