QUASAR 4-bit QAT checkpoints for Qwen3.5, Gemma 4 and Muse Glimmer: closer to BF16 than Google’s and NVIDIA’s 4-bit releases

QUASAR has released new 4-bit quantization-aware training (QAT) checkpoints for Qwen3.5-4B, Gemma 4 E4B/12B, and Muse Glimmer 30B. These models are available in NVFP4 for vLLM and Q4_0 GGUF for llama.cpp, Ollama, and LM Studio, requiring no custom kernels or conversion steps.

Performance vs. Reference Releases

The checkpoints utilize quantization-aware distillation, where the original BF16 model acts as the teacher to "heal" the quantized model. This approach results in lower KL divergence (closer to BF16) compared to specific reference releases:

• Muse Glimmer 30B (NVFP4): KL divergence of 0.0183 vs. 0.0300 for NVIDIA's NVFP4 release.
• Gemma 4 12B (W4A16): KL divergence of 0.034 vs. 0.044 for Google's QAT.
• Qwen3.5-4B (NVFP4): KL divergence of 0.0181 vs. 0.0419 for cyankiwi's AWQ.

Benchmark Highlights

• Muse Glimmer 30B: Matches or exceeds BF16 performance on GPQA-Diamond (63.6 vs 62.1), MMLU-Pro (83.3 vs 83.1), and AIME'25 (90.0 vs 86.7). It is the only 4-bit build to maintain BF16 performance on RULER at 32K, 64K, and 128K context lengths.
• Qwen3.5-4B: Achieves the best AIME'25 (74.2) and GPQA-Diamond (77.9) scores among tested 4-bit builds.
• Gemma 4 E4B/12B: Outperforms Google's QAT on all downstream benchmarks tested. On H100 hardware, E4B sees a +36% increase in batched throughput using vLLM's Machete kernel.
• Qwen3.8-27B: Previously released NVFP4 checkpoint scores 90.9 on GPQA-Diamond (vs 91.4 BF16) with a memory footprint of 19.7 GB.