KODA-35B: Kernel Optimization via Domain Adaptation

KODA-35B is a CUDA kernel generation and optimization model based on Qwen3.6-35B-A3B, with 35B total parameters and approximately 3B active at inference. Given a reference implementation, it proposes a CUDA kernel and revises it across rounds using compiler diagnostics, correctness tests, runtime measurements, and profiler feedback from the KODA agent.

Benchmark results

Correctness and median speedup versus total parameter count on KernelBench and SOL-ExecBench

The charts compare KODA-35B with the Qwen3.6 base model and other models on KernelBench and SOL-ExecBench. The horizontal axis shows total parameter count on a logarithmic scale. The red star marks KODA-35B.

Best values are bold, including ties. Compared with its Qwen3.6 base model, KODA-35B improves correctness by 35.2 percentage points on KernelBench and 31.1 points on SOL-ExecBench, based on the displayed scores. It ties for the highest KernelBench correctness and reaches the second-highest KernelBench median speedup in this comparison.

KernelBench (270 problems)

Model Correctness Fast@1 Fast@2 Median speedup
KODA-35B 95.6% 88.9% 61.9% 3.15ร—
Qwen3.6-35B-A3B 60.4% 24.1% 10.7% 0.94ร—
Qwen3.8-Flash-Next 94.8% 82.2% 51.9% 2.27ร—
DeepSeek-V4.1-Flash-1017 92.6% 90.0% 70.4% 4.31ร—
DeepSeek-V4-Flash-0731 95.6% 77.4% 52.6% 2.26ร—
GLM-5.3 94.1% 84.4% 62.6% 2.97ร—
GLM-5.3-Flash 95.6% 80.0% 48.1% 2.02ร—
Kimi-K3 93.7% 84.8% 64.1% 2.99ร—
MiMo-v2.6-Flash-MOPD 89.6% 71.1% 39.6% 1.72ร—
Recursive-Flint-40B 68.9% 18.1% 9.3% 0.48ร—
MusaCoder-27B 73.0% 26.7% 14.1% 0.95ร—
Kevin-32B 5.2% 1.1% 0.7% 0.64ร—
IndustrialCoder-32B-Thinking 23.7% 3.0% 1.1% 0.27ร—

SOL-ExecBench (202 problems with fixed workload sampling)

Model Correctness Fast@1 Fast@2 Median speedup
KODA-35B 95.5% 78.7% 48.0% 2.10ร—
Qwen3.6-35B-A3B 64.4% 17.8% 8.9% 0.73ร—
Qwen3.8-Flash-Next 95.5% 74.8% 44.1% 1.70ร—
DeepSeek-V4.1-Flash-1017 91.6% 81.7% 54.5% 3.15ร—
DeepSeek-V4-Flash-0731 98.0% 67.3% 41.1% 1.37ร—
GLM-5.3 93.6% 78.7% 50.0% 2.28ร—
GLM-5.3-Flash 96.0% 72.8% 43.6% 1.57ร—
Kimi-K3 94.6% 78.2% 44.6% 1.81ร—
MiMo-v2.6-Flash-MOPD 82.2% 57.4% 32.7% 1.43ร—
Recursive-Flint-40B 65.8% 13.4% 8.4% 0.30ร—
MusaCoder-27B 71.3% 26.7% 14.9% 0.87ร—
Kevin-32B 2.5% 0.0% 0.0% 0.74ร—
IndustrialCoder-32B-Thinking 25.7% 3.5% 2.0% 0.59ร—

Usage

Serve KODA with a compatible inference engine such as vLLM or SGLang, and connect its model API to the KODA agent. The agent compiles and evaluates candidates in managed sandboxes and feeds the results back to the model.

Model serving and kernel evaluation use separate GPU resources. Refer to the agent repository for configuration and evaluation instructions.

Model provenance

This repository provides the KODA model weights, tokenizer, and configuration in Hugging Face format. KODA is fine-tuned from Qwen3.6-35B-A3B.

License

The model is distributed under Apache License 2.0. See LICENSE for the license text and upstream notices. We acknowledge the Qwen team for the base model. The agent code and benchmark datasets have their own licenses.

Downloads last month
-
Safetensors
Model size
35B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Baidu-Baige/KODA-35B

Finetuned
(532)
this model
Quantizations
1 model