Reflection AI's first model, and the first thing the lab trained itself — a sparse MoE with 501B total parameters and 23B active across 52 layers, aimed at coding, reasoning and agentic work. Announced on October 5, 2026 with an early-access signup rather than a download: Beam is still undergoing final red-teaming, and the weights, technical report, model card and fine-tuning stack are promised later in October under Apache 2.0. The API serves it as Beam-501B-A23B with a 262,144-token context, 131,072-token maximum output and mandatory reasoning at five effort levels; knowledge cutoff is June 30, 2026. Reflection reports midtraining extended the effective context to 1M tokens, which the beta API does not yet expose.

The weights landed on 2026-10-07, two days after the announcement and ahead of the "later in October" promise, as reflection/Beam-1.0-501B-A23B — 188 safetensors shards, ungated, genuinely Apache 2.0 with no community-licence rider, alongside FP8, NVFP4 and DFlash2 builds. Two details only the release settles. The published config.json carries max_position_embeddings: 1048576, so the 1M context is now exposed in the weights even though the beta API capped at 262K; and it confirms the routing as 256 experts with 8 active per token over 52 layers. Artificial Analysis scores it 18 on Intelligence Index v5.0 (30 on v4.3.2, when it ranked 19th of 120 open-weights models in its size class), with no public API so the figure is from self-hosted inference.

The architecture is built for RL to run on top of it without falling over. Interleaved local and global attention, fine-grained routed experts, and three stacked stability choices: auxiliary-loss-free load balancing after DeepSeek, extended with cosine decay of the expert-bias updates to quiet routing churn late in training, plus sequence-level balancing to keep utilization even on data outside the pretraining distribution — the finished base reportedly reaches a busiest expert load of just 1.04×. The residual stream gets depth-based scaling, SandwichNorm, elementwise attention gating and FP32 accumulation, which Reflection says holds residual RMS bounded across all 52 layers. Pretraining ran on 23.8T tokens over 6,144 NVIDIA GB300 NVL72 GPUs in under four weeks, after a ladder of smaller models established a scaling recipe the lab says held across four orders of magnitude of compute. The RL phase is the larger number: over 100 million rollouts on 10.5K GB300s across four weeks. Alignment was trained as a separate model from the same base, then merged with the capability teacher by multi-teacher on-policy distillation, using deliberative alignment and an adversarially-rebuilt SFT mixture; they report forecasting alignment-RL gains at r = 0.79 against 0.46 for a Best-of-N ceiling.

On where it actually lands, Reflection is more careful than the coverage around it. Its own table puts Beam roughly level with GLM 5.2 and ahead of Inkling and Nemotron 3 Ultra, while GLM 5.3, Kimi K3, Qwen 3.8-Max and DeepSeek V4.1 Flash stay ahead on raw capability: Terminal Bench v2.1 80.1 against 88.2 / 88.3 / 86.6 / 90.6, HLE without tools 36.2 against 42.3 / 46.9 / 43.6 / 39.1, and DeepSWE v1.1 44.4 against DeepSeek's 74.2. Beam's best showings are SWE-Bench Verified 80.9, SWE-Bench Pro v2-Hard 77.2, SWEBench Multilingual 78.0, AIME 2026 97.8 and GPQA Diamond 90.5. The actual pitch is efficiency: GLM-5.2-comparable reasoning at 3–4× less inference compute, which is a claim about active-parameter count rather than accuracy, and one Reflection computes itself as 2 × active parameters × mean generated tokens using Artificial Analysis and DataCurve figures. All benchmark numbers are self-reported. (At launch no capabilities technical report had been published; see the last paragraph.) Laskin told Semafor that CAISSI and the UK AI Security Institute are assessing Beam, which neither the launch post nor Reflection's safety page mentions and neither body has confirmed.

The release package was completed on October 16, 2026 ("Beam is now available"): a model card, a Beam Safety Technical Report, a speculative-decoding drafter, and day-zero fine-tuning support across NVIDIA NeMo RL / AutoModel / Megatron, RadixArk Miles, Hugging Face TRL, Anyscale SkyRL, Axolotl and Unsloth, with self-distillation recipes promised "in the coming days". The post puts RL scale at about 1.3 billion sandboxes across training, evaluation and ablations, peaking at 883,006 vCPUs over 23 clusters on GCP and AWS. The model card restates the 1M-token context and reports per-precision results (BF16 / FP8 / NVFP4): SWE-Bench Verified 81.6 / 80.2 / 80.7, Terminal Bench v2.1 80.1 / 78.9 / 78.4, HLE without tools 36.0 / 36.0 / 34.8 (51.1 with tools), AIME 2026 97.7 / 98.0 / 97.5 and GPQA Diamond 90.6 / 90.3 / 89.3. The safety report is an evaluation document, not a training report: Reflection rates Beam "Limited" in all four systemic-risk categories of its Open Safety Framework, with external evaluation inputs from SecureBio, Irregular, Nemesys Insights, Andon Labs and Gray Swan. It does not mention CAISI or the UK AI Security Institute.

The capabilities technical report followed in early November 2026: the 156-page Beam Model Technical Report (Reflection Team et al., PDF last modified 2026-11-07 UTC; its citation block points at reflection.ai/beam-technical-report, which did not resolve on 2026-11-10, and the HuggingFace card still read "coming"). It gives the exact scale as 501.3B total and 23.4B active parameters over 52 layers (49 MoE and 3 dense), 256 routed experts with 8 active and no shared expert, a sigmoid top-k router, a 151,936-token vocabulary, 48 query and 8 KV heads, and a 5:1 mix of local (512-token sliding window, RoPE) and global (NoPE) attention with QK-norm and gated attention. The one new architectural piece it names is a depth-scaled residual connection (DSRes), which rescales the residual stream by 1/√n of the accumulated sublayers to keep its norm bounded; length extrapolation uses scalable softmax. The scaling recipe was fitted on a ladder from 300M to 23.4B active parameters across four orders of magnitude of pretraining compute. Pretraining on 6,144 GB300s reached 92.3% goodput late in the run (up from 73.2% before an infrastructure relaunch) and 21% MFU in steady state. The RL platform ran 110K concurrent rollouts with 12-second weight updates, and the report says capabilities "continued to improve as we increased RL compute, with no sign of a plateau". It also documents the DFlash 2 drafter (a 1.93B-parameter block-diffusion model) and reports that scaling the alignment teacher from 100B to 501B parameters cut false-positive over-refusals by 48.7% relatively. The fine-tuning cookbook, including the promised self-distillation recipes, is at reflection-oss/cookbook (repo created 2026-10-19).

Model Details

Architecture MOE
Parameters 501.3B
Active params 23.4B
Experts 256 (top-8)
Context window 1,048,576
Training tokens 23.8T
AA Intelligence 18 was 30 on v4.3.2
Training hardware 6,144 NVIDIA GB300 NVL72 (pretraining); 10.5K GB300 (RL)
Training time under 4 weeks pretraining; 4 weeks RL
License Apache 2.0

Benchmark Scores

Benchmark Score Mode
SWE-Bench Verified 80.9 —
SWE-Bench Pro v2-Hard 77.2 —
SWEBench Multilingual 78.0 —
Terminal Bench v2.1 80.1 —
AIME 2026 97.8 —
GPQA Diamond 90.5 —
HLE (no tools) 36.2 —
MCP Atlas 78.7 —
moeopen-weightreasoningagenticcodelong-context

Related