最新模型
Reflection Releases Beam 1.0 Weights, a 501B MoE with Three Precision Options and a Speculative Decoding Model
On October 16, Reflection released Beam 1.0 weights under the Apache 2.0 license in BF16, FP8, and NVFP4 versions. The model, aimed at software development and long-horizon agent tasks, also comes with a DFlash2 draft model, though its Chinese-language capabilities and real-world deployment costs still need validation.
Reflection officially released Beam 1.0 weights on October 16, moving the model from its preview announcement in early October to a downloadable, self-hosted release. The release includes BF16, FP8, and NVFP4 versions, as well as a speculative decoding draft model. The weights are licensed under Apache 2.0, allowing teams to deploy and customize the model themselves. Official release announcement
Beam uses a sparse mixture-of-experts architecture, with about 501 billion total parameters and about 23 billion activated per token. It supports a context window of up to one million tokens and accepts and generates text. The model card lists software development, tool use, and long-horizon agent tasks as its primary use cases, while explicitly noting that the model was developed primarily for English. Chinese engineering teams should evaluate coding ability separately from Chinese-language requirement understanding and document retrieval. Official model card
A notable deployment component in this release is the DFlash2 draft model, with about 1.93 billion parameters. In a single forward pass, it proposes a block of candidate tokens, which Beam then verifies. According to Reflection, when rejection sampling is used, the output distribution can match that of the original model. The draft model reads Beam’s hidden states from multiple layers, reuses its frozen token embeddings and output head, and uses a 4,096-token sliding window to control context costs on the draft side. The training block length is 16; 8 is recommended for serving. DFlash2 model card
For evaluations, the official model card reports 81.6 on SWE-bench Verified and 80.1 on Terminal-Bench 2.1. These are results reported by the publisher and should not be treated as success rates that any agent framework can reproduce. Reflection’s efficiency comparison also has limits: an earlier technical note estimated compute based on the number of activated parameters and generated tokens, excluding prompt prefill, attention costs that vary with context length, and serving overhead. Model evaluations, Efficiency calculation notes
The next engineering step is to hold the agent framework, tool set, and task budget constant, compare success rates across the three precision options, and then measure speculative decoding’s acceptance rate and end-to-end latency. The figure of 23 billion parameters activated per token describes the compute path; memory budgets for the full weights, KV cache, and parallel serving must be accounted for separately. A one-million-token context limit also does not mean that using a long context is inexpensive.