Muse Spark Open 816B A42B DFlash2 released

This release is not a new model to chat with. It is a small helper model, a speculative-decoding drafter, that NVIDIA trained to make Meta's 816-billion-parameter Muse Spark Open generate text faster. Speculative decoding works by having a cheap model guess several upcoming tokens and letting the large model check them all at once; every guess the large model accepts is a token it did not have to produce one step at a time, and the output still comes from the large model's verification.

Per the model card, this drafter uses NVIDIA's DFlash 2 method, which proposes a whole block of candidate tokens in a single parallel pass rather than one after another, then picks one coherent path through the top candidates at each position for the target model to verify. The drafter has five layers and about 4.38 billion parameters of its own, reusing Muse Spark Open's embedding and output layers. It was trained on about 3.48 million prompts whose responses were generated by Muse Spark Open itself, so its guesses imitate the model it serves.

The card reports results as acceptance length, the average number of tokens emitted per verification step. With seven drafted tokens on a GB300 node, that averaged 3.87 on MT-Bench with greedy decoding and 3.49 with sampling, and fell to 2.52 and 2.35 respectively on the high-entropy category of SPEED-Bench's long-input throughput split, where text is less predictable. The card does not translate these into an end-to-end speedup. It targets NVIDIA Blackwell hardware, ships with vLLM and SGLang serving recipes, and is governed by the NVIDIA Open Model Agreement alongside Meta's Muse license.