Reflection AI released Beam's weights on 16 October 2026 with a 1M-token context window and a 128K output cap, after an API debut on 5 October. It suits teams that want a self-hostable coding and agent model, and it replaces a closed endpoint only where you can run 501B parameters or rent a partner host.
What changed
Beam is an open-weight sparse mixture-of-experts model from Reflection AI, built for coding and agent work, with 501B total and 23B active parameters. Reflection's own model card reports 81.6% on SWE-bench Verified and 90.6% on GPQA Diamond, and the weights ship under the Apache 2.0 licence.
Where it sits
- $1.50/M$ per 1M tokensBlended price (3:1)Lower is better#35 / 83peer median $1.71/Mvendor price, checked by HokAI
- 81.6%% solvedSWE-bench VerifiedHigher is better#10 / 34peer median 77.6%per source, see benchmark scores
- 90.6%% correctGPQA DiamondHigher is better#18 / 54peer median 87.6%per source, see benchmark scores
Priced around the middle of the 83 GA models with a published price (rank 35), and in the top third on SWE-bench Verified (rank 10 of 34).
Ranks are against GA models on HokAI that publish the same figure; ties share a rank.
Provider: Reflection AI · Family: Beam
More about Reflection AI on HokAI
Context window: 1,048,576 tokens · Max output: 131,072
Input modalities: text, tool-calls · Output: text, tool-calls
About Beam
Beam is the first open-weight model from Reflection AI, a US lab that also publishes the Mirror CLI coding harness. The API model (Beam-1.0-501B-A23B) was created on 5 October 2026, and the weights followed on Hugging Face on 16 October 2026 under Apache 2.0. Reflection describes it as built for coding, reasoning and long-horizon agent work, with a stated emphasis on spending fewer reasoning tokens than other open models.
The architecture is a sparse mixture-of-experts transformer with 501.3B total and 23.4B active parameters, 256 experts, and 52 layers (3 dense, 49 MoE). Attention is gated grouped-query attention with a sliding window, the vocabulary is 152K tokens, and the knowledge cutoff is 30 June 2026. Input and output are text only. The release ships in BF16, FP8 and NVFP4, plus a speculative-decoding drafter.
The context window is 1M tokens (1,048,576 in the API listing) and the maximum output is 128K tokens (131,072). In the API docs example, reasoning is mandatory and defaults to the highest effort level, with five levels from low to max. Tool calling and structured outputs are supported, and the API follows the OpenAI chat completions format, so existing OpenAI SDK code points at it by changing the base URL.
On Reflection's own model card, the BF16 build scores 81.6% on SWE-bench Verified, 66.1% on SWE-bench Pro v1, 80.1% on Terminal Bench v2.1, 90.6% on GPQA Diamond, 97.7% on AIME 2026 and 36.0% on Humanity's Last Exam without tools (51.1% with tools). It also reports 80.8% on MCP Atlas and 76.8% on BrowseComp with context management. These are vendor-run figures at temperature 0.7 and top_p 0.99, and the FP8 and NVFP4 builds land within roughly one to two points of BF16 on most rows. Reflection says Beam reaches comparable quality to leading open models with about three times fewer reasoning tokens; we have not seen an independent test of that ratio, so treat it as a claim. For open models in the same size class, compare GLM-5.2 and GLM-5.3, and for another open coding-focused option see Kimi K3 and DeepSeek V4 Pro.
Access is broad on day one. You can use Reflection's first-party API (free to start during the public beta), download the weights, or call a partner endpoint: the launch post lists AWS, Google Cloud, Microsoft Foundry, Cloudflare, Databricks, Fireworks AI, Together AI, Baseten, Modal, Nebius, Oracle, Vercel, OpenRouter, Ollama and LM Studio among others. The launch post names vLLM, SGLang, Hugging Face Transformers, Unsloth, Axolotl and NVIDIA Dynamo as day-zero serving and tuning paths. Third-party per-token prices vary by host; the OpenRouter listing on 19 October 2026 shows $1.00 input, $3.00 output and $0.10 cached input per 1M tokens.
On safety, Reflection published a separate safety technical report and ran its Iris evaluations: the model card reports an 88.9% refusal or pivot rate on harmful requests and an 81.5% compliance rate on harmless ones for BF16. The usage section of the card explicitly allows distillation, fine-tuning, quantization and good-faith adversarial testing. Pre-release testing partners named in the launch post include Gray Swan, Irregular, SecureBio and Perimeter Bio.
Beam fits teams that want a coding or agent model they can run on their own hardware or rent from several hosts, and that value an Apache 2.0 licence with no usage carve-outs on the weights. It is a poor fit for anyone who needs image, audio or video input, who cannot spare multi-GPU capacity for a 501B-parameter model, or who needs a vendor SLA and published data-retention terms, since we found none for the first-party API. Closed frontier models still lead on some agentic suites, so compare on your own repository before switching.
Pricing
Per-token prices are the OpenRouter listing read on 19 October 2026 (1.00 input, 3.00 output, 0.10 cached input per 1M tokens), not a Reflection rate card; other hosts price differently. Reflection's own API is in public beta and free to start. The weights are Apache-licensed, so self-hosting costs only your compute.
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.030 | $0.0030 | $0.033 |
| Support reply | $0.0020 | $0.0009 | $0.0029 |
| One coding agent run | $0.200 | $0.060 | $0.260 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Key Features
- Sparse MoE with 23.4B active: 501.3B total parameters across 256 experts, with 23.4B active per token.
- 1M-token context: Context window of 1,048,576 tokens and up to 131,072 output tokens.
- Three numerics plus drafter: Weights in BF16, FP8 and NVFP4, with a speculative-decoding drafter model.
- Tool calling and structured output: Supports function calling, JSON schema output and MCP-style tool use through an OpenAI-compatible API.
- Mirror CLI agent: Open-source terminal agent that Beam was trained alongside; also tested with OpenCode, Pi and Hermes.
Pros
- Permissively licensed weights with no usage carve-outs listed on the model card.
- Strong vendor-reported agentic coding scores for an open model, from a small active-parameter count.
- Available on more than 20 platforms at launch, so you can avoid lock-in to one host.
Cons
- Text-only; no vision or audio.
- Self-hosting needs multi-GPU hardware.
- No published first-party price list, SLA or data-retention terms found.
Benchmarks
- IFBench: 79.6% vendor-reported · 19 Oct 2026 — How precisely the model follows detailed instructions, % passing.
- AIME 2026: 97.7% vendor-reported · 19 Oct 2026 — Competition-level maths problems from the 2026 exam, % solved.
- MCP Atlas: 80.8% vendor-reported · 19 Oct 2026 — Tasks solved by calling external tools over the Model Context Protocol, % solved.
- GPQA Diamond: 90.6% vendor-reported · 19 Oct 2026 — PhD-level science questions that are hard to search for, % correct.
- SWE-bench Pro: 66.1% vendor-reported · 19 Oct 2026 — Harder, longer real-repository coding tasks, % solved.
- SWE-bench Verified: 81.6% vendor-reported · 19 Oct 2026 — Real GitHub issues fixed end to end, % solved.
- Terminal-Bench 2.1: 80.1% vendor-reported · 19 Oct 2026 — Multi-step tasks completed in a real command line, % solved.
- Humanity's Last Exam: 36% vendor-reported · 19 Oct 2026 — Expert-written questions across many fields, % correct.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Frequently Asked Questions
What does it cost to use Beam?
Reflection says its first-party API is free to start during the public beta and has not published a rate card on the pages we read. A third-party host, OpenRouter, charges $1.00 per 1M input tokens, $3.00 per 1M output tokens and $0.10 per 1M cached input at last read. Because the weights are Apache 2.0, self-hosting costs only your own GPU time.
Where does Beam do best on Reflection's own tests?
The model card puts the BF16 build at 80.1% on Terminal Bench v2.1, 66.1% on SWE-bench Pro v1 and 80.8% on MCP Atlas, all vendor-run at temperature 0.7. FP8 and NVFP4 builds land within about one to two points on most rows, for example 65.9% and 65.6% on SWE-bench Pro v1. No independent replication was found yet, so check these against your own repository.
Is Beam open source or only open weights?
The weights are released on Hugging Face under Apache 2.0 in BF16, FP8 and NVFP4, with a drafter model. Training data and full training code are not released, so open-weight is the accurate label even though the licence is permissive. The model card allows distillation, fine-tuning and quantization.
Does Beam train on my data?
The model card and developer docs we read do not state a retention or training policy for the first-party API. If that matters, self-host the weights or read the terms on platform.reflection.ai before sending sensitive data. Partner hosts set their own policies.
Who should pick Beam, and who should skip it?
Pick it if you run coding agents and want an open model you can self-host or rent from many hosts. Skip it if you need image or audio input, or lack GPUs and need a contractual SLA; GLM-5.2 or a closed frontier model may suit that better.