Language modelPreview
Reflection Beam
Reflection AI
- Released
- October 5, 2026
- Data date
- October 7, 2026
- Input
- Text
- Output
- Text
Reflection announced Beam as its first open-weight model. This text-only sparse MoE has 501 billion total parameters, with 23 billion active per token. Reflection targets coding and agentic tasks.
Beam has no mode that disables reasoning. Beta API access is through a waitlist. Weights, a model card, and a technical report are promised for later in October 2026; Apache 2.0 is the planned weights license, but it has not been published yet.
Specifications and access
| Specification | Value and source |
|---|---|
| API model ID | Beam-501B-A23BSource |
| Architecture | Sparse mixture of experts, 501 billion total parameters and 23 billion active per tokenSource |
| Modality | Text input, text outputSource |
| API context window | 262,144 tokens total for input and output. The limit may change during the beta.Source |
| Maximum output | 131,072 tokensSource |
| Knowledge cutoff | June 30, 2026Source |
| Reasoning | Always enabled. Adjustable from low to max; medium is the default.Source |
| Weights | Not publicly available yet. Reflection announced a release for later in October 2026.Source |
| Announced license | Apache 2.0 is planned for the announced weights. The license text has not been published yet.Source |
| Public model weights | Reflection’s verified Hugging Face organization does not yet list public models.Source |
Page 1 of 2
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
DeepSWE (Show measurement, test conditions, and source)
- Source value
- 44.4
- Score
- 44.4
- Metric
- reported score
- Unit
- No unit provided
- Benchmark version
- 1.1
- Category
- coding
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Reflection Beam launch
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Beam launch evaluation, excluding Inkling and Beam Base. DeepSWE v1.1.
- Context
- Reflection reports the Beam column. Effort, harness, run date, sample size and score units are not specified in the table. The raw score is retained without adding a percentage unit.
SWE Bench Pro v2-Hard (Show measurement, test conditions, and source)
- Source value
- 77.2
- Score
- 77.2
- Metric
- reported score
- Unit
- No unit provided
- Benchmark version
- 2-Hard
- Category
- coding
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Reflection Beam launch
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Beam launch evaluation, excluding Inkling and Beam Base. Hard version 2 subset, distinct from version 1.
- Context
- Reflection reports the Beam column. Effort, harness, run date, sample size and score units are not specified in the table. The raw score is retained without adding a percentage unit.
SWE Bench Pro (Show measurement, test conditions, and source)
- Source value
- 65.5
- Score
- 65.5
- Metric
- reported score
- Unit
- No unit provided
- Benchmark version
- 1
- Category
- coding
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Reflection Beam launch
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Beam launch evaluation, excluding Inkling and Beam Base. Version 1, distinct from the version 2 Hard subset.
- Context
- Reflection reports the Beam column. Effort, harness, run date, sample size and score units are not specified in the table. The raw score is retained without adding a percentage unit.
Terminal Bench (Show measurement, test conditions, and source)
- Source value
- 80.1
- Score
- 80.1
- Metric
- reported score
- Unit
- No unit provided
- Benchmark version
- 2.1
- Category
- coding
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Reflection Beam launch
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Beam launch evaluation, excluding Inkling and Beam Base. Terminal Bench v2.1.
- Context
- Reflection reports the Beam column. Effort, harness, run date, sample size and score units are not specified in the table. The raw score is retained without adding a percentage unit.
SWE Atlas Codebase QnA (Show measurement, test conditions, and source)
- Source value
- 34.6
- Score
- 34.6
- Metric
- reported score
- Unit
- No unit provided
- Category
- coding
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Reflection Beam launch
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Beam launch evaluation, excluding Inkling and Beam Base. Codebase question answering.
- Context
- Reflection reports the Beam column. Effort, harness, run date, sample size and score units are not specified in the table. The raw score is retained without adding a percentage unit.
