Reflection releases Beam 501B open-weight model with 23B active parameters, 1M context and 81.6% SWE-bench Verified

Reflection AI has officially released Beam 1.0, a 501-billion-parameter open-weight language model designed for coding, advanced reasoning and autonomous AI agents. Following its initial announcement on October 5, the company published model weights on October 16, 2026, under the Apache 2.0 license, alongside deployment integrations, a model card and safety documentation.

Beam uses a sparse Mixture-of-Experts (MoE) architecture with 501.3 billion total parameters and approximately 23.4 billion activated during token processing. Its context window reaches one million tokens, while released instruction-tuned checkpoints support BF16, FP8 and NVFP4 precision formats.

The final model card reports 81.6% on SWE-bench Verified, 80.1% on Terminal-Bench 2.1 and 97.7% on AIME 2026 for BF16. FP8 reaches 80.2% on SWE-bench Verified and NVFP4 reaches 80.7%. Reflection is positioning Beam as an efficient alternative to larger open-weight models, emphasizing the computational cost of generating successful solutions rather than model size alone.

The release moves Beam beyond an early-access announcement: developers can obtain weights, deploy the model on their own infrastructure, customize it and integrate it into software engineering workflows.

··········

BEAM 1.0: TECHNICAL SPECIFICATIONS AND RELEASE DETAILS.

Beam combines a large total parameter count with sparse activation, allowing only a fraction of its parameters to participate in processing each token. The architecture includes 52 layers, comprising three dense layers and 49 MoE layers, with 256 experts and a vocabulary of approximately 152,000 tokens.

........

Specification	Reflection Beam 1.0
Developer	Reflection AI
Initial announcement	October 5, 2026
Public weights release	October 16, 2026
Model type	Sparse Mixture-of-Experts
Total parameters	501.3 billion
Active parameters	23.4 billion
Transformer layers	52
Dense / MoE layers	3 / 49
Number of experts	256
Attention architecture	Gated GQA and sliding-window attention
Maximum context	1 million tokens
Vocabulary	Approximately 152,000 tokens
Supported modalities	Text input and output
Available precisions	BF16, FP8, NVFP4
Knowledge cutoff	June 30, 2026
License	Apache 2.0

........

The released model is primarily developed for English, although its software engineering capabilities extend to multiple programming languages. Beam is text-only: its million-token context should not be confused with native image, audio or video understanding. Its intended uses include software development, long-horizon agentic workflows, mathematical reasoning, scientific problems and tool-assisted information retrieval.

··········

THE 501B MIXTURE-OF-EXPERTS ARCHITECTURE ACTIVATES ONLY 23.4B PARAMETERS.

Beam separates total model capacity from the parameters actively used during inference. A dense 501-billion-parameter model would ordinarily involve nearly all of its weights in processing each token. Beam instead routes computation through selected experts, activating approximately 23.4 billion parameters.

23.4B ÷ 501.3B × 100 = 4.67%.

Approximately 4.7% of Beam's total parameters are active per token, corresponding to a total-to-active ratio of 21.4:1. Sparse activation reduces computation relative to a similarly sized dense model but does not eliminate the need to store the complete expert set, since different tokens may activate different experts.

Beam also uses gated grouped-query attention and sliding-window attention to manage attention computation and memory. Reflection describes normalization and residual-stream stabilization intended to support training stability across its 52 layers, particularly during long-sequence and reinforcement-learning workloads.

··········

SWE-BENCH VERIFIED REACHES 81.6%, WITH DIFFERENT RESULTS ACROSS PRECISION FORMATS.

Reflection's final model card provides benchmark results for three numerical formats. Quantization does not produce identical outcomes across tasks: lower precision can improve deployment efficiency, but performance effects vary.

........

Coding benchmark	BF16	FP8	NVFP4
SWE-bench Verified	81.6%	80.2%	80.7%
SWE-bench Pro v2-Hard	76.5%	78.4%	78.6%
SWE-bench Pro v1	66.1%	65.9%	65.6%
Terminal-Bench 2.1	80.1%	78.9%	78.4%
SWE-bench Multilingual	78.3%	78.1%	77.5%
DeepSWE v1.1	43.3%	43.0%	42.7%
SWE Atlas Codebase QnA	31.8%	32.2%	31.4%

........

BF16 achieves the highest reported SWE-bench Verified result, while NVFP4 slightly exceeds BF16 on SWE-bench Pro v2-Hard. These differences explain why a single benchmark result should not be attributed to every checkpoint.

SWE-bench Verified, SWE-bench Pro and Terminal-Bench evaluate different aspects of software engineering; their scores are not interchangeable measures of general coding ability. Production performance also depends on the coding harness, repository access, testing environment, tool permissions and inference settings. The figures are Reflection's reported results, not independent tests by Data Studios.

··········

BEAM ALSO TARGETS ADVANCED REASONING, SCIENCE AND TOOL USE.

