Reflection AI releases Apache 2.0-licensed weights for Beam, a 501B-parameter model, trained with RL that penalized unnecessary tokens
VentureBeatMatt Marshall
Context & Ripple Effects
Reflection introduced Beam on October 5–6, claiming reasoning performance near GLM 5.2 with three to four times less compute and agentic-task performance approaching Qwen3.8-Max. Publishing the weights turns those claims from a model announcement into an artifact developers can independently run and assess; the company is also releasing the training approach behind its token-efficiency objective.
Developers and enterprise teams can download, modify, and commercially deploy Beam under Apache 2.0 rather than depending on Reflection AI to operate it as a service.
Beam adopters can evaluate whether its reinforcement-learning penalty for unnecessary tokens improves the quality-versus-token trade-off in their own reasoning and agent workflows.
Second-order effects
Teams choosing between Beam and Arcee’s commercially usable Trinity-Large-Thinking model gain another large open-weight option, increasing the importance of independently reproducible performance and inference-cost tests over vendor claims.
Inference-platform and tooling providers have a reason to support Beam where customers want self-hosted deployments, while model buyers gain more leverage to separate a model’s weights from its serving stack.
Third-order effects
If more frontier-scale models are released under permissive terms, model vendors will face greater pressure to differentiate through training methods, runtime tooling, and support rather than access to weights alone.
Token-efficient training objectives could make output-token use a more explicit competitive metric alongside benchmark scores, especially for agentic workloads that generate long traces.
The trend: Open-weight AI competition is moving toward commercially reusable model artifacts whose practical value depends on independent deployment, serving tooling, and inference efficiency.
I think Reflection AI's Beam is the first open model to adapt Harmony format ands acknowledge it (Chat says muse use something similar to Harmony but they call it ATEM)
@ragzoi
Raghu on x
one number runs the other way. reflection's model card has beam at 39.3% on automationbench (public split only). artificial analysis got 48.5% on its version. probably different task sets, but you don't often see the outside eval come in higher
@defaiscope
@defaiscope on x
Reflection is measuring Beam against US open models. That's an easier bar than open models overall. Its launch post described the goal as advancing the “Western open-weight frontier.
@jimmy_longbow_
JimiLonbo on x
Reflection's first model is a proper open-weights drop: Beam, 501B-A23B MoE, Apache 2.0, weights in BF16, FP8 and NVFP4, plus the harness they trained it on. vLLM, SGLang and Ollama all shipped day-0 support. Ollama Cloud has it free for now with zero data retention, so you can w…
@stephzhan
Stephanie Zhan on x
Beam is yours to build on. Open weights, Apache 2.0. Alongside Mirror CLI, the harness Reflection built and trained it on. Proud of this team. Let's build! 🇺🇸
@eliebakouch
Elie on x
reflection beam oss architecture details IT'S SLIDING WINDOW AGAIN, same window size and pattern as inkling (window 512 and 5:1, meta is 2k and 3:1). they use a similar global attention scaling as llama4 ( https://arxiv.org/...) which still makes it position dependent even if the…
@chetantekur
Chetan Tekur on x
1/ Working on @reflection_ai Beam has been incredibly satisfying. It feels like I've been working on this for years. It's incredible what a small, motivated team can do when everyone is pulling in the same direction. Beam 1.0 is out today. Here's what we're releasing 👇
@mikelecheve
Mikel on x
Reflection just open-sourced a 501B model under Apache 2.0. Not “open-ish
@hrantkostanyan
Dr Hrant Kostanyan on x
Today we released Beam, Reflection's first open-weight model, under Apache 2.0. It's a 501B-parameter Mixture-of-Experts model with 23B active parameters, built for coding, reasoning, and long-horizon agentic work.
