Beam vs Kimi K3: what the challenger's own table shows

On October 5, Reflection AI, the Nvidia-backed lab that raised on the promise of an American open-weight frontier model, shipped Beam: a 501B-parameter sparse MoE with 23B active, trained on 23.8T tokens. The launch post includes a benchmark table, and the table has a target painted on it. Of the 14 benchmarks where Beam and Kimi K3 both report scores, every one lists Kimi K3 as the column to beat.

We sell Kimi K3 inference, so we read that table the way you read a bid on your own house. Twice.

Disclosure up front, because it shapes everything below: Wallaby Token operates an OpenAI-compatible inference API that serves Kimi K3 at 10% off the official rate card. We are not a neutral observer. What we can offer instead of neutrality is provenance: every number in this post carries its source, and every claim we could not verify gets marked as such.

The scoreline

Across the 14 shared benchmarks in Reflection's own table, updated October 8, Kimi K3 leads 13. Beam edges ahead on exactly one: tau3 banking, 38.0 to 37.1, a narrow margin inside any reasonable noise band.

The gaps that matter for agentic coding are not subtle:

Two things about this table before anyone quotes it.

First, both columns were run by Reflection, on Reflection's harness. That cuts both ways. It is the strongest kind of concession, a challenger measuring itself against you and publishing the loss. It is also not independent measurement. Moonshot's own model card scores K3 slightly differently on shared rows: 67.5 versus 68.0 on DeepSWE, the same 88.3 on Terminal-Bench. We have not merged the two sources, and you shouldn't either.

Second, read the other columns while you're in there. The same table has DeepSeek V4.1 Flash ahead of both Beam and K3 on DeepSWE at 74.2 and Terminal-Bench at 90.6, and Qwen 3.8-Max far ahead on tau3 banking at 55.2. Kimi K3 wins this particular matchup. It does not win the whole page.

The sentence that matters most

The launch post contains its own summary, and it is more precise than the coverage:

"Where frontier open models like Kimi K3 remain ahead on raw capability, Beam's advantage is efficiency at inference time."

That is the actual claim worth evaluating. Not "Beam beats K3"; Reflection never says that. The claim is that Beam gets you GLM-5.2-class reasoning at 3–4× less inference compute.

Here is the methodology behind that efficiency figure, from their Figure 2 notes: FLOPs are estimated as 2 × active parameters × generated tokens, using Artificial Analysis and DataCurve evals for other models. It excludes prompt prefill, attention over context, and serving overhead, which are the dominant cost drivers in long-context agent workloads. Their own caption says it plainly: "an approximate compute comparison rather than measured inference cost."

Efficiency per FLOP is genuinely interesting for self-hosters. But most buyers don't pay in FLOPs. They pay in dollars per accepted task, and there the Beam column is empty: no published price, no self-serve access, no weights to cost against. Until one of those lands, there is no number to multiply by.

What you cannot do with Beam today

This is the section the announcement coverage skipped.

You cannot download it today. The weights, technical report, and model card are promised under Apache 2.0 "later this month." As of October 10, nothing is on Hugging Face, and Beam is absent from OpenRouter, Ollama, and the models.dev catalog. You cannot get a third-party score: Artificial Analysis has no Beam page, so every capability number in circulation traces to one lab's table. You cannot try it without joining a waitlist. If you get in, the beta serves a 256K context window with a footnote that it may change; Beam's midtraining extends effective context to 1M tokens, but that is not what the API exposes today. And it is text-only at launch.

Contrast with the model it aimed at. Kimi K3's weights have been downloadable since July 27. It is served by 20 providers on OpenRouter alone, plus Moonshot's own API, at a published $0.30 / $3.00 / $15.00 per million tokens for cache hit, cache miss, and output. It takes text, images, and video. Its context window is 1M tokens in production, not in a training footnote.

None of this is permanent. Beam's Apache 2.0 release will close most of these gaps within weeks, and Reflection says it will launch with distribution partners and open-source harness integrations already wired. For anyone choosing a model this week, though, the gap is the decision.

The traffic already voted

Benchmark tables are claims; routed traffic is behavior. On OpenRouter, daily Kimi K3 volume ran about 16B tokens in late September and peaked at 63.5B on October 7, the week of the Beam launch. The biggest K3 consumers are coding agents: Hermes Agent at 306B tokens over the trailing window, Claude Code at 297B, pi at 210B, Codex at 132B, OpenClaw at 71.5B. Codex's K3 usage roughly quadrupled in two weeks, helped along by Moonshot publishing an official Codex integration guide on September 28.

Some of that spike is event heat and will cool. October 9 already settled back to 31B, still about 40% above the September baseline. The floor moved.

What happens next

Beam's weights landing will be the real test, and we'll run it. Same method we use for our own evaluations: identical tasks through the same agent harness, cost per passing task at published rates, first-response latency metered, cache behavior logged. We ran four reasoning-effort tiers through that pipeline on our own gateway last week, and Beam goes through it the day a key or a checkpoint is available. If it is genuinely strong at its price point, it clears the same bar we set for anything we consider serving.

Until then, the honest scoreboard reads: capability, K3 by Reflection's own count; efficiency, an unpriced estimate; availability, K3 in production today.

Sources and further reading

• Introducing Beam: Reflection's 501B open-weight model: the announcement and the benchmark table, updated October 8
• Reflection developer docs: Beam-501B-A23B model ID, OpenAI-compatible endpoint, beta access notes
• Moonshot AI: Kimi K3 API pricing: the official rate card
• OpenRouter: Kimi K3 model page: provider list and app-level traffic, pulled October 10
• Fortune: Reflection AI unveils Beam: launch coverage and the US-vs-China open-weight framing

Deeper: Kimi K3 API pricing across 18 providers, refreshed daily by our own monitor. Wider: Open-weight LLM API pricing, October 2026, the four major open models on the same workload. Hands-on: Kimi K3 in Codex, the setup that Codex's K3 traffic is flowing through.