GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
Bottlepick
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Bottlepick
Model Tree
Input Modalities
Output Modalities
Supported Functionality
| base | king v96 (dendriteholdings/albedo-qwen3.6-35b-king-XCVI @ 2a6708b8, 22 shards) |
| adapter | onpolicy_v96_run1 DPO LoRA, checkpoint-100 (r32, alpha32) |
| merge | 350 modules resolved; config.json matches GENESIS_MODEL_CONFIG with 0 field differences |
| arch | qwen3_5_moe, 22 safetensors shards |
The adapter was trained against king v96, so v96 is the only correct merge base — merging it onto a different base produces garbage weights rather than a weaker model.
Scored locally under the SN97 regime that went live on mainnet 2026-07-30 ~13:52 (new evaluator
prompt, z-ai/glm-5.2 as sole SOTA anchor and sole judge), on the 100 tasks published with eval run
a3fd0092. See train_methods/gate_runs/glmonly_20260730/ in the sn97 repo.