Reflection AI's first open-weight model — a 501B-parameter sparse mixture-of-experts system with ~23B active and a 1M-token context, announced October 5, 2026 and built for coding, reasoning, and agentic workloads.
Last verified · 2026-10-05 · by Moe Ameen
Beam is a large language model from Reflection AI, announced on October 5, 2026 as the company's first open-weight release. It is a text-only sparse mixture-of-experts (MoE) model with 501 billion total parameters and about 23 billion active per token, paired with a context window Reflection lists at 1 million tokens. It is aimed squarely at technical and agentic work — software engineering, terminal operations, tool use, web search, and complex STEM reasoning — rather than at chat or creative content. Reflection frames the release around "frontier open intelligence" and positions Beam against the leading open-weight models, including this year's Chinese releases.
On training, Reflection says Beam was pretrained on 23.8 trillion tokens and then put through a large reinforcement-learning run — more than 100 million rollouts across roughly 10,500 NVIDIA GB300 GPUs over about four weeks. It reports strong figures on coding and reasoning benchmarks (including SWE-Bench Pro, Terminal-Bench, AIME 2026, and GPQA Diamond) and claims parity with Z.ai's GLM-5.2 on advanced reasoning while using three to four times less inference compute. Those are Reflection's own numbers and await independent verification.
Availability matters here. At announcement, the weights were not yet public — access was an early-access waitlist while the model finished red-teaming and evaluations. Reflection said it will release the weights under an Apache 2.0 license later in October 2026, along with a technical report, a model card, and the full stack to run, evaluate, and fine-tune the model. Confirm the spec, benchmarks, and release terms on Reflection's own pages before relying on them.
The honest framing for a creator: Beam is a text-in, text-out reasoning and coding engine, not a content studio. It produces text and code — no images, video, audio, captions, or publishing. Its strength is reasoning, long-context analysis, and agentic behavior like tool use and search, which makes it a capable research-and-automation front end. Turning that into finished, on-brand, published content is a separate job.
Beam's most distinctive edge for creators is not writing — it is acting. Its agentic abilities (tool use, terminal, web search) mean it can run an autonomous research pass over a topic, pull and synthesize source material, and hand back a structured brief or script; and once the Apache 2.0 weights ship, you can run that whole step on your own hardware, keeping proprietary source in-house. That is a powerful front end for a content pipeline — but it is only the front end. Beam ends at researched, correct text. It puts no face on the idea, builds no feed-ready format, and publishes nothing. That downstream half is exactly what [Kompozy](/) is built for, and the pairing is a clean division of labor: Beam's agent gathers and drafts, Kompozy produces and ships.
Wire Beam's research or draft into Kompozy as a source and the engine does what a reasoning model cannot. A [Persona Short](/glossary/persona-shorts) narrates the finding on camera with a face-locked HeyGen avatar; a brand-exact [Carousel](/glossary/hyperframes) walks the argument through step by step; Quote Graphics lift the one stat worth remembering; and a matching Blog Article and Email Newsletter go long for the audience that wants depth — every piece rewritten under a [Persona Brief](/glossary/persona-brief) so research-grade source lands in your voice, not the model's. Then [Autopilot](/glossary/autopilot) reframes each asset to 9:16, 1:1, and 16:9 and schedules and publishes across the eight social platforms plus blog and email behind a per-post review. Beam runs the research and writes the brain-work; Kompozy turns it into a week of on-brand, published content.
Beam is Reflection AI's first open-weight model, announced October 5, 2026. It is a text-only sparse mixture-of-experts model with 501 billion total parameters (about 23 billion active) and a 1M-token context, built for coding, reasoning, and agentic workloads like tool use, terminal operations, and web search. Reflection says it will release the weights under an Apache 2.0 license later in October 2026.
No. Beam is a text-in, text-out model — it outputs text and code only, with no images, video, or audio. To turn its scripts, briefs, and research into visual content you pair it with a generation-and-publishing engine like Kompozy, which makes persona/avatar video, carousels, images, quote graphics, blogs, and newsletters.
Reflection describes Beam as open-weight and says it will publish the weights under an Apache 2.0 license later in October 2026, with a technical report and fine-tuning tools. At announcement, access was an early-access waitlist and no API price had been published. Expect the practical cost to be GPU compute if you self-host. Confirm the license and current pricing on Reflection's pages.
It is tuned for coding, reasoning, and agentic work rather than English short-form voice, so it works best as an upstream research, drafting, and automation brain — especially for gathering or digesting source material. For on-brand captions, formatting, and distribution, hand that output to a content engine like Kompozy.
Research or draft in Beam, then bring the text into Kompozy as a source. Kompozy generates 18 formats from that one input — persona/avatar video, carousels, quote graphics, photo posts, a blog, and a newsletter — holds a consistent voice and face across them via the Persona Brief, and schedules and publishes across eight social platforms plus blog and email.