Reflection AI's Beam Bets That Efficiency Beats Benchmarks in the Open-Weight Race

Most model launches this year have followed the same script: ship a chart, claim a new state-of-the-art number, repeat next month. This week's most interesting release broke that pattern. On October 5, Reflection AI published benchmark results for Beam, a 501-billion-parameter open-weight model aimed at coding and agentic work — and the numbers it published are, by its own admission, mixed. The pitch isn't "we're the best." It's "we're the most efficient," and that's a more interesting bet than it sounds.

What Beam actually is

Beam is a sparse mixture-of-experts model: 501 billion total parameters, but only 23 billion active per forward pass, which is what makes the efficiency argument possible in the first place. It's purpose-built for coding, reasoning, and agentic workloads, and Reflection plans to release the weights under an Apache 2.0 license later this month — meaning anyone will be able to self-host, fine-tune, or audit it, not just call it through an API.

The scorecard: ahead on one axis, behind on another

On straight coding benchmarks, Beam leads: 80.9 on SWE-bench Verified and 78.0 on SWE-bench Multilingual, ahead of rival open model Inkling across four shared coding tests. But on the two benchmarks that measure multi-step agentic behavior rather than single-shot code fixes, Beam falls behind: it trails DeepSeek V4.1 Flash by roughly 10 points on Terminal-Bench and by nearly 30 points on DeepSWE, with Kimi K3 and GLM 5.3 also beating it on most rows.

Worth flagging: these are Reflection's own reported numbers. Since the public checkpoint hasn't shipped yet, none of this is independently verifiable until the weights land — a caveat that applies to basically every pre-release benchmark table and is easy to forget when a chart looks convincing.

The actual pitch: compute, not leaderboard position

Reflection's real claim is that Beam matches Z.ai's GLM-5.2 on advanced reasoning benchmarks while using 3-4x less inference compute. For a lab, that's a strange thing to lead with instead of a leaderboard win. For anyone actually running agents in production, it's arguably the more useful number — inference cost and latency compound across millions of calls in a way that a few benchmark points don't. A model that's "good enough" at a quarter of the compute cost can beat a model that's "slightly better" at full price, once you're the one paying the GPU bill instead of reading the launch post.

Seven billion dollars in compute, before shipping a flagship model

Reflection AI is barely two years old. It was founded in 2024 by two Google DeepMind alumni — CEO Misha Laskin, who led reward modeling on Gemini, and CTO Ioannis Antonoglou, a co-creator of AlphaGo. The company raised a $2 billion round led by Nvidia in late 2025 that pushed its valuation to $8 billion, and is reportedly now in talks to raise further at a valuation north of $20 billion.

What's more striking is the compute Reflection has locked up before Beam has even fully shipped: roughly $150 million a month in capacity from SpaceX through 2029 (worth on the order of $6 billion over the life of the deal), plus a separate $1 billion-plus, multi-year compute agreement with Nebius that includes access to Nvidia's GB300 chips. That's over $7 billion in committed compute for a company whose first flagship open model is still mid-rollout.

Why this is bigger than one model

Open-weight leaderboards have been dominated by Chinese labs for most of this year — DeepSeek, Kimi, GLM, Qwen. Reflection is explicitly positioning itself as one of the few well-funded U.S. labs trying to compete there rather than cede the category, while OpenAI and Anthropic stay focused on closed, API-only frontier models. Whether Beam actually closes that gap once the weights are public and independently testable is still an open question — on the published numbers, it doesn't uniformly win.

For engineering teams evaluating open models to self-host or fine-tune, that's the detail worth tracking over raw leaderboard position: license terms, real-world inference cost per token, and whether the benchmark story holds up once outsiders can run it themselves. Beam's weights are due later this month — that's when this story actually gets tested.

Sources

• TechCrunch — Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute cost

• Reflection AI — Introducing Beam: Reflection's 501B open-weight model

• Help Net Security — Reflection's Beam trails top open models on coding tests but claims lower inference compute

• The Next Web — Reflection signs a $1bn AI compute deal with Nebius, its second big capacity grab in a month

• Mobile World Live — SpaceX inks computing deal with Reflection AI

• Folio3 — Reflection AI Secures $2 Billion in Massive Funding Round, Valuation Soars to $8 Billion