Jasper-K-32B

Jasper-K-32B is a 32B-parameter model obtained by RL post-training QwQ-32B, developed by Ragentile for GPU kernel optimization. It learns to generate correct kernel implementations and improve their performance through repeated interaction with an execution environment.

Read the Jasper-K technical blog for the training method, evaluation results, and real-workload experiments.

RL post-training

We post-train Jasper-K using multi-rollout, multi-turn GRPO: the model generates candidate kernels, receives compilation, correctness, and measured-performance feedback, and revises its implementations over successive turns. Across trajectories, it explores different optimization strategies; improvement-based rewards encourage it to keep searching for faster implementations after finding a correct one.

RL training uses 180 KernelBenchFixed Level 1 and Level 2 tasks, with another 20 tasks reserved for training monitoring and checkpoint selection. KernelBenchFixed revises tasks that can otherwise admit shortcuts. LeetCuda is our in-house benchmark of 50 kernel optimization tasks, originating from the reinforcement learning team at the University of Alberta, Canada. Its tasks are used only for testing and are excluded from RL training and checkpoint selection.

Results reported in the technical blog

Benchmark	QwQ-32B	Jasper-K-32B	Relative improvement
KernelBench (250 tasks, L1–L3)	1.265×	1.301×	+2.85%
LeetCuda (50 tasks)	1.323×	1.399×	+5.74%

Scores are METR Best GM Speedup @16 × 4: for each task, select the fastest correct implementation found across 16 trajectories of up to four turns, use the reference implementation when no faster correct candidate is found, and take the geometric mean across tasks. The reported comparisons use the same hardware platform, execution environment, correctness checks, timing protocol, and search budget. KernelBench includes training tasks; LeetCuda is held out.

Jasper-K-32B improves over QwQ-32B on both benchmarks, providing evidence that the benefits of RL post-training extend to independently authored tasks. These results use the evaluation settings described in the technical blog.

Other models in the family: Jasper-K-27B · Jasper-K-35B-MoE.

Model files

This repository contains the Jasper-K-32B model weights, tokenizer files, and model configuration.

Download

python3 -m pip install --upgrade huggingface_hub
hf download ragentile/Jasper-K-32B --local-dir ./Jasper-K-32B

Inference harness and usage

The inference harness, evaluator template, feedback examples, and self-test are maintained in Jasper-Harness on GitHub. Follow the GitHub instructions to configure your model path, serving environment, and evaluator.

License

The Jasper-K-32B model weights are released under the Apache License 2.0. See NOTICE for base-model attribution. External dependencies retain their respective licenses and copyright notices.