technically dispatch kimi liquid stacked prs

Why this is in the vault

A three-story roundup — Kimi K3's revenue-gated open-weight license, Liquid AI's on-device small model, and GitHub's native stacked-PR feature — worth keeping mainly for the last item, which lands directly on RDCO's own PR-only repo workflow.

Mapping against Ray Data Co

The concrete connection is the stacked-PRs story: RDCO runs a PR-only workflow on its repos (branch + PR + autonomous review/merge, no human gate — see CLAUDE.md hard rule and feedback_pr_only_workflow). GitHub's public-preview stacked-PR feature (shipped July 30, 2026) exists to solve exactly the tension RDCO's agent-driven repo work already produces: agents generate code faster than a human (or even an autonomous reviewer) can absorb in one shot, so breaking a large change into an ordered stack of small, individually-reviewable PRs that still merge in dependency order is a workflow RDCO's own review/merge automation should track — it's an argument for chunking agent-authored changes rather than shipping one large diff per task, which is a real, checkable adjustment to how the autonomous-review gate is structured.

The other two stories map more weakly. Kimi K3's license — Moonshot requiring $20mm+/year "Model as a Service" providers to negotiate commercial terms while individual devs stay free — is a useful open-weight licensing precedent to have on file, but RDCO isn't self-hosting or reselling open-weight models today; it's background, not an active build input. Liquid AI's LFM2.5-2.6B (2.6B params, CPU-only, 220 tok/s on an M5 Max) is a genuine on-device economics data point, but RDCO's agent stack runs on cloud Claude API calls, not local small models, so the connection is speculative rather than something changing a current decision. Both are filed for the open-weights/on-device threads already tracked in the vault, not because they change anything RDCO is doing this week.

Curation section

Kimi K3 licensing

Moonshot AI's Kimi K3 (2.8T params, 1M-token context) ships under a custom license, not standard open-source terms: MaaS providers over $20mm/year revenue must negotiate a commercial license unless certified; individual users/devs stay free through licensed providers. Inference providers (Modal, Baseten, Fireworks, DigitalOcean, RunPod, Together AI, Databricks) converged on identical pricing — "$3 per million input tokens, $15 per million tokens out" — undercutting Anthropic's rates, with Moonshot leaning on providers' inference optimization since its own infra couldn't handle demand.

Liquid AI on-device

Liquid AI released LFM2.5-2.6B: 2.6B params, ~2.5GB on disk, built to run on CPU rather than GPU clusters — reported at "220 tokens/second on an M5 Max." License permits commercial use under $10mm revenue; larger companies negotiate. Framed as evidence that, below a certain size, CPU inference is competitive with GPU inference for edge/laptop/phone deployment.

Stacked PRs on GitHub

GitHub shipped native stacked pull requests in public preview July 30, 2026 — codifying dependency order across a chain of related PRs so large changes can ship as small, individually reviewable, correctly-ordered chunks. Graphite pioneered the pattern in 2023 (later acquired by Cursor, then SpaceXAI); this is the first major platform to build it natively. Framed explicitly around the problem of coding agents outpacing team review capacity.

No sponsor block, "brought to you by" line, or affiliate link was present in the issue; the newsletter credited a correspondent (Sam Schneider) rather than running a paid placement. Zero third-party primary sources were deep-fetched beyond the newsletter's own reconstructed body (plaintext_body was empty in Gmail, so this note is built from a web reconstruction of the published post per source_fidelity above) — the newsletter's own summaries were detailed enