GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
AlephFunk
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
AlephFunk
Model Tree
Qwen/Qwen3-1.7B-Base
Input Modalities
Output Modalities
Supported Functionality
Parent vs this model, precision used for both: bf16. Evaluation suite digest 3865a9dda944, sweep-size evaluation (PIQA/ARC: 400 items each, zero-shot log-likelihood; synthetic probes: 160 items each, few-shot, seeded).
| metric | parent | this model |
|---|---|---|
| wikitext-2 test perplexity (1024-token windows) | 10.229 | 12.660 |
| pile-10k sample perplexity (512 tokens) | 9.025 | 11.082 |
| PIQA acc_norm | 0.740 | 0.735 |
| ARC-Easy acc_norm | 0.698 | 0.645 |
| ARC-Challenge acc_norm | 0.427 | 0.378 |
| swap-tracking (chance 0.25) | 0.500 | 0.269 |
| location-tracking (chance 0.20) | 0.675 | 0.531 |
| multi-step variable arithmetic, ranked (chance 0.20) | 0.725 | 0.487 |
| multi-step greedy-generation exact match (n=80) | 0.550 | 0.362 |
Paired change in wikitext loss: +0.213 +/- 0.010 nats/token (95% CI). Probe accuracies have 95% CIs of roughly +/-0.04 (PIQA/ARC) to +/-0.12 (synthetic probes); differences smaller than that are noise. Evaluation was run through an in-memory layer view of the parent; on the 8B recipe this was verified to be bit-identical to loading the written checkpoint.
Produced by the laco-lab subproject (CLI lacolab, 81 tests, audit of Mercity's LaCo-8B, full report and experiment manifest) in the jev-qwen repository of MoralityLabAI (branch add/laco-lab-layer-compression). Rebuild: lacolab compress --model Qwen/Qwen3-1.7B-Base --recipe runs/c17/d3.json --out <dir>.