Reflection Beam: Is It Free, and How Do You Use It?

2026-10-17 · 5 min read

Reflection Beamopen weightsfree accessmodel-release

Reflection released Beam, its first open-weight model, on October 16, 2026. The weights are free to download under Apache 2.0, and Ollama Cloud offers a limited-time free tier — but running it yourself needs a multi-GPU node, and hosted APIs are paid. There is no single “free” answer: which one applies depends on whether you download it, use the free cloud tier, or pay a provider. aiwith has not integrated Beam. If you want to work in aiwith today, Kimi K3 and DeepSeek are available alternatives, not access to Reflection’s model.

Last checked: October 17, 2026. This is a source-checked access guide, not a hands-on benchmark.

What is Reflection Beam?

Beam is a text-only sparse mixture-of-experts model with 501 billion total parameters and 23 billion active per token. It has a 1M-token context window and is released under the Apache 2.0 license — one of the more permissive open-weight licenses, unlike the custom licenses some labs use.

The Hugging Face launch post describes 52 layers (3 dense plus 49 MoE), 256 experts per MoE layer with 8 routed per token, and three checkpoint formats: BF16, FP8 and NVFP4. Reflection lists the knowledge cutoff as June 30, 2026. Beam is built for coding, reasoning and long-horizon agentic tasks, and it ships with Mirror, Reflection’s own open-source CLI harness.

Reflection’s own reported numbers include 81.6% on SWE-bench Verified, 97.7% on AIME 2026 and 90.6% on GPQA Diamond. Those are the vendor’s evaluations. The independent Artificial Analysis review scores Beam at 30 on its Intelligence Index — second among US open-weight models, behind Meta’s Muse Spark Open (39) — and highlights its token efficiency while noting Terminal-Bench 4.0 as a weak point (3.0%).

Is Reflection Beam free?

There are three different questions behind “free”:

Are the weights free? Yes. Reflection released the Beam checkpoints on Hugging Face under Apache 2.0, so you can download, inspect, run and fine-tune them without a license fee. “Open weights” still means you supply the hardware to run them.
Is there a free hosted option? Ollama is a day-zero partner and lists Beam on Ollama Cloud “for free (for a limited time only)” with full zero data retention. Ollama has not published an end date for that offer, so treat it as temporary and check the Ollama blog post before relying on it. Reflection’s own API platform is in public beta and says it is “free to start” for testing.
Are the paid APIs free? No. OpenRouter and the other launch providers bill per token, with their own account and credit rules.

How do you use Reflection Beam?

The route depends on what you need:

Fastest free test — Ollama Cloud. If you already have Ollama, the cloud tier avoids downloading anything: ollama run beam:cloud. Ollama also lists one-line launchers such as ollama launch claude --model beam:cloud and ollama launch opencode --model beam:cloud. Access runs through Ollama Cloud, so this is a hosted option, not local inference.
Paid API. OpenRouter lists the model as reflection/beam-1.0-501b-a23b. Reflection’s first-party platform and the other launch providers (AWS, Azure, Google Cloud, Baseten, Fireworks AI, Modal, Nebius, Together AI, Vercel and more) have their own pricing and access rules.
Self-host the weights. Free to download, but see the hardware section below before committing.

What does OpenRouter charge?

The OpenRouter model page lists $1 per million input tokens, $3 per million output tokens and $0.10 per million cached input tokens, with a 1M-token context. We checked these against OpenRouter’s public model catalog; there was no free variant of reflection/beam-1.0-501b-a23b in the catalog at the time of this check. Prices and provider availability can change.

For example, 10,000 uncached input tokens plus 2,000 billed output tokens cost about $0.016 at those rates, before any account-level fees. That is an arithmetic example, not a prediction of how many tokens a task will use.

Can you run Reflection Beam on your own hardware?

Yes, but the headline number is not a single-GPU job. All 256 experts stay in memory, so even though each token computes through only ~23B parameters, deployment needs the memory of a 501B model.

The HF launch post’s guidance:

NVFP4 checkpoint (recommended): shrinks the 501B weights from ~1 TB in BF16 to about 280 GB, and runs natively on NVIDIA Blackwell.
Four B200s (720 GB) is the recommended setup; two B200s (360 GB) can hold the weights with little room left for activations and the KV cache. 2×B300 at 288 GB each, or 8×RTX Pro 6000, are also listed.
Pre-Blackwell GPUs such as the H200 cannot run NVFP4; use the FP8 checkpoints instead.

If you want to test Beam without that hardware, use a hosted option (Ollama Cloud, OpenRouter, HF serverless inference providers, or Reflection’s beta) instead of downloading the weights.

Is Reflection Beam available on aiwith?

No, as of this check. An aiwith subscription does not unlock Reflection Beam. We checked the configured model list and the alternative landing pages below.

If you mainly want help with a coding question or a long document in a browser, you can use:

Kimi K3 on aiwith for a long-context or reasoning task (1M-token context).
DeepSeek on aiwith for an everyday writing or coding task. Check the model selector for the exact version; the landing-page headline uses broader DeepSeek branding.

These are task alternatives, not a benchmark ranking or a claim that they match Beam’s capabilities. If your goal is specifically to evaluate Reflection’s new release, use a provider that explicitly lists beam-1.0-501b-a23b.