stealthmodels.

MODEL BENCHMARKS

Beam 1.0 501B A23B: Pelican

Reflection’s open-weight reasoning model activates 23.4B of its 501.3B parameters per token, with a million-token context for coding and long-running agents.

A good fit for

Software engineering, multi-step tool workflows and mathematical or scientific reasoning across large codebases and documents.

Context window
1,048,576tokens[6] [7]
Active parameters
23.4Bparameters[1]
License
Apache 2.0[1]
Model factsArchitecture, context & capabilities
First available
2026-10-16[2]
Developer
Reflection AI[2] [1]
Origin
United States[3]
Family
Beam 1.0[1]
Context window
1,048,576 tokens[6] [7]
Maximum output
131,072 tokens[6] [7]
Architecture
Sparse mixture-of-experts transformer; 52 layers (3 dense, 49 MoE), 256 experts with 8 selected per token; gated GQA and sliding-window attention[1] [4]
Total parameters
501.27B[5]
Active parameters per token
23.4B[1]
Input
text[1]
Output
text[1]
Normal speed
972.09 tokens/sec avgMin 883.62 · Max 1,148.11 tokens/secOur measurement
API access & pricingOpenCode · $0 input / $0 output · per 1M tokens

OpenCode Beam Free: $0 input / $0 output for a limited time. OpenRouter / Nebius: $0.25 input / $1 output per 1M tokens. Reflection advertises free-to-start public beta access; no exact first-party rate verified.

[9] [10] [2]
  • OpenCodeObserved 2026-10-16
    Input $0Output $0Cached read $0

    per 1M tokens

    limited-time promotion

    Limited-time free offer; account access and regional availability follow OpenCode’s terms.

    beam-free[9]
  • OpenRouter / NebiusObserved 2026-10-16
    Input $0.25Output $1

    per 1M tokens

    Available

    Listed Nebius FP8 endpoint; input and output rates per million tokens.

    reflection/beam-1.0-501b-a23b[10] [7]
Run it locallyWeights, memory & deployment

large multi-GPU server or substantial host-memory offload

[1] [14]
  • Official BF16 checkpoint1,002.66 GB storage · BF16 with F32 tensors[15]
  • Official FP8 checkpoint503.28 GB storage · FP8[16]

Weight payload size excludes KV cache, runtime overhead and additional serving components; it is not a minimum VRAM requirement.

[15] [16]
Measurements & sources16 primary references

StealthMark score

StealthMark measures fluid and visual intelligence of AI models, resisting any pre-existing benchmark focused training.

General and Visual

General Intelligence

The ability to reason, solve problems and understand how things relate and behave.

Visual Intelligence

The ability to imagine, create and accurately reconstruct complex images in code form.

Judging

Scores reflect how well each result meets the benchmark requirements.

Overall score

Overall combines weighted task scores on a scale from 0 to 100. Estimated means the total includes an estimate for one missing scene.

Scoring categories

Task & likeness

Subjects, actions and likeness.

Anatomy & construction

Coherent bodies, joints and machinery.

Space & placement

Perspective, scale and contact.

Light & reflections

Lighting, shadows and reflections.

Purposeful detail

Clear, purposeful small features.

Artistry & character

Composition, expression and drawing skill, including complexity and its execution.

Materials & effects

Surfaces, texture and movement.

Visual integrity

Clean shapes, layers and edges.

BENCHMARK RESULTS

Compare results

This link opens the same results in the same order.

StealthMark benchmark results

Reset benchmark preferences?

Return to the default view, clear comparison selections and restore the advertisement.