Reflection AI Beam: 501B Apache 2.0 Open Model, Specs, Benchmarks

Reflection AI, the Nvidia-backed startup, released the weights for Beam 1.0 on October 16, 2026, an open-weight model with 501 billion total parameters and 23 billion active per token under the Apache 2.0 license [1][2]. Beam is text only and has a 1 million token context window. Reflection says it matches GLM-5.2 on advanced reasoning with 3 to 4 times less inference compute, and Artificial Analysis scores it 30 on its Intelligence Index, second among US open-weight models [2][3]. The hosted API has no published price, and the efficiency claim is Reflection's own estimate.

What is Reflection AI Beam and what is new

Beam is the first open-weight model from Reflection AI, a sparse mixture-of-experts system built for coding and agent work. Reflection announced it on October 5, 2026, and released the weights and model card on October 16 under Apache 2.0 [1][2]. It describes Beam as a workhorse in the class of fast mid-sized flash models, and says it trained from new weights without using other models' outputs [4].

The pitch is token efficiency. Reflection says Beam scores comparably to GLM-5.2 on advanced reasoning tests while using 3 to 4 times less inference compute, with a larger gap against models above 2 trillion parameters such as Qwen 3.8 Max [2]. The estimate multiplies active parameters by mean generated tokens, so it leaves out prompt processing, attention and serving overhead. Reflection counts GLM-5.2 at 40 billion active parameters against 23 billion for Beam [4].

"Beam advances the Western open-weight frontier."

Reflection AI, "Introducing Beam," October 5, 2026 [2]

Nvidia (NASDAQ: NVDA) is part of the story. Beam trained on 10,500 Nvidia GB300 GPUs, and the Financial Times reported on October 10 that Nvidia was in early talks to buy Reflection or deepen its stake, as Depth Grid covered in Nvidia in Talks to Deepen Stake in or Buy Reflection AI. Chief technology officer Ioannis Antonoglou called Nvidia a "close partner and investor" and dismissed the acquisition report as "baseless rumors" without confirming or denying talks [4].

Under the hood: architecture, training and hardware

Beam has 501.3 billion total parameters and 23.4 billion active per token across 52 layers, 3 dense and 49 mixture-of-experts, with 256 experts. It uses gated grouped-query attention with sliding-window attention and a 152,000 token vocabulary [1]. It was pretrained on 23.8 trillion tokens of web and licensed data on 6,144 GB300 GPUs in under four weeks, then extended to 1 million tokens of context in midtraining, though reinforcement learning ran at a maximum of 256,000 tokens [2].

Reinforcement learning ran for four weeks and produced more than 100 million rollouts, with a penalty for unnecessary tokens [2][4]. The model is text only, was developed predominantly for English and has a knowledge cutoff of June 30, 2026 [1]. Weights come in BF16, FP8 and NVFP4, at about 469 GiB for FP8 and about 273 GiB for NVFP4, with a 1.93 billion parameter drafter for speculative decoding [5].

Hardware guidance varies by source. The model card gives none. VentureBeat reports that the 8-bit and 4-bit builds fit on one eight-GPU server using H100, H200 or GB300 chips, while OrcaRouter lists a reference node of four B300 GPUs and day-zero support for vLLM, SGLang, Transformers, Unsloth and NVIDIA Dynamo [1][4][5]. Reflection also ships Mirror, an open-source coding agent in Rust and Python, and an Apache 2.0 operating layer with Kubernetes containers for deployment [4][8].

Benchmarks: Reflection's table against an independent index

Reflection's own table, updated October 8, puts Beam level with GLM-5.2 on coding and behind newer Chinese open models. These scores are company-published, and harness choices shift the same benchmark by several points [2][6].

The final model card lists slightly different BF16 figures, including 43.3% on DeepSWE v1.1, 66.1% on SWE-Bench Pro v1 and 81.6% on SWE-bench Verified, and the FP8 and NVFP4 builds land within roughly two points on most tests [1]. Reflection's table puts Kimi K3 at 88.3 on Terminal-Bench v2.1, while Meta's model card lists it at 85.0 on the same test [2][6]. That gap shows why self-reported tables need an outside check.

Artificial Analysis gives Beam an Intelligence Index score of 30. That ranks second among US open-weight models, behind Meta's Muse Spark Open at 39 and ahead of Nvidia's Nemotron 3.5 Super VL at 26, and tenth among all open-weight models, one point behind the UAE's K2 Horizon at 31. GPT-6 Astra at maximum effort scores 53 [3]. Beam uses about 26,000 output tokens per index task, roughly 19,000 of reasoning and 7,000 of answer, against about 52,000 for K2 Horizon [3].

