KE

KernelBench RLVR 120b

KernelbenchGGUF by mradermacherUpdated 2026-08-23
Download on Hugging Face
Parameters117B
Context128K
Layers36
Architecturegpt-oss
Smallest quant0.2 GB
Quick answer

KernelBench RLVR 120b needs about 89.0 GB of memory at Q4_K_M (a 87.9 GB file) with an 8K context window, which is more than a 48 GB GPU holds. This repository has 23 quants, from 0.2 GB to 93.9 GB (Q5_K_M). Longer context needs more: at 32K tokens the KV cache alone adds about 2.4 GB.

KernelBench RLVR 120b – 120B‑parameter GPU kernel generation model

KernelBench RLVR 120b is a 120B‑parameter model fine‑tuned with GRPO for GPU kernel generation. It is trained on a 131072‑token context and supports coding and tools abilities. The model is released under the Apache‑2.0 license.

Good at

  • GPU kernel generation
  • Speed‑optimized code
  • Execution‑grounded RL training

Limits

  • No vision support
  • Large VRAM requirement for full precision
  • Requires 131072‑token context

Best for

  • Optimizing CUDA kernels
  • Research on compute‑optimal strategies
  • Testing GPU performance

Who it’s for. Developers and researchers needing high‑quality GPU kernel generation. The model is suitable for those who can allocate sufficient VRAM and have access to a 131072‑token context window.

Quants & memory

QuantFileEst. memoryvs VRAMResultDownload
—0.2 GB 1.3 GB
Fits Get file →
IQ1_M66.1 GB 67.2 GB
Too big Get file →
IQ1_S66.1 GB 67.2 GB
Too big Get file →
IQ2_XXS66.1 GB 67.2 GB
Too big Get file →
IQ2_XS66.1 GB 67.2 GB
Too big Get file →
Q3_K_S66.1 GB 67.3 GB
Too big Get file →
IQ2_M66.2 GB 67.3 GB
Too big Get file →
IQ2_S66.2 GB 67.3 GB
Too big Get file →
IQ3_S66.2 GB 67.3 GB
Too big Get file →
IQ3_XS66.2 GB 67.3 GB
Too big Get file →
IQ3_XXS66.2 GB 67.3 GB
Too big Get file →
Q2_K66.2 GB 67.3 GB
Too big Get file →
IQ4_XS66.2 GB 67.3 GB
Too big Get file →
Q4_066.5 GB 67.6 GB
Too big Get file →
Q2_K_S66.6 GB 67.7 GB
Too big Get file →
IQ3_M66.7 GB 67.8 GB
Too big Get file →
Q3_K_M71.1 GB 72.2 GB
Too big Get file →
Q3_K_L73.4 GB 74.5 GB
Too big Get file →
Q4_173.5 GB 74.6 GB
Too big Get file →
Q4_K_S80.9 GB 82.0 GB
Too big Get file →
Q4_K_M87.9 GB 89.0 GB
Too big Get file →
Q5_K_S88.0 GB 89.1 GB
Too big Get file →
Q5_K_M93.9 GB 95.0 GB
Too big Get file →

For 8 GB VRAM, use Q4_K_M (87.85 GB file, ~89 GB memory at 8 k context). For 16 GB VRAM, Q5_K_M (93.94 GB file, ~95 GB memory). For larger budgets, IQ1_M or IQ1_S (66.08 GB file, ~67.2 GB memory) provide better quality. Estimates are approximate.

What GPU do you need for KernelBench RLVR 120b?

These figures add the file size, an fp16 KV cache at an 8K context window (worked out from this model’s own header) and about 0.5 GB for the runtime. They are estimates, so leave a little headroom, and expect to use less context or a smaller quant if you are right at the limit.

  • 4 GB VRAM: the best quant that fits is the main file (0.2 GB file, about 1.3 GB in use).
  • 6 GB VRAM: the best quant that fits is the main file (0.2 GB file, about 1.3 GB in use).
  • 8 GB VRAM: the best quant that fits is the main file (0.2 GB file, about 1.3 GB in use).
  • 12 GB VRAM: the best quant that fits is the main file (0.2 GB file, about 1.3 GB in use).
  • 16 GB VRAM: the best quant that fits is the main file (0.2 GB file, about 1.3 GB in use).
  • 24 GB VRAM: the best quant that fits is the main file (0.2 GB file, about 1.3 GB in use).
  • 32 GB VRAM: the best quant that fits is the main file (0.2 GB file, about 1.3 GB in use).
  • 48 GB VRAM: the best quant that fits is the main file (0.2 GB file, about 1.3 GB in use).

Which KernelBench RLVR 120b quant should you choose?

Q4_K_M is the usual starting point for KernelBench RLVR 120b: it balances file size and quality well. Here it is a 87.9 GB file.

If memory is tight, the smallest file is the first one in the table at 0.2 GB, with the quality trade-offs that come with fewer bits.

New to quant names? Read GGUF quantization explained.

Specifications

Architecture from the GGUF header

Architecturegpt-oss
Parameters117B (approximate, mixture of experts)
Layers36
Trained context131,072 tokens
KV cache (fp16)72 KB per token

Source

Quantized bymradermacher
Licenseapache-2.0
Downloads (HF)686
Last updated2026-08-23

Frequently asked questions about KernelBench RLVR 120b

How much VRAM does KernelBench RLVR 120b need?

About 89.0 GB at Q4_K_M with an 8K context window. That is the 87.9 GB file, plus the KV cache and roughly 0.5 GB of runtime overhead. Smaller quants need less and longer context needs more.

Can I run KernelBench RLVR 120b on an 8 GB GPU?

Yes, with the smaller files: the 0.2 GB file needs about 1.3 GB at 8K context. Larger quants in this repository do not fit in 8 GB.

What is the file size of KernelBench RLVR 120b GGUF?

It depends on the quant: file is 0.2 GB, IQ1_M is 66.1 GB, IQ1_S is 66.1 GB, IQ2_XXS is 66.1 GB, IQ2_XS is 66.1 GB, Q3_K_S is 66.1 GB and 17 more.

What context length does KernelBench RLVR 120b support?

The GGUF header lists a trained context length of 131,072 tokens (128K). The KV cache grows with the context you use: about 0.6 GB at 8K and 2.4 GB at 32K.

Where can I download KernelBench RLVR 120b GGUF?

The files are on Hugging Face in the mradermacher/KernelBench-RLVR-120b-i1-GGUF repository. The download buttons on this page open the file on Hugging Face. ggufmodels.com does not host any model files.

What license does KernelBench RLVR 120b use?

Hugging Face lists the license as apache-2.0. Licenses can differ between the quantized files and the original model, so check the repository and the base model before commercial use.

Can I run KernelBench RLVR 120b on 8 GB VRAM?

The Q4_K_M quant requires about 89 GB of VRAM at 8 k context, so 8 GB VRAM is insufficient.

Does the model support vision tasks?

No, the model has no vision support.

What license applies to this model?

The model is released under the Apache‑2.0 license.

Other GGUFs of KernelBench-RLVR-120b

Similar models

Browse all →

More from mradermacher

All by mradermacher →

More Kernelbench models

All Kernelbench →

Other large models

All large models →