AI Briefing
KO

Unsloth Tops the Benchmark

·2026.04.18 01:17

Key point

In the Qwen3.6-35B-A3B GGUF benchmark, Unsloth quants claim the best size efficiency relative to KLD.

Details

They released a KLD performance benchmark for Qwen3.6-35B-A3B GGUF, and stated that Unsloth quants took the Pareto frontier in 21 out of 22 cases based on KLD versus disk size.

The GGUF release is available at unsloth/Qwen3.6-35B-A3B-GGUF on Hugging Face.

Additionally, regarding the controversy over GGUF re-uploads, they explained that in most cases the cause was external factors—such as llama.cpp bug fixes or template improvements from upstream models—rather than internal mistakes.

Examples cited include:

  • Gemma 4 was re-uploaded a total of 4 times, 3 of which were due to reflecting various llama.cpp fixes.
  • The last of those 4 was due to reflecting Google's official chat template improvement.
  • In the investigation of MiniMax 2.7 NaNs, they noted that NaNs were found in 38% (10/26) of Bartowski quants and 22% (5/23) of Unsloth quants.
  • Unsloth stated that a fix for this issue has already been applied.
benchmarkqwenquantizationunslothgguflocal-llmllama-cpp
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

RedditHugging Face Introduces Native GGUF Support in TransformersRedditDynamic Quantiser Released, Improving Accuracy Over Standard Quantization in VRAM-Constrained EnvironmentsHuggingFace BlogHugging Face Adds GGUF Model Support to TransformersTLDR AIQwen Intelligence Launches Three Mobile AI Agents and Open Benchmark SuitePyTorchKRPrismML Releases 'Ternary Bonsai 2 27B', a Ternary Quantized Model Maintaining 98.2% of FP16 PerformanceHacker NewsApple Releases LensVLM-9B for Processing Compressed Text Images