Jan-nano Technical Report
This addresses the tradeoff between capability and computational resources for language model users, though it appears incremental as it builds on Qwen3-4B.
The paper tackles the computational efficiency problem in language models by introducing Jan-nano, a 4B parameter model that achieves 83.2% on the SimpleQA benchmark while running on consumer hardware.
Most language models face a fundamental tradeoff where powerful capabilities require substantial computational resources. We shatter this constraint with Jan-nano, a 4B parameter language model that redefines efficiency through radical specialization: instead of trying to know everything, it masters the art of finding anything instantly. Fine-tuned from Qwen3-4B using our novel multi-stage Reinforcement Learning with Verifiable Rewards (RLVR) system that completely eliminates reliance on next token prediction training (SFT), Jan-nano achieves 83.2% on SimpleQA benchmark with MCP integration while running on consumer hardware. With 128K context length, Jan-nano proves that intelligence isn't about scale, it's about strategy.