Beam's training emphasizes mathematical reasoning, scientific problem solving and multi-step tool interactions alongside software engineering.

........

Reasoning benchmark	BF16	FP8	NVFP4
AIME 2026	97.7%	98.0%	97.5%
GPQA Diamond	90.6%	90.3%	89.3%
LiveCodeBench	88.3%	88.1%	87.4%
Humanity's Last Exam, no tools	36.0%	36.0%	34.8%
Humanity's Last Exam, with tools	51.1%	51.1%	51.1%
SciCode	49.2%	48.3%	46.9%

........

Humanity's Last Exam illustrates the contribution of external tools: BF16 increases from 36.0% without tools to 51.1% with tools, or 15.1 percentage points. That increase should not be attributed solely to the underlying model, because tool access changes the overall problem-solving system. Effective external resource use is nevertheless a central capability for autonomous agents.

··········

AGENTIC BENCHMARKS MEASURE AUTOMATION, BROWSING AND TOOL INTEGRATION.

Reflection also evaluates automated workflows, MCP-based tool use, browsing and information retrieval.

........

Agentic benchmark	BF16	FP8	NVFP4
AutomationBench	39.3%	38.2%	39.8%
MCP Atlas	80.8%	81.5%	80.5%
Tau3 Banking	36.1%	36.8%	35.4%
BrowseComp with context management	76.8%	76.7%	75.6%
DeepSearchQA with context management	80.5%	79.6%	81.1%

........

These benchmarks address workflows involving several external systems rather than isolated question answering. An agent may need to locate information, call an API, interpret a response, update a document and verify the outcome. Each step creates opportunities for tool-selection errors or incomplete execution.

Strong reasoning scores do not guarantee reliable automation. Reflection's figures provide evidence under tested conditions, but enterprises should validate performance on their own workflows and permission structures.

··········

THE ONE-MILLION-TOKEN CONTEXT SUPPORTS LARGE REPOSITORIES AND LONG AGENT SESSIONS.

Beam supports up to one million context tokens, accommodating extensive software repositories, long technical documents and multi-stage agent interactions. Reflection developed long-context capabilities during midtraining with structured repositories, long-horizon tasks and extended documents.

The objective was to identify and connect relevant information across lengthy sequences before reinforcement learning. However, a large advertised context window does not guarantee equal accuracy at every position. Retrieval quality, attention behavior, prompt organization and irrelevant material affect results.

A million-token prompt also requires significant prefill computation and attention-related memory. For many coding workflows, targeted retrieval and selective context management may remain cheaper than repeatedly loading an entire repository. Large context is flexibility, not a requirement to use its maximum capacity.

··········

REFLECTION TRAINED BEAM ON 23.8 TRILLION TOKENS AND MORE THAN 100 MILLION RL ROLLOUTS.

Reflection reports pretraining on 23.8 trillion tokens from web content, technical documentation, code, educational material and licensed datasets. Large-scale reinforcement learning subsequently targeted coding, reasoning and agentic performance.

........

Training metric	Reported scale
Pretraining corpus	23.8 trillion tokens
Reinforcement-learning rollouts	More than 100 million
NVIDIA GB300 GPUs	Approximately 10,500
Major RL campaign duration	Four weeks
Training and evaluation sandboxes	Approximately 1.3 billion
High-quality task environments	Approximately 1 million
Maximum RL rollout context	256,000 tokens

........

The 256K maximum RL rollout context and 1M supported inference context describe different development stages and are not contradictory. Reflection also reports extensive sandbox-based training and evaluation across large GPU and CPU fleets.

The company says capabilities continued improving as RL compute increased, although observations within the tested range do not establish unlimited scaling. Substantial training computation was invested to improve the efficiency of later inference-time reasoning.

··········

REFLECTION CLAIMS APPROXIMATELY THREE TIMES GREATER REASONING-TOKEN EFFICIENCY.

Reflection reports that Beam can achieve comparable solution quality to leading open models while using approximately three times fewer inference-time reasoning tokens in certain comparisons. This is not equivalent to a threefold reduction in total serving cost.

Generation length is only one expense. Model activation, memory bandwidth, prompt prefill, batch size, hardware utilization and serving overhead all affect cost.

Reflection estimates generation computation with FLOPs ≈ 2 × active parameters × generated tokens. For Beam and 10,000 generated tokens:

2 × 23.4B × 10,000 = 468 trillion FLOPs.

This approximate forward-pass calculation excludes attention, prefill, memory movement and infrastructure overhead. It nevertheless illustrates why sparse activation and shorter reasoning traces can be economically significant for recurring workloads.

··········

BF16, FP8 AND NVFP4 CREATE DIFFERENT MEMORY REQUIREMENTS.

The full 501.3-billion-parameter count remains relevant regardless of how many parameters are active per token. Raw weight storage can be estimated by multiplying total parameters by bytes per weight.

........

Precision	Approximate bits per weight	Theoretical raw weight storage
BF16	16	1,003 GB
FP8	8	501 GB
NVFP4	4	251 GB

........

