GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
newtype-2038
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
newtype-2038
Model Tree
Qwen/Qwen3.5-122B-A10B
Input Modalities
Output Modalities
Supported Functionality
v10 LoRA — multi-turn masking applied to the same 825-row v9 corpus.
Loss is computed only on assistant content tokens; user/tool/observation
tokens are masked to -100. Hypothesis derived from SWE-Hero (2025):
sequence-only training pollutes the loss signal with tool-output noise,
which we believe explains v7/v9's high empty_patch FAIL rate (~33%).
Identical hyperparameters and corpus as v9 — masking is the single independent variable.
Hyperparams: r=16 α=16 lr=5e-5 epoch=2 max_length=4096. Eval: 30-task seed4 smoke result will be posted on completion.
See SFT method doc for full corpus/iteration history.