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
Part of the skip-lens project (multi-token NL activation lenses on Qwen3.6-27B).
The future-lens further trained with GRPO RL, where the reward is how well a frozen activation reconstructor (AR) recovers the layer-62 activation from this lens's generated explanation.
Naming note. This is the checkpoint we informally called the "yolo NLA." To be precise: it is RL'd against a frozen AR reward model (a separate, pre-trained AR passed as
--ar-ckpt) — it is not jointly/co-trained with the AR. The frozen-AR reward is exactly the setup used throughout; no checkpoint here bundles a co-trained AR.
Sibling: skip-lens-qwen36-27b-naive-futurelens is the plain SFT future-lens (no RL).
Load as a PEFT adapter (PeftModel.from_pretrained) + inject a layer-62 activation via the Karvonen hook.