SWEBench Multilingual (Show measurement, test conditions, and source)
- Source value
- 78
- Score
- 78
- Metric
- reported score
- Unit
- No unit provided
- Category
- coding
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Reflection Beam launch
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Beam launch evaluation, excluding Inkling and Beam Base. Multilingual software-engineering evaluation.
- Context
- Reflection reports the Beam column. Effort, harness, run date, sample size and score units are not specified in the table. The raw score is retained without adding a percentage unit.
SWEBench Verified (Show measurement, test conditions, and source)
- Source value
- 80.9
- Score
- 80.9
- Metric
- reported score
- Unit
- No unit provided
- Category
- coding
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Reflection Beam launch
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Beam launch evaluation, excluding Inkling and Beam Base. Verified software-engineering task set.
- Context
- Reflection reports the Beam column. Effort, harness, run date, sample size and score units are not specified in the table. The raw score is retained without adding a percentage unit.
AIME 2026 (Show measurement, test conditions, and source)
- Source value
- 97.8
- Score
- 97.8
- Metric
- reported score
- Unit
- No unit provided
- Benchmark version
- 2026
- Category
- reasoning
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Reflection Beam launch
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Beam launch evaluation, excluding Inkling and Beam Base. AIME 2026 mathematics.
- Context
- Reflection reports the Beam column. Effort, harness, run date, sample size and score units are not specified in the table. The raw score is retained without adding a percentage unit.
HLE (no tools) (Show measurement, test conditions, and source)
- Source value
- 36.2
- Score
- 36.2
- Metric
- reported score
- Unit
- No unit provided
- Category
- reasoning
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Reflection Beam launch
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Beam launch evaluation, excluding Inkling and Beam Base. No tools, explicitly stated by the source.
- Context
- Reflection reports the Beam column. Effort, harness, run date, sample size and score units are not specified in the table. The raw score is retained without adding a percentage unit.
SciCode (Show measurement, test conditions, and source)
- Source value
- 49.7
- Score
- 49.7
- Metric
- reported score
- Unit
- No unit provided
- Category
- coding
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Reflection Beam launch
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Beam launch evaluation, excluding Inkling and Beam Base. Scientific coding.
- Context
- Reflection reports the Beam column. Effort, harness, run date, sample size and score units are not specified in the table. The raw score is retained without adding a percentage unit.
CriPT AA (Show measurement, test conditions, and source)
- Source value
- 16.3
- Score
- 16.3
- Metric
- reported score
- Unit
- No unit provided
- Category
- reasoning
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Reflection Beam launch
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Beam launch evaluation, excluding Inkling and Beam Base. Exact row spelling CriPT AA retained.
- Context
- Reflection reports the Beam column. Effort, harness, run date, sample size and score units are not specified in the table. The raw score is retained without adding a percentage unit.
GPQA Diamond (Show measurement, test conditions, and source)
- Source value
- 90.5
- Score
- 90.5
- Metric
- reported score
- Unit
- No unit provided
- Category
- reasoning
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Reflection Beam launch
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Beam launch evaluation, excluding Inkling and Beam Base. Diamond subset.
- Context
- Reflection reports the Beam column. Effort, harness, run date, sample size and score units are not specified in the table. The raw score is retained without adding a percentage unit.
Page 1 of 2
Sources and data date
| Type | Evidence and data date |
|---|---|
| Input | Documented modalities. Retrieved October 7, 2026. |
| Output | Documented modalities. Retrieved October 7, 2026. |
| Research status | Research date October 7, 2026. 2 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not found in the inspected sources | independent-benchmarks Original wording The inspected official launch supplies Reflection's results. A separately published exact-model independent evaluation was not found. |
| Additional source | Reflection developer model documentation (retrieved October 7, 2026) · Reflection developer model documentation · Editorial description reviewed October 7, 2026 |
| Additional source | Reflection reasoning documentation (retrieved October 7, 2026) · Reflection reasoning documentation · Editorial description reviewed October 7, 2026 |
| Additional source | Reflection’s verified Hugging Face organization (retrieved October 7, 2026) |
| Additional source | Reflection commercial terms (retrieved October 7, 2026) |