Qwen 3.6 35B A3B leads local models in niche code understanding

In a recent hands-on evaluation, a researcher tested four new small local LLMs on their ability to understand highly niche academic code. The models—Qwen 3.6 35B A3B, Qwen 3.6 27B, Gemma 4 26B A4B, and Nemotron 3 Nano—all showed dramatic improvement over earlier small models like Devstral Small 2. The key technical innovation is the combination of gated delta net, hybrid Mamba2, and sliding window attention, which allows these models to handle very long contexts efficiently. This enabled the researcher to feed an entire academic paper along with its accompanying code and ask the model to explain how the code maps to the paper. The Qwen 3.6 35B A3B (35B total parameters, 3B active) emerged as the top performer, significantly outperforming its peers in understanding niche code that was likely not in its training set.

The findings suggest that an intelligent human paired with any of these four local models is now more capable than using a larger cloud model like Opus 4.7 alone. The researcher noted that even the Qwen 3.6 27B and Gemma 4 26B A4B demonstrated strong comprehension, though memory limitations (32GB RAM) prevented testing Devstral Small 2 with long contexts. The researcher expressed hope that Mistral will release a new small model with a gated delta net architecture, as it could potentially dethrone Qwen. This progress marks a major milestone: small, locally-run models can now serve as true research assistants for specialized domains, democratizing access to advanced AI without relying on API calls or expensive hardware.