GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
cds-jb
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
cds-jb
Model Tree
Base
Qwen/Qwen3-8B
Adapter
this model
Input Modalities
Text
Output Modalities
Text
Supported Functionality
Dedicated Endpoints
Container
| metric | value |
|---|---|
| operating accuracy (organism mask, all dots) | 0.893 |
| ablated control (dots blinded to the prompt) | 0.073 (chance 0.10) |
| load-bearing gap | 0.820 |
| train accuracy (provably-seen training instances) | 0.922 |
| test accuracy (fresh instances) | 0.920 |
Per-query ensemble completeness (same instances, every possible query — the latents serve ALL queries from one fixed computation, including threads never named in any text):
Final value of chain 1: : 1.000Final value of chain 2: : 0.993Final value of chain 3: : 1.000Chain number with the largest final value: : 0.047Chain number with the smallest final value: : 0.147Last digit of the sum of the three final values: : 0.813Training data is an infinite procedural stream (every instance seen at most once); "train accuracy" evaluates instances bit-exactly replayed from the training seed, "test" a disjoint seed — the match shows the organism runs the algorithm rather than memorizing instances.
raw workspace, not an answer cache: finals v1/v2/v3 decode at ~1.0 across the block but the SUM is at chance (0.06-0.16) at every dot - yet behavioral sum accuracy is 0.81, so the aggregate is assembled downstream in the masked suffix from raw latent operands. Training history: suffix-side aggregation over a latent operand needed ~3x the stage budget to train (v2 stalled at the first latent operand; a dot-side sum variant never trained at all).
See examples.md for a full operating transcript (the model sees the scenario,
emits only dots in <think>, then the query is revealed and it answers from the dot
activations) and the surface-CoT version of the same instance that the curriculum started from.
lt_cfg.json (full config)training_code/ — complete training/eval/probe code snapshot (see its README)plots/ — load-bearing eval, probe grid + logit lens, training curveActivation-oracle / interpretability research target: the reasoning trace exists ONLY in the
dot activations (logit lens is near-dark), with deterministic ground truth from the task
generator. Wired into the cds-jb/AVBench recog eval as suite latent_threads (token-exact
rows; row_metadata.selected=False rows query threads with no textual trace anywhere in the
transcript). Part of the latent-threads collection together with its four siblings.
Trained 2026-06-12 (wandb group lt3g, MATS10-CS-JB/cot-oracle). Built on the bottleneck-mask methodology of the pointer-chase filler organism (cds-jb/qwen3-8b-pointer-chase-filler-cot).