Qwen3.5-27B
A 27B-parameter version of Alibaba's Tongyi Qianwen (Qwen) series, belonging to the family of open-source large language models.
Timeline (last 90 days)
RL can let a 4B model close more than half the gap to a 27B model, but the gains have a clear ceiling.
In the experiments, the 4B model's accuracy before training was about fifty percent, while the 27B model reached over seventy-five percent.
After training for most of a day using only about 1,000 edge tasks that are 'sometimes done right and sometimes done wrong,' the 4B model recovered more than half the gap to the 27B model.
After using the 27B model to generate correct trajectories to distill the 4B model, the distilled model solved some problems that the pure-RL model had never gotten right.
Some users pointed out that Nvidia's 550B-parameter Nemotron 3 Ultra (released in June) is only on par with Qwen3.5-27B, which was released in February.
The Authoritative Actor model in SMC is a large, accurate model responsible for generating the final adopted execution trajectory; in the experiments, a Qwen3.5-27B INT4 quantized version was used.
All Facts (6)
RL can let a 4B model close more than half the gap to a 27B model, but the gains have a clear ceiling.
50%UnverifiedIn the experiments, the 4B model's accuracy before training was about fifty percent, while the 27B model reached over seventy-five percent.
50%UnverifiedAfter training for most of a day using only about 1,000 edge tasks that are 'sometimes done right and sometimes done wrong,' the 4B model recovered more than half the gap to the 27B model.
50%UnverifiedAfter using the 27B model to generate correct trajectories to distill the 4B model, the distilled model solved some problems that the pure-RL model had never gotten right.
50%UnverifiedSome users pointed out that Nvidia's 550B-parameter Nemotron 3 Ultra (released in June) is only on par with Qwen3.5-27B, which was released in February.
50%UnverifiedThe Authoritative Actor model in SMC is a large, accurate model responsible for generating the final adopted execution trajectory; in the experiments, a Qwen3.5-27B INT4 quantized version was used.
50%