GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
SohamProgrammer
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
SohamProgrammer
Model Tree
Input Modalities
Output Modalities
Supported Functionality
~356K steps / ~1.1B tokens. Chain: Stage C → CoT (math+code reasoning) →
Refresh (240M topper) → Realization (judgment tasks) → SFT (31.7M incl.
OpenAssistant) → Wiki entities+math (13.35M) → SFT-refresh → DPO
(UltraFeedback, β=0.5) → diverse-SFT (smol-smoltalk + UltraChat + task
synthetics) → pristine anchors. AdamW throughout. Full log:
docs/TRAINING_REPORT.md.
step_355789)evals/benchmarks/final/results.json
(ARC-Easy, PIQA, HellaSwag-sample, Winogrande, TruthfulQA-MC1,
MMLU-easy, GSM8K-sample, MBPP-sample)docs/EVAL_REPORT.mdpython
from transformers import AutoModelForCausalLM, AutoTokenizertok = AutoTokenizer.from_pretrained("REPO_ID")model = AutoModelForCausalLM.from_pretrained("REPO_ID")
On Apple Silicon the native MLX checkpoint + scripts/spot_infer.py is fastest.
Suggested decode: temperature 0.4, top_p 0.9, repetition_penalty 1.25.
GGUF quants (Q8_0 recommended, Q4_K_M small-device, F16 reference) + a
local-Ollama Modelfile ship alongside. (Ollama library listing excluded:
publisher under 18; local use unaffected.)
Weak: large-number word problems, deep code, biology, multi-step reasoning, instruction-following beyond basics, non-English. May hallucinate entities and dates. 190M memorization model, not a reasoner.
Weights: Apache-2.0. Training-data attribution: docs/DATA_REPORT.md
(CC BY-SA content attributed; 0.6% GPL-family trace in code data measured
and documented; no verbatim reproduction possible at this scale).