NVIDIA Nemotron 3.5 Bolt 74B A9B 24GB LiLiCorr released

This is not a model you chat with. NVIDIA's model card describes it as a draft checkpoint of about 1.1 billion parameters that works only alongside Nemotron 3.5 Bolt 74B-A9B 24GB, the company's 2-bit-quantized model built to run agentic tasks locally in 24GB of memory. The technique is speculative decoding: the small drafter proposes a block of upcoming tokens, the large model verifies the block in parallel, and generation can advance by several accepted tokens per verification step. Because the large model does the checking and produces the visible text, the card says correct verification preserves its output distribution, and the drafter adds no safety controls of its own.

What is specific to this drafter is LiLiCorr. According to its paper, drafters of the DFlash type propose a whole block in one pass but predict each position separately, so the tokens can look fine one by one yet not fit together. LiLiCorr keeps several candidates per position (eight, in this checkpoint, using a 37-million-parameter head) and scores which candidates at neighboring positions pair up, so a more coherent block reaches the large model. The authors report 7 to 19 percent longer accepted runs than plain DFlash in their own tests. Every response the drafter was trained on was generated by Bolt itself.

On NVIDIA's SPEED-Bench test with seven drafted tokens per block, acceptance length, the number of tokens each check advances, averaged 4.30, from 5.26 on coding down to 3.41 on writing. The card gives no tokens-per-second figure and says these results are not tied to a particular serving runtime; it was tested with vLLM on DGX Spark, and the Bolt card's vLLM recipe loads it with seven speculative tokens. Bolt is compressed from Nemotron 3.5 Super VL, which has its own LiLiCorr drafter.