Ollama's Beam free period ends November 9: pricing, the 501B model behind it, and your options

If you have been running Beam on Ollama for free, the ride ends Sunday night, US Pacific time. Ollama emailed users on November 7: starting November 9, 2026 at 12:01 AM PST, Beam requires usage credits or a subscription, billed at $0.125 per million input tokens, $0.015 cached input, and $0.40 per million output tokens. One wording catch: the email calls the cutoff "Sunday, November 9", but November 9, 2026 is a Monday. The model page's timestamp (free until November 9, 2026, 12:01 AM PST) and the email's date agree, so the cutoff is Monday 12:01 AM PST, 10:01 AM in Israel; the email's weekday is simply off. Requests without credits will return payment-required errors. Beam is Reflection AI's first open-weight model, a 501B-total, 23B-active mixture-of-experts model aimed at coding and agentic work, and it has been free on Ollama Cloud since October 16. If your agents or scripts call beam:cloud, you have a decision to make this weekend: pay, self-host the open weights, or switch.

What changes on November 9

Per Ollama's email and the Beam model page:

• Free access ends at 12:01 AM PST on Monday, November 9, 2026 (10:01 AM Israel time). The model page carries the same timestamp. Ollama's email words it as "Sunday, November 9"; the weekday is wrong, the date and time are consistent across both.
• After that, Beam is priced like other Ollama Cloud models: $0.125 per million input tokens, $0.015 per million cached input tokens, $0.40 per million output tokens.
• Free-tier requests will fail with payment-required errors until you add credits on the Ollama dashboard. Ollama notes its subscription plans include a discount on usage credits.
• Nothing changes about the model itself: same beam:cloud tag, same 1M token context, same API shape through ollama.com/api/chat and the Ollama client libraries.

This is a hard cutover, not a grace period. If you have CI jobs, coding assistants, or cron agents pointed at beam:cloud, they will start erroring Monday morning US Pacific time, late morning in Israel.

What Beam actually is

Beam comes from Reflection AI, which introduced it on October 5, 2026 as its first open-weight model. The architecture, per Reflection's announcement and the Ollama model page:

• Sparse mixture-of-experts: 501.3B total parameters, 23.4B active per token, 256 experts, 52 layers (3 dense, 49 MoE).
• 1M token context window, text only, developed predominantly for English. Knowledge cutoff June 30, 2026.
• Gated GQA plus sliding window attention, modified Harmony chat template, released in BF16, FP8, and NVFP4 numerics.
• Trained on 23.8 trillion tokens, then a reinforcement learning run Reflection describes as over 100 million rollouts on 10,500 NVIDIA GB300 GPUs across four weeks.

Reflection positions Beam against larger open models like GLM 5.2 and Qwen 3.8-Max on coding and agentic tasks, and concedes frontier open models like Kimi K3 lead it on raw capability, arguing Beam's advantage is inference efficiency. The Ollama page claims Beam reaches comparable solution quality to leading open models with roughly 3x fewer inference-time reasoning tokens. Both are vendor claims.

On Ollama, Beam is cloud-only: ollama run beam:cloud, and ollama launch integrations for Claude Code, OpenCode, Hermes Agent, and OpenClaw. Because the weights are open, self-hosting is a real option, though a 501B MoE is not a laptop model; plan for serious GPU memory even in FP8 or NVFP4.

The benchmark picture, with warnings attached

The Ollama model page publishes Beam's eval table across BF16, FP8, and NVFP4 builds (all run at temperature 0.7, top_p 0.99). Selected BF16 numbers:

• SWE-Bench Verified 81.6, SWE-Bench Pro v2-Hard 76.5 (78.6 on the NVFP4 build), Terminal Bench v2.1 80.1, DeepSWE v1.1 43.3.
• AIME 2026 97.7, GPQA Diamond 90.6, HLE with tools 51.1.
• MCP Atlas 80.8, BrowseComp with context management 76.8, Tau3 Banking 36.1.

