reflection/Beam-1.0-501B-A23B-DFlash2

Beam 1.0 DFlash2

DFlash2 draft model for Beam. It is not a standalone language model: a speculative decoding server runs it beside Beam, and Beam verifies every drafted token. With rejection sampling, the output distribution matches Beam's exactly.

Model

DFlash2 drafts a whole block of tokens in one forward pass. It keeps the top candidates at each position in the block, and a small selector picks one path through them. Two-tap dynamic convolutions within the block keep accuracy from decaying toward its end.

The drafter conditions on hidden states from Beam and reuses Beam's frozen embedding and LM head. Its attention over the context uses a 4,096-token sliding window, so drafting cost does not grow with context length.

• Parameters: 1.93B
• Architecture: 5 layers, hidden size 5120, 40 query and 8 KV heads of dimension 128, SwiGLU intermediate size 17,408
• Beam layers read: 4, 16, 28, 40 and 46
• Vocabulary: 151,936 tokens, shared with Beam
• Block size: trained with 16, 8 recommended for serving

Serve

Serve this drafter with Beam in bf16, FP8 or NVFP4:

• SGLang recipe
• vLLM recipe
• TokenSpeed recipe

Evaluation

All results use Beam NVFP4 with tensor parallelism 4 on B300 GPUs, and a serving block size of 8, so each verification step checks 7 drafted tokens.

Acceptance length

Mean number of tokens generated per verification step, out of a maximum of 8, with up to 4,096 new tokens per request. Sampled runs use temperature 0.7 and top-p 0.99, Beam's recommended settings.

Benchmark	Sampled	Greedy
GSM8K	5.60	5.88
MATH-500	5.05	5.39
HumanEval	4.58	4.68
MBPP	4.19	4.32
MT-Bench	3.93	4.11

Benchmark	Sampled	Greedy
GSM8K	5.60	5.88
MATH-500	5.05	5.39
HumanEval	4.58	4.68
MBPP	4.19	4.32
MT-Bench	3.93	4.11

Long context

Acceptance length on agentic coding traces from DeepSWE, by prompt length. We replay 446 recorded Beam turns, originally sampled at temperature 0.7 and top-p 0.99, with up to 64 turns per bucket.

Prompt length	Acceptance length
0 to 2k	5.72
2k to 4k	5.39
4k to 8k	4.91
8k to 16k	4.66
16k to 32k	4.22
32k to 64k	4.58
64k to 128k	4.23
All	4.82

Prompt length	Acceptance length
0 to 2k	5.72
2k to 4k	5.39
4k to 8k	4.91
8k to 16k	4.66
16k to 32k	4.22
32k to 64k	4.58
64k to 128k	4.23
All	4.82

Decode speedup

Output tokens per second without and with the drafter, at 1, 16 and 32 concurrent requests, with greedy decoding.

Benchmark	Concurrency 1	Concurrency 16	Concurrency 32
GSM8K	288 → 874 (3.03×)	2,515 → 6,312 (2.51×)	3,931 → 10,799 (2.75×)
MATH-500	289 → 864 (2.99×)	2,483 → 6,012 (2.42×)	3,869 → 9,756 (2.52×)
HumanEval	289 → 733 (2.54×)	2,459 → 5,406 (2.20×)	3,799 → 8,896 (2.34×)
MBPP	289 → 665 (2.30×)	2,434 → 5,031 (2.07×)	3,796 → 8,399 (2.21×)
MT-Bench	289 → 565 (1.96×)	2,305 → 3,967 (1.72×)	3,487 → 6,573 (1.89×)

Benchmark	Concurrency 1	Concurrency 16	Concurrency 32
GSM8K	288 → 874 (3.03×)	2,515 → 6,312 (2.51×)	3,931 → 10,799 (2.75×)
MATH-500	289 → 864 (2.99×)	2,483 → 6,012 (2.42×)	3,869 → 9,756 (2.52×)
HumanEval	289 → 733 (2.54×)	2,459 → 5,406 (2.20×)	3,799 → 8,896 (2.34×)
MBPP	289 → 665 (2.30×)	2,434 → 5,031 (2.07×)	3,796 → 8,399 (2.21×)
MT-Bench	289 → 565 (1.96×)	2,305 → 3,967 (1.72×)	3,487 → 6,573 (1.89×)

Citation

If you find our work helpful, please consider citing it:

@misc{reflection2026introducingbeam,
  author = {{Reflection}},
  title  = {Introducing {Beam}: {Reflection}'s 501B open-weight model},
  year   = {2026},
  month  = oct,
  url    = {https://reflection.ai/blog/introducing-beam}
}

@inproceedings{chen2026dflash,
  title     = {{DFlash: Block Diffusion for Flash Speculative Decoding}},
  author    = {Chen, Jian and Liang, Yesheng and Liu, Zhijian},
  booktitle = {International Conference on Machine Learning (ICML)},
  year      = {2026}
}

@misc{inco2026dflash2,
  title  = {{DFlash 2: Keep Drafting Parallel}},
  author = {{Inco AI}},
  year   = {2026},
  month  = aug,
  url    = {https://inco.ai/blog/dflash2/}
}