@iamgrigorev
George Grigorev on x
Reflection Beam checkpoint is finally here and hence details of the architecture Several important observations - 48 q heads, 8 kv heads, SWA-512, swa theta 10k (conceptually same as laguna :D ) - pretty high aspect ratio 98.5 (not enough depth, harms reasoning) - 1:5 SWA:GA rati…
@chahatusharma
Chahat Sharma on x
Reflection AI: Beam's open weights are on Hugging Face under Apache 2.0. It launches with over 40 partners and integrations, which Reflection calls an open operating system for agents. Would you self-host a 501 billion parameter model?
@1littlecoder
@1littlecoder on x
Super Fast ⚡️ @reflection_ai Beam is right now Live on @nebiustf 🔥 Beam is Reflection's open-weight, instruction-tuned text model for coding, reasoning and agentic tool use. 🔥 Its mixture-of-experts architecture has 501.3B total parameters and 23.4B active parameters. 🔥 It suppor…
@bradshannon
@bradshannon on x
I got early beta access to @reflection_ai's Beam model and it's SUPER fast. I gave it a couple simple questions and gave it the inevitable question of ‘Why is the sky blue?’ It gave the correct answer, Rayleigh scattering, so I asked it to create an HTML page to explain it. It ga…
@brandondamos
Brandon Amos on x
We've released the weights for Beam!! 🤗 https://huggingface.co/... 📄 Model card: https://reflection.ai/... 🧪 Try Beam in our playground & get an API key: https://platform.reflection.ai/ ... 📚 AND we take open model safety seriously and have posted our safety report: https://refle…
@alvarobartt
Alvaro Bartolome on x
Reflection just released Beam 1.0 501B open-weight model on the @huggingface Hub 🔥 - 501B MoE with A23B - 1M context - Text in, text out - Built for coding, reasoning, and agentic work - ~3x fewer reasoning tokens than leading open models - BF16, FP8, and NVFP4 - Apache 2.0
@fgbelidji
Florent Gbelidji on x
Crazy week for open-source AI, right? Now it's our turn! Beam, Reflection's first open model, is out: -501B MoE, 23B active -1M context -Built for coding, reasoning & agentic workloads -Apache 2.0 🧵
@reflection_ai
@reflection_ai on x
Beam is now available across platforms, with open weights under the Apache 2.0 license - in collaboration with an ecosystem of 40+ partners. We're releasing: - Weights, tokenizer, and reference implementation - BF16, FP8 and NVFP4 weights and a DFlash2 drafter - The open source h…
@sugarmissa1
Missa on x
@Techmeme @mmarshall Hot take: Apache 2.0 matters more than 501B. A model you can ship inside a client product with no licensing call beats a benchmark lead you can only rent. “Open” passes one test: can a small team run it and sell the output? The weights are free. The GPUs to h…
@metacurity.com
Cynthia Brumfield on bluesky
Foreign governments would rather opt into Chinese open-weight models than buy from US tech companies they have been trying to escape. — Western open-weight models still can't beat China — restofworld.org/2026/china-o...
@artificialanlys
@artificialanlys on x
Beam uses ~26k output tokens per Intelligence Index task: ~19k for reasoning and ~7k for answers. Only 26% of its output goes to final answers, the 28th-lowest share of 219 models tested on Intelligence Index v4.3
@_ghorbani
Behrooz Ghorbani on x
Really proud of Beam, our first-ever release! Among US open-weight models tested by Artificial Analysis, it ranks #2 on the Intelligence Index and #1 on the Cyber Index. So grateful to the team for the work that went into this. We're just getting started 🚀
@artificialanlys
@artificialanlys on x
Reflection's first model, Beam, scores 30 on the Artificial Analysis Intelligence Index, placing it #2 among US open weights models Beam is an open weights model with 501B total and 23B active parameters, released on @huggingface Congratulations to @reflection_ai on the release! …
@artificialanlys
@artificialanlys on x
Beam scores 34% of on the Artificial Analysis Cyber Index, making it the leading US open weights model that we've tested Driven by its token efficiency, it also sits on the Cyber Index vs. Output Tokens per task frontier, alongside GPT-6 Sol (Daybreak Blue)