Two cautions. First, these are self-reported numbers from the model's own hosting page; Reflection ran the evals. Second, Reflection's October launch post shows slightly different figures for the same benchmarks (DeepSWE v1.1 44.4, SWE-Bench Verified 80.9), and its comparison table has Beam trailing models like DeepSeek V4.1 Flash and Kimi K3 on several coding evals. The honest summary: Beam looks competitive with the current open-weight pack on coding and agentic work, not clearly ahead of it, and its differentiator is meant to be token efficiency.

What I would check

These are my untested suggestions, not findings:

• Grep your repos, dotfiles, and cron jobs for beam this weekend, before the Monday cutoff. The failure mode is silent: payment-required errors in a pipeline you forgot used the free tier.
• If you keep using Beam through a coding harness, measure your actual token mix. The output rate ($0.40) is 3.2x the input rate, and reasoning models inflate output tokens, so effective cost depends heavily on how much the model thinks.
• Compare Beam's cost per completed task against whatever you used before October 16, not against the sticker price of frontier APIs.
• If you are considering self-hosting to avoid the bill, price the GPU hours first. For most solo builders, Ollama's rates will beat renting the hardware to run a 501B MoE.

What I would not conclude

• That Beam is free after November 9 with a plan. It is paid; plans discount credits, they do not make Beam free.
• That the benchmark table predicts your workload. Vendor evals at temperature 0.7 tell you little about a 30-minute agentic coding session with retries.
• That the "3x fewer reasoning tokens" claim has been independently verified. It comes from the model's own materials.
• That self-hosting is cheaper. Open weights make it possible, not economical.

Frequently asked questions

When exactly does Beam stop being free on Ollama?

Monday, November 9, 2026 at 12:01 AM PST (10:01 AM Israel time), per the Beam model page's timestamp and the date in Ollama's email. The email says "Sunday, November 9", but November 9 is a Monday; treat the timestamp as authoritative. After it, requests without credits or a subscription return payment-required errors.

What does Beam cost after the free period?

$0.125 per million input tokens, $0.015 per million cached input tokens, and $0.40 per million output tokens, matching Ollama's other cloud models. Ollama says subscription plans include a discount on usage credits.

What is Beam, technically?

Reflection AI's first open-weight model: a sparse MoE with 501.3B total and 23.4B active parameters, 256 experts, 1M token context, text-only, knowledge cutoff June 30, 2026. Trained on 23.8T tokens plus a large RL run Reflection describes at over 100 million rollouts on 10,500 GB300 GPUs.

Can I keep using Beam without paying Ollama?

The weights are open, so you can self-host or use another host. On Ollama itself, no: beam:cloud is a cloud model and will bill from November 9. Keep in mind a 501B-parameter MoE needs substantial GPU memory to self-host, even quantized.

How do I run Beam with a coding agent?

Per the Ollama model page: ollama launch claude --model beam:cloud for Claude Code, with the same pattern for OpenCode, Hermes Agent, and OpenClaw, or call ollama.com/api/chat directly with an API key.

How good is Beam for coding?

Vendor-reported numbers (Ollama's model page, BF16 build): SWE-Bench Verified 81.6, Terminal Bench v2.1 80.1, MCP Atlas 80.8. Reflection's own launch table shows it behind DeepSeek V4.1 Flash and Kimi K3 on several coding evals. Treat all of these as vendor evals and test on your own tasks.

Does this affect local Ollama models?

No. This is about Beam on Ollama Cloud. Locally pulled open models remain free to run as before.

Sources

• Beam model page | Ollama - read November 7, 2026. Pricing cutoff, per-token prices, architecture table, benchmark table, launch integrations, and API examples.
• Introducing Beam: Reflection's 501B open-weight model - read November 7, 2026. Training details (23.8T tokens, 100M+ RL rollouts on 10.5K GB300 GPUs), positioning against other open models, and the launch benchmark table.
• Ollama email to users, November 7, 2026 - the free-period end date, pricing, payment-required errors, and where to add credits. This is the newsletter item that prompted the article.
• Welcome Beam | Hugging Face blog - confirmed via search listing that the weights were released openly on October 16, 2026.

Not read: the Beam technical report and model card, Reflection's early-access terms, and Ollama's plan pricing details. All benchmark figures above are vendor-reported. I have not run Beam myself.