AI FRONTIER
Signals worth understanding before they become obvious.
OpenRouter new model: Reflection: Beam 1.0 501B A23B
Beam 1.0 is Reflection's open-weight reasoning model for coding and long-horizon agentic workflows. It uses a sparse mixture-of-experts architecture with 501B total and 23B active parameters per token, and a 1M-token context window.
- Sparse mixture-of-experts architecture with 501B total and 23B active parameters per token
- 1,048,576-token context window with support for up to 131,072 completion tokens
- Released under Apache 2.0 with configurable reasoning effort, function calling and structured outputs
- Priced at $0.25 per 1M input tokens and $1.00 per 1M output tokens; served by 4 providers on OpenRouter (Nebius Token Factory, Together, CoreWeave, Baseten), released Oct 16, 2026 with a June 2026 knowledge cutoff
With only 23B active parameters per token against 501B total, builders can aim for large-model capacity at low per-token compute, but should measure whether accuracy holds on long-horizon coding agent tasks.
DeepSeek-V4.1-Flash-1017 released: 763B multimodal model under MIT
DeepSeek released DeepSeek-V4.1-Flash-1017 as a minor update to DeepSeek-V4.1-Flash, reporting gains across reasoning, coding, agentic tasks, and visual benchmarks while reducing task execution time and output length.
- Model size is 763B parameters with tensor types BF16, F32, F8_E4M3, and I8, and the pipeline is image-text-to-text.
- The technical report states HLE without tools rose from 36.8 to 42.6, Terminal-Bench 4.0 from 31.2 to 42.3, and DeepSWE v1.1 from 74.2 to 77.0.
- The report states Chartography with tools rose from 78.9 to 83.0 and BabyVision with tools from 89.6 to 92.5.
- The repository and model weights are MIT licensed, with recommended sampling of temperature 1.0, top_p 1.0, a 1M token context window, and max_tokens of at least 256K.
Builders get 763B multimodal weights under MIT with a 1M context window and a Rust/Python prompt-encoding toolkit, lowering integration cost for agentic and visual pipelines.
DeepSeek Releases V4.1-Flash-1017, Scores 77.0 on DeepSWE, Opens Search API
DeepSeek released V4.1-Flash-1017, which is reported to score 77.0 on DeepSWE, and opened a search API.
- DeepSeek released V4.1-Flash-1017.
- The model is reported to score 77.0 on DeepSWE.
- DeepSeek opened a search API.
Builders can pair V4.1-Flash-1017's search API with its reported DeepSWE 77.0 score for retrieval-augmented agent workflows.