the first open-weights model from Reflection, a 501B-total, 23B-active mixture-of-experts built for agentic tool use, coding and general reasoning. it reads text only and its weights ship under Apache 2.0.
pick it if
you want a long-window text model priced with the flash tier
you want Apache 2.0 weights, with BF16, FP8 and NVFP4 builds to self-host
you run tool-calling agents on text and repeat long prompts, so cached input matters
skip it if
you need image, audio or video input; Beam reads text only
long autonomous coding is the job; its reported DeepSWE v1.1 score sits below GLM-5.3 Flash and DeepSeek V4.1 Flash
DeepSeek's larger V4 model: 1.6T total parameters (49B active), text-only, with a 1M-token context and selectable thinking effort. previewed in April 2026 and generally available from 2026-08-13, it sits above V4.1 Flash, which costs less and also reads images.
pick it if
want DeepSeek's largest V4 model with MIT-licensed open weights
resend long prompts: cache hits bill $0.044 per 1M versus $1.32
need low / high / max thinking effort for agent workflows
skip it if
you need image input; V4.1 Flash accepts images
V4.1 Flash at $0.30 / $1.20 already covers your workload
faq
which is better, Beam or DeepSeek V4 Pro?
on Terminal Bench 2.1, as each maker reports it, Beam scores 80.1% and DeepSeek V4 Pro 87.9. these are each maker's own numbers, not a controlled head-to-head; the spec table lists every benchmark both publish.
which is cheaper, Beam or DeepSeek V4 Pro?
Beam lists lower input and output prices. on a coding-agent session (2M tokens in, 60,000 out) Beam costs $0.27 and DeepSeek V4 Pro $2.88.
can i switch between Beam and DeepSeek V4 Pro mid-task?
yes. in an octomind session, /model octohub:beam and /model octohub:deepseek-v4-pro switch models without losing the conversation, so you can draft on one and review on the other.
try either model
switch between models inside the same octomind session.