Updated

Updated · Unite.AI · Aug 25

Multiverse Computing Says 4-Bit 60B Model Beats 16-Bit Parent on 7 of 9 Benchmarks

Updated

Updated · Unite.AI · Aug 25

Multiverse Computing Says 4-Bit 60B Model Beats 16-Bit Parent on 7 of 9 Benchmarks

Q: How did a heavily compressed 4-bit AI model manage to outsmart its massive full-precision predecessor on complex reasoning tasks?

It did so by changing the recovery process after compression. Multiverse says the 60B model was not merely shrunk and quantized to 4-bit MXFP4; it was then “healed” using Quantization-Aware Healing, a distillation method that trains the compressed student directly against the original 120B model’s output probabilities. That matters because the usual pipeline recovers a compressed model from an already degraded intermediate checkpoint, typically a bfloat16 version of the smaller architecture. In Multiverse’s view, that creates a ceiling: the 4-bit student can only learn from a weaker teacher. QAH removes that bottleneck by matching the student to the original 120B model with a KL-divergence loss, so the smaller model inherits richer signal about how the larger model weighs alternative answers, not just the final label. The result, according to the company’s benchmarks, is that the healed 60B 4-bit model, HyperNova-60B, beat the recovered bfloat16 version of the same architecture on 7 of 9 tests. The biggest gains were in long-context reasoning and competition math, where it scored 42.7 versus 35.3 on AA-LCR and 76.3 versus 70.7 on AIME 2025. Losses on MMLU-Pro and SciCode were small. Multiverse also reports that QAH trained faster and more stably than standard quantization-aware training, reaching peak quality in about 100 steps in a 9B test, versus roughly 700 for QAT, while avoiding the post-peak collapse it observed with the standard method.

1 articles · Updated · Unite.AI · Aug 25

Multiverse Computing said its Quantization-Aware Healing method let a GPT-OSS 120B derivative shrink to 60B and 4-bit MXFP4 while outperforming the recovered 60B bfloat16 checkpoint on 7 of 9 tests.
7.4 points on long-context reasoning and 5.6 on AIME 2025 math drove the gains, while the two losses—MMLU-Pro and SciCode—were under 1.5 points.
QAH changes the recovery step after compression by distilling the 4-bit student directly from the original 120B model, rather than from an intermediate full-precision checkpoint that Multiverse argues caps performance.
100 training steps were enough for QAH to reach peak scores in a 9B test, versus about 700 for quantization-aware training; Multiverse said QAT then lost nearly 19 points by step 1,200 while QAH stayed within two points of peak.
HyperNova-60B, released open-weight under Apache 2.0, uses roughly one-quarter the weight memory of a bfloat16 version and about half the teacher’s compute per token, though the results come from Multiverse’s own paper and blog.

Sources

Unite.AI4h ago

Multiverse Computing's QAH Enables 4-Bit LLM to Outperform Full-Precision Model

How did a heavily compressed 4-bit AI model manage to outsmart its massive full-precision predecessor on complex reasoning tasks?

If stripping an AI to 4-bit precision actually improves its math skills, are we fundamentally misunderstanding how neural networks store knowledge?

What is the hidden cost of bypassing intermediate checkpoints to train a compressed AI directly from its massive original teacher?

Sources

1 total

Unite.AI4h ago

Multiverse Computing's QAH Enables 4-Bit LLM to Outperform Full-Precision Model

Multiverse Computing Says 4-Bit 60B Model Beats 16-Bit Parent on 7 of 9 Benchmarks

Summary

Sources

Insights

Related Stories

Sources

Related Stories