Beam scores 48.5% on AutomationBench-AA against 37.2% for K2 Horizon, and 13.1% on CritPt against 4.6%. On AA-Omniscience it answers 33% correctly against 18% for K2 Horizon, but its hallucination rate is 47% against 26%. It completes 34% of tasks on the Cyber Index, the best result among US open-weight models Artificial Analysis tested [3]. Its weak spots are 3.0% on Terminal-Bench 4.0, a different version from the v2.1 Reflection cites, and 976 Elo on AA-Briefcase. In Mirror, its Coding Agent Index score is 28, against 43 for Claude Code with Qwen3.8 Max and 68 for the leaders [3].

Pricing, availability and how to access it

Beam's weights are free to download from Hugging Face under Apache 2.0, and Reflection places no restriction on distillation, fine-tuning, quantization or merging [1][5]. Community builds for llama.cpp, LM Studio, Jan and Ollama are listed on the model card, and Reflection's open-source code is on GitHub [1][8].

Reflection has not published a hosted API price. Its platform at platform.reflection.ai is free during the public beta, and Baseten began serving Beam through OpenRouter on October 16 with no price listed at publication [4]. OrcaRouter reports a served context of 1,048,576 tokens, a 131,072 token output cap and mandatory reasoning with five effort levels from low to max, defaulting to max. Reasoning tokens count toward rate limits, so it advises setting the effort level explicitly [5]. The developer documentation covers the API [7].

Reflection lists 33 named partners and about nine more pending approval. Dell and HPE will install Beam on customer hardware with Reflection engineers on site, and Shinsegae Group in South Korea is building an AI factory with it. Oura, which tested Beam through Modal, said it showed fewer context leaks and was more grounded than other frontier models it evaluated [4].

Limitations, trade-offs and alternatives

Fewer tokens do not guarantee a lower bill. Reflection told VentureBeat it has nothing concrete to share on cost per finished task, its compute estimate leaves out prompt processing, which dominates long agent and coding prompts, and the default maximum reasoning effort makes the out-of-box setup the most expensive one [4][5]. VentureBeat is running its own cost-per-task tests comparing Beam with Muse Spark Open, DeepSeek V4.1 Flash and GLM-5.2, and results are pending [4].

On safety, the model card reports that BF16 Beam refused or redirected 88.9% of harmful requests and complied with 81.5% of harmless ones [1]. Reflection published a Beam Safety Report, and Gray Swan, Irregular and SecureBio tested the model before release [4][9]. VentureBeat also reports that Beam used outside AI models and OCR services to read images when given web access, which Reflection calls acceptable problem-solving. Antonoglou says Beam is not meant to replace frontier models "on everything" [4].

Alternatives are plentiful. Muse Spark Open from Meta scores 39 on the same index under a custom license, and its sibling model is covered in Meta Muse Wonder Launch: 1M Context, $4 API Price, Benchmarks [3]. DeepSeek V4.1 Flash posts 90.6 on Terminal-Bench v2.1 in Reflection's own table [2]. Teams that want another open model can wait for Mistral, whose Large 4 weights are due by the end of October and are detailed in Mistral Large 4 Preview: 1.05T Open-Weight MoE, Specs and Pricing.

Sources

1. Reflection AI. "Beam-1.0-501B-A23B," model card, Hugging Face. huggingface.co
2. Reflection AI. "Introducing Beam," October 5, 2026, updated October 8. reflection.ai
3. Artificial Analysis. "Reflection has released Beam," October 16, 2026. artificialanalysis.ai
4. VentureBeat. "Reflection's Beam aims to undercut Chinese open models with fewer tokens, not cheaper tokens." venturebeat.com
5. OrcaRouter. "Beam 1.0 Open Weights Are Out and Scored 30 Independently." orcarouter.ai
6. Meta. "Muse-Spark-Open-816B-A42B," model card, Hugging Face. huggingface.co
7. Reflection AI. Developer documentation. developers.reflection.ai
8. Reflection AI. Open-source repositories, GitHub. github.com
9. Reflection AI. "Beam Safety Report." reflection.ai

Read More on Depth Grid

• Nvidia in Talks to Deepen Stake in or Buy Reflection AI
• Mistral Large 4 Preview: 1.05T Open-Weight MoE, Specs and Pricing
• Meta Muse Wonder Launch: 1M Context, $4 API Price, Benchmarks
• Claude Haiku 5.5 Launch: $0.10 Input, 1M Context, Benchmarks

Article by Mahesh | Depth Grid

;