Use on OpenRouter · Use on Vercel AI Gateway · Release blog post poolside/Laguna S 2.1 DFlash NVFP4 DFlash speculator for the NVFP4 target poolside/Laguna S 2.1 NVFP4. The speculator is a 6 layer Laguna style draft model (BF16); pair it with the NVFP4 base for lower latency serving via speculative decoding. Trained: e0630 rhiemann baseline SFT, DFlash Stage 2, 15k steps. Recommended serving setting: num speculative tokens=7 . DFlash upstream support is in progress (vLLM 46853, SGLang 29446, TRT LLM 15666). Use poolside/Laguna S 2.1 NVFP4 as the target model. Benchmarks Measured with TP=2, temperature=0 , and num speculative tokens=15 . Throughput speedup Concurrency GSM8K MATH 500 HumanEval MBPP MT Bench : : : : : : 1 3.324x 2.893x 3.692x 2.426x 2.338x 4 2.498x 2.174x 2.742x 1.848x 1.831x 8 2.279x 1.948x 2.634x 1.719x 1.772x 16 2.302x 1.965x 2.626x 1.731x 1.672x Acceptance length Concurrency GSM8K MATH 500 HumanEval MBPP MT Bench : : : : : : 1 5.775 4.942 6.438 4.171 4.017 4 5.804 4.935 6.298 4.161 4.091 8 5.719 4.885 6.537 4.193 4.373 16 5.758 4.896 6.412 4.123 3.981
We use cookies for essential functionality and analytics. You can accept or reject analytics cookies.Cookie policy