Beam 1.0 501B-A23B

Beam 1.0 501B-A23B

Quick answer: Beam 1.0 is Reflection AI's first open-weight model, a 501B-total/23B-active-parameter sparse MoE with a 1M-token context window, released under Apache 2.0. It reaches 81.6% on SWE-bench Verified and 97.7% on AIME 2026 (BF16 numerics), with FP8 and NVFP4 variants released alongside.

At a Glance

Where Beam 1.0 leads

• Token-efficient reasoning: Reflection claims comparable solution quality to leading open models using ~3x fewer inference-time reasoning tokens
• Strong agentic coding: 81.6% SWE-bench Verified, 80.1% Terminal Bench v2.1 (self-reported), 76.5% SWE-bench Pro v2-Hard
• Full 1M-token context window with Apache 2.0 licensing and day-one BF16/FP8/NVFP4 releases

Where it lags

• Text-only modality — no vision/multimodal support
• Technical report not yet published ("coming") at release time, limiting independent methodology verification

Best for: Teams wanting an open-weight, Apache-licensed alternative to closed frontier models for long-context agentic coding and tool-use workloads.

What Beam 1.0 Is

Beam 1.0 is the debut open-weight release from Reflection AI, a lab previously known for closed reasoning research. It uses a sparse mixture-of-experts architecture (501.3B total / 23.4B active parameters, 256 experts, 52 layers: 3 dense + 49 MoE) with gated GQA and sliding-window attention, targeting a 1M-token context window.

The model is positioned around three capability clusters: software engineering (resolving real GitHub issues, CLI/DevOps workflows, agentic coding harnesses), agentic tool use (long-horizon task completion, MCP-based tool integration, browsing), and reasoning (graduate-level science, theorem proving, applied math). Reflection emphasizes "token-efficient reasoning" — reaching comparable quality to other open models while using meaningfully fewer reasoning tokens per response.

Released simultaneously in BF16, FP8, and NVFP4 numerics, Beam 1.0 targets builders who want open weights without giving up frontier-class agentic coding performance.

Specifications

Pricing

Open weights: free to download and self-host (Apache 2.0); cost is hosting/inference infrastructure only. Hugging Face model card.

Context Window

Beam 1.0 has a 1,000,000-token context window — roughly 1,500+ pages of text in a single request. This supports whole-codebase analysis, long multi-document agentic workflows, and extended tool-use sessions without aggressive context truncation.

Public Benchmark Scores

Scores above are reported by Reflection AI and shown for context. They are not Benchgen measurements. Figures use BF16 numerics, temp=0.7/top_p=0.99 — see the model card for methodology. The model card's "HLE (with tools)" (51.1%) and "LiveCodeBench" (88.3%) figures, and the "Terminal Bench v2.1" / "AA-LCR" / "AA Omniscience" / "BrowseComp w/ context management" / "DeepSearchQA w/ context management" / "IFBench" / "LongBench v2" rows, are excluded here — see footnote.

Excluded benchmark rows (bucket C, documented per data-ops policy):

• HLE (with tools) — excluded: our humanitys-last-exam entity's existing sibling entries for comparable models use a no-tools condition; using the tools-enabled figure here would create a methodology mismatch, so the no-tools 36.0% figure was used instead and the with-tools number omitted.
• Terminal Bench v2.1 — excluded: no confirmed local match. Our repo carries versioned successors (Terminal-Bench 3.0, 4.0, Hard, Science 0.1) but none is confirmed equivalent to "v2.1" as the vendor names it — mapping to any of them would risk a silent methodology/version mismatch.
• AutomationBench - public split only — mapped cautiously to automationbench (Zapier) since it's the only local entity with that exact name, but note the vendor's score is a "public split only" subset, which may not be fully comparable to full-dataset scores from other models on our leaderboard; flagged, not excluded, since no alternative entity exists.
• AA-LCR, AA Omniscience (index), BrowseComp w/ context management, DeepSearchQA w/ context management, IFBench, LongBench v2 — excluded: Artificial Analysis (AA)-branded and context-management-qualified variants weren't independently confirmed to match our local generic entities' methodology (e.g. local browsecomp/deepsearchqa don't specify a "w/ context management" condition) within this run's time budget; carried forward as an open item for a future pass.
• LiveCodeBench — excluded this pass pending confirmation of which local livecodebench* variant/version the vendor's unqualified "LiveCodeBench" score corresponds to.

Beam 1.0 vs Alternatives

Beam 1.0 trades closed-API convenience for open weights and a 1M-token window at a competitive agentic-coding score — a relevant option for teams that need to self-host or fine-tune rather than call a hosted API.

How Beam 1.0 Performs on Real Agent Tasks

Beam's self-reported scores on SWE-bench Verified, MCP Atlas, and DeepSWE suggest strong tool-use and long-horizon coding capability, and its emphasis on token-efficient reasoning is a meaningful signal for cost-sensitive agentic deployments where inference-token count compounds across long tool-use chains. However, a public leaderboard score establishes general capability on a fixed task distribution — it does not establish reliable repeated completion on a specific domain's workflows. Builders evaluating Beam 1.0 for a particular agentic use case should validate against their own harness and task distribution rather than extrapolating directly from these vendor-reported numbers, especially since Reflection's full technical report (methodology detail) was marked "coming" at release.

Use Beam 1.0 via API

# Open-weight model — serve via vLLM, SGLang, or similar.
from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")

response = client.chat.completions.create(
    model="reflection/Beam-1.0-501B-A23B",
    messages=[{"role": "user", "content": "Summarize this contract clause..."}],
)
print(response.choices[0].message.content)

Related Models & Benchmarks

• Compare: DeepSeek V4.1 Flash · Qwen3-VL 235B-A22B Instruct
• Benchmarks: SWE-bench Verified · GPQA Diamond · AIME 2026

Frequently Asked Questions

Specs and scores sourced from Reflection AI's official Hugging Face model card; third-party benchmark scores attributed inline. Last updated 2026-10-20.