These simplified decimal-gigabyte estimates exclude quantization metadata, scaling factors, runtime buffers, KV cache and serving overhead. They are not official minimum GPU-memory requirements.

Actual deployment can require additional memory for expert routing, attention states, concurrency and long-context processing. NVFP4 reduces weight storage substantially, but a model with approximately 251 GB of theoretical four-bit weights remains a major infrastructure workload. Sparse activation primarily reduces computation per token; it does not make Beam a small local model.

··········

BEAM IS AVAILABLE THROUGH OPEN WEIGHTS, HOSTED APIS AND ENTERPRISE INFRASTRUCTURE.

Reflection has published Beam's weights under Apache 2.0, enabling download and self-hosting rather than exclusive reliance on a proprietary API. The company has announced an ecosystem of more than 40 partners and integrations.

Supported inference and development technologies include Hugging Face Transformers, vLLM, SGLang, Unsloth, Axolotl and NVIDIA Dynamo. Infrastructure partners span hosted inference, cloud and enterprise deployment services.

Obtaining weights is only the first step toward production deployment. Organizations also need compatible inference engines, optimized kernels, monitoring, security controls and operational processes. Reflection is distributing a speculative-decoding drafter model intended to improve generation speed under suitable configurations.

Costs, throughput and latency will vary across providers and hardware configurations.

··········

MIRROR CLI CONNECTS BEAM TO REAL SOFTWARE ENGINEERING WORKFLOWS.

Reflection has introduced Mirror CLI, a coding-agent environment for Beam and compatible models. Developers can interact through a terminal, inspect agent activity, approve tool calls and guide execution.

Features include resumable conversations, parallel agents in separate worktrees, project instructions, skills, MCP connections and Python extensions. Its architecture separates a Python frontend from a Rust agent core.

Reflection says Mirror originated in internal engineering workflows, including code review and background agents. Beam has also been tested with OpenCode, Pi and Hermes.

Interoperability matters because a coding model's usefulness depends on reading repositories, editing files, executing commands and responding to tests. The same model may perform differently across agent harnesses, so comparisons should normalize execution environments as well as model settings.

··········

OPEN WEIGHTS ENABLE FINE-TUNING, DISTILLATION AND PRIVATE DEPLOYMENT.

Apache 2.0 allows developers to inspect, modify and redistribute Beam under the license's terms. Reflection describes support for fine-tuning, parameter-efficient adaptation, quantization and distillation, alongside integrations for post-training and reinforcement learning.

One proposed customization method is self-distillation: generate task-specific traces, select verified examples and train an adapter for the target workflow. Reflection has indicated that further self-distillation recipes will follow the initial release.

Customization matters for specialized codebases, internal terminology and repeated operational procedures. Private deployment can also offer control over data handling and infrastructure location, but open weights do not automatically guarantee security, compliance or low cost.

··········

SAFETY EVALUATIONS SHOW BOTH REFUSAL BEHAVIOR AND LEGITIMATE-REQUEST COMPLIANCE.

Reflection has published safety documentation for harmful requests, benign interactions and agentic risks. Its model card reports the following Iris Safety Evaluations.

........

Safety metric	BF16	FP8	NVFP4
Harmful requests: refusal or pivot rate	88.9%	89.7%	88.5%
Harmless requests: compliance rate	81.5%	81.6%	82.1%

........

The figures illustrate a trade-off between rejecting unsafe instructions and answering legitimate ones. Higher refusal is not automatically preferable if it also increases rejection of harmless tasks; high compliance is insufficient if unsafe requests are followed.

For coding agents, security also depends on the surrounding harness. Shell commands, file editing and external service access create operational risks regardless of model benchmark scores. Enterprises should evaluate sandboxing, credential access, permissions and auditability across the complete agent system.

··········

WHAT BEAM CHANGES FOR OPEN-WEIGHT CODING MODELS.

Beam enters a market where benchmark performance, reasoning-token efficiency, memory footprint, deployment flexibility and agentic tooling all affect model selection.

Its 501.3B total parameters, 23.4B active parameters, million-token context and three precision formats combine substantial capacity with sparse computation. The final model card reports 81.6% SWE-bench Verified for BF16 alongside strong reasoning and tool-use scores.

Reflection's release strategy extends beyond weights, with inference integrations, customization workflows and Mirror CLI. Important limitations remain: infrastructure requirements are substantial, long contexts carry significant prefill costs, and the claimed reasoning-token advantage is not a universal reduction in total inference expense.

Benchmark results also depend on evaluation harnesses and must be validated on production tasks before deployment decisions.

Beam combines a 501B-parameter open-weight model, approximately 23B active parameters and an ecosystem designed for autonomous coding agents. Its commercial value will depend on converting those capabilities into reliable completed tasks at a competitive total cost.

··········

FOLLOW US FOR MORE.

DATA STUDIOS

datastudios.org