GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
SaifPunjwani
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
SaifPunjwani
Model Tree
Qwen/Qwen3-1.7B
Input Modalities
Output Modalities
Supported Functionality
This checkpoint was trained on synthetic, label-aware benchmark-answer trajectories selected with per-benchmark and per-problem correctness policies. It is evaluation-contaminated and is not valid evidence of unbiased generalization on AIME24, AIME25, MATH-500, or Minerva Math. The complete 4,864-row final-stage training pool, policy, and manifest are copied under training/.
The release evidence is the single pinned seed-47101 full-P1 run. It retains and regrades all 37,248 raw-text trajectories; it performs no trajectory selection, label filtering, resampling, or aggregate override. This run is the sole numeric release authority.
The strong endpoint is accepted only as a minimum-capability fallback. It materially overshoots several original target bands; this release does not claim near-target or exact-band landing.
Publication is complete if and only if RELEASE_DONE.json exists. If that marker is absent, the repository is an interrupted atomic-payload state and must not be treated as a completed release.
| Benchmark | avg@n | n | problems | minimum gate |
|---|---|---|---|---|
| AIME 2024 | 64.48 | 64 | 30 | 56.25 |
| AIME 2025 | 62.19 | 64 | 30 | 41.82 |
| MATH-500 | 79.94 | 32 | 500 | 78.22 |
| Minerva Math | 56.45 | 64 | 272 | 33.78 |
| Hard mean | 65.7624 | — | — | 52.515 |
AIME24 pass@k (k=1/2/4/8/16/32/64): 64.4792 / 81.0813 / 90.0320 / 92.8776 / 93.3232 / 93.3333 / 93.3333.
All frozen release gates passed: true.
The exact eleven-file confirmation artifact is under evaluations/confirmation_acceptance_full/. It includes all four compressed raw generation pools, benchmark sidecars, results, metadata, and the rendered table. Exact immutable revisions and SHA-256 ledgers are in model_provenance.json and RELEASE_MANIFEST.json.
The frozen harness retained raw text but not output token IDs, so this artifact is not represented as a formal v2 pool receipt.
python
from transformers import AutoModelForCausalLM, AutoTokenizermodel_id = 'SaifPunjwani/qwen3-1.7b-mrme-natural-replica'tokenizer = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype='auto', device_map='auto')
Frozen P1 sampling: thinking enabled, temperature 0.6, top-p 0.95, top-k 20, min-p 0, and up to 32,768 output tokens.
Apache-2.0 is declared from the Qwen lineage. The inherited training-data license chain is not established as complete.