Description
Beam 1.0 DFlash2 is a draft model that operates in conjunction with the Beam language model, providing a unique approach to language processing through speculative decoding. This model is not standalone; instead, it functions as a drafter, generating a block of tokens in one forward pass while maintaining top candidates at each position. A small selector then chooses the optimal path through these candidates. The model's architecture includes five layers with a hidden size of 5120, and it utilizes two-tap dynamic convolutions to prevent accuracy decay toward the end of the block.
Read More
The drafter conditions on hidden states from Beam, reusing Beam's frozen embedding and language model head. It features attention over the context with a 4,096-token sliding window, ensuring that drafting costs do not increase with context length. With 1.93 billion parameters, Beam 1.0 DFlash2 is trained with a block size of 16, although a size of 8 is recommended for serving.
In terms of performance, Beam 1.0 DFlash2 demonstrates impressive results across various benchmarks, including GSM8K, MATH-500, HumanEval, MBPP, and MT-Bench. The model's acceptance length varies with prompt length, maintaining efficiency even with long contexts. Additionally, the model offers significant decode speedup, with output tokens per second increasing substantially across different levels of concurrency.
Beam 1.0 DFlash2 is designed for integration with Beam in bf16, FP8, or NVFP4 formats, using recipes such as SGLang, vLLM, and TokenSpeed. This model is particularly suited for applications requiring high accuracy and efficiency in language processing, making it a valuable tool for developers and researchers in the field of artificial intelligence.
- CategoryAI Models & LLMs
- DeploymentCloud
- PlatformsWeb
Beam 1.0 DFlash2 Highlights
Speculative decoding server
Token drafting with top candidate selection
Two-tap dynamic convolutions
4,096-token sliding window attention
1.93 billion parameters
Five-layer architecture
Integration with Beam's frozen embedding
Support for bf16, FP8, NVFP4
Getting Started with Beam 1.0 DFlash2
Access model: Visit Hugging Face
Authenticate: Use your credentials
Set up environment: Configure system for Beam
Integrate via API: Connect using provided recipes
Optimize: Adjust settings for performance
Beam 1.0 DFlash2's Use Cases
- Language Processing
- Token Drafting
- Contextual Attention
- AI Research
- Model Integration










