Reflection's Beam Is a 501B Open-Weight Coding Model That Isn't Out Yet — That Gap Is a Paid Evaluation-Readiness Service

Reflection just told the market exactly what to expect from Beam — and exactly when they'll be able to get it.

That gap between announcement and weights is the business.

The headline claim, verbatim

Beam is a sparse Mixture-of-Experts model: "501 billion" total parameters with "23 billion" active per token, pretrained on "23.8 trillion diverse, curated, high-quality tokens" in "under four weeks" using "6,144 NVIDIA GB300 NVL72 GPUs." It is licensed Apache 2.0.

On Reflection's own benchmark table: "SWEBench Verified: 80.9," "Terminal Bench v2.1: 80.1," "SWE Bench Pro v2-Hard: 77.2," and "AIME 2026: 97.8." Context runs to "256K tokens" during RL, extending to "1M tokens" post-midtraining.

The catch: it isn't shipping today

Reflection's own framing: weights, technical report, model card, and developer artifacts arrive "later this month." The model is currently "undergoing final red-teaming and evaluations." Right now there's an early access signup — not a download.

Why this is a service, not a weekend project

Reflection published its own rubric: DeepSWE, SWE Bench Pro, Terminal Bench, SWEBench Verified and Multilingual, MCP Atlas, BrowseComp. Reproducing that evaluation against a client's actual codebase and actual coding-agent workflow — not a generic demo repo — takes setup: harness wiring, task selection, cost accounting against their current provider, and a clear pass/fail bar agreed before the test runs.

That's a fixed-scope engagement that can be sold and delivered before Beam's weights even exist.

The moneyPlay in practice

1. Sell a pre-launch 'open-weight coding agent eval-readiness' package to teams currently running Claude- or GPT-based coding agents, built around Beam's own published benchmark suite

2. Baseline the client's current agent today — token cost per task, pass rate on their own repos, context usage — so there is a same-task comparison the moment Beam's weights ship

3. Pre-wire the client's harness for Apache 2.0 licensing and the 256K/1M-token context window so swapping in Beam is a config change, not a rebuild

4. Offer a 'day-one eval sprint' retainer: the moment early access or open weights land, run the client's existing benchmark suite against Beam and deliver a swap/no-swap recommendation inside an agreed window

5. Extend into an ongoing benchmark-monitoring retainer, since Reflection shipped this in under four weeks of pretraining and will likely iterate fast

Bottom line

Reflection told the world the benchmark numbers before the weights were available to verify them. That's a trust gap most buyers won't close alone — and a fixed-scope readiness engagement is exactly the thing that closes it for them.

Sources:

https://reflection.ai/blog/introducing-beam