GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
ceselder
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
ceselder
Model Tree
Qwen/Qwen3.6-27B
Input Modalities
Output Modalities
Supported Functionality
Twin of ceselder/skip-lens-qwen36-27b-agreelens-a: exact same 244,367
(position, on-policy 12-token span) pairs, but the input activation is the layer-62
residual instead of layer-42. Intended test-time feed: J_42->62 @ h42 (the official released
Jacobian, camilablank/workspace-lenses) — the skip-lens trick. Same recipe otherwise
(bs 64, lr 1e-4, LoRA r64 α16 rsLoRA scope-all, 1 epoch, seed 0).
Fedlayer workspace agreement (Sonnet vs official J, 353 items): 0.755 J-fed / 0.705 raw-h42-fed (L42 twin: see sibling repo). The J-fed L62 arms are the strongest workspace readers in the agreelens series — skip-lens+Jacobian adds +0.14–0.17 over matched-L42 training.
Usage: PEFT LoRA on Qwen3.6-27B; capture h42 at the read position, apply J_42->62, inject at marker ㈜ (id 158983) via Karvonen norm-matched hook; template in bundled nla_meta.yaml.