Efficiency Over Leaderboards: Reflection AI's Beam and the American Open-Weight Bet

On Monday, 5 October 2026, New York-based Reflection AI announced Beam, its first open-weight language model. Reflection was founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou. It has raised more than $4 billion, most recently at a $25 billion pre-money valuation, with Nvidia and Sequoia among its backers. The company has spent most of its existence promising that America would get an open-model lab to compete with DeepSeek, Qwen, Moonshot and Z.ai. Beam is the first product it has released.

You cannot download it yet. Reflection says the weights, technical report, model card and fine-tuning tooling will ship “later this month” under Apache 2.0, after final red-teaming. Until then, access is through an early-access waitlist at platform.reflection.ai. Every number below comes from Reflection’s own announcement, reported consistently by TechCrunch and SiliconANGLE. None of it has been independently reproduced.

What Beam is

Beam is a sparse mixture-of-experts (MoE) model with 501 billion total parameters, of which 23 billion are active per token. Reflection’s post lists the main design choices:

52 layers with interleaved local and global attention. Most layers attend over a sliding window, and periodic layers attend over the full sequence. This is the standard way to make very long contexts affordable.
Fine-grained routed experts with auxiliary-loss-free load balancing, the approach DeepSeek popularised. Instead of adding a penalty term to the loss to force even expert usage, the router’s biases are adjusted directly. Reflection says the busiest expert ran at only 1.04× the average load.
FP32 residual accumulation, depth-based scaling and SandwichNorm. These are numerical-stability choices that matter when you train with very large batches.
A context window of 1 million tokens, extended during mid-training. Reinforcement learning ran at contexts of up to 256K tokens.

The ratio of active to total parameters is what defines the model. At 23B active, each generated token costs roughly as much compute as a mid-sized dense model, while 501B parameters’ worth of stored knowledge sits behind the router. For comparison, Reflection lists Z.ai’s GLM-5.2 at about 750B total, Qwen 3.8-Max at more than 2 trillion, and Thinking Machines’ Inkling at 975B total with about 41B active.

How it was trained

The engineering write-up is arguably more interesting than the model. According to Reflection:

Pre-training: 23.8 trillion tokens on 6,144 Nvidia GB300 GPUs, finished in under four weeks with goodput near 92% at the end of the run. The company says its filters discarded about 95% of raw web tokens.
Reinforcement learning: about 10,500 GB300 GPUs over four weeks, generating more than 100 million rollouts across 1 million coding, agentic and STEM environments. Grading used roughly 1.3 billion sandboxes, with up to 170,000 running concurrently. Reflection calls this “one of the largest scale RL runs conducted by any open lab to date”.
Asynchronous RL with heavy staleness tolerance. The trainer learned from rollouts up to a day old, as many as 107 weight versions behind the current policy, and Reflection says numerics stayed stable. New weights reached the inference fleet in a median of 12 seconds. The run survived 71 inference incidents with a median recovery time of 8 minutes and never stopped training.
Safety: a separate safety “teacher” model was trained from the same checkpoint and merged into the final model through multi-teacher on-policy distillation, alongside deliberative-alignment techniques.

For engineers, the takeaway is that frontier-style post-training is now largely an infrastructure problem: sandbox orchestration, weight broadcast and fault recovery. A two-year-old startup is publishing those numbers the way labs once published parameter counts.

The benchmarks, read honestly

Reflection’s headline claims are 80.9 on SWE-bench Verified, 65.5 on SWE-bench Pro v1 (against GLM-5.2’s 62.1) and 97.8 on AIME 2026. It also says Beam matches GLM-5.2 on advanced reasoning “while using 3–4× less inference compute”.

The full table in the announcement tells a more mixed story, and it deserves credit for publishing one. Beam’s comparison set includes newer Chinese open models, and against those it often trails by a wide margin:

Benchmark	Beam	GLM 5.3	Kimi K3	DeepSeek V4.1 Flash
DeepSWE v1.1	44.4	61.0	68.0	74.2
Terminal Bench v2.1	80.1	88.2	88.3	90.6
SWE Atlas Codebase QnA	34.6	61.0	68.0	—
HLE (no tools)	36.2	42.3	46.9	39.1
GPQA Diamond	90.5	91.7	93.5	90.9

Reflection says as much: “Where frontier open models like Kimi K3 remain ahead on raw capability, Beam’s advantage is efficiency at inference time.” DeepSWE is the benchmark this blog has used to compare open models. On it, Beam’s 44.4 sits far below StepFun’s Step 5 Preview at 67.7. The model Beam clearly beats is Inkling, the other well-funded American entrant.

The efficiency claim needs its footnote

The “3–4× less compute” figure is calculated, not measured. Reflection estimates FLOPs as 2 × active parameters × mean generated tokens. That calculation explicitly leaves out prompt prefill, context-dependent attention and serving overhead. It rewards two things: a small active-parameter count and short answers, which Reflection encouraged with a length penalty during RL. Both are real advantages. But it is a model-compute comparison and does not tell you your bill. On agentic workloads with long contexts, prefill and attention costs are often the larger share, and they scale with context length rather than active parameters. A 23B-active model serving 500K-token repository contexts will not be 3–4× cheaper than GLM-5.2 on that alone.

Why it matters anyway

Beam’s significance is less about the leaderboard and more about who released it, and on what terms:

Licensing and provenance. Apache 2.0 weights from a US company give enterprises and governments that cannot run Chinese-origin models a reasonably strong open option for coding agents. In interviews, Laskin described this demand directly as “sovereign AI” buyers who do not want to depend on Chinese models. In practice, that may matter more than the 20-point gap on DeepSWE.
Self-hosting economics. With 23B active parameters, Beam runs fast per token on a single multi-GPU node, provided you can hold 501B parameters in memory. At FP8 that is roughly half a terabyte of weights, so it is not a laptop model, but it fits on one 8-GPU H200/B200 server.
A fine-tuning base. Reflection is pitching Beam as a “workhorse” for companies building their own “AI factories” on proprietary data. An efficient MoE base with published RL tooling suits that better than a larger, stronger model you cannot afford to post-train.

Open questions

Will the weights match the scorecard? All results come from the company itself. The real test starts when the weights land and third parties run SWE-bench and DeepSWE in their own harnesses.
The safety report. Beam is still in red-teaming. The same week, Anthropic published research showing that safeguards on Z.ai’s open GLM-5.3 could be bypassed 64–100% of the time for building cyber exploits. How Reflection handles dual-use evaluation for a downloadable model will be closely watched.
Can the follow-on model close the gap? Laskin has said a more powerful successor is in training. Beam’s DeepSWE score suggests long-horizon software engineering is where the gap to Chinese open models remains largest.

For now, Beam is a credible, efficient, permissively licensed model with a refreshingly detailed engineering report. It is not, by its own numbers, the best open model in the world. Whether “good enough and American” is a winning product is the question the next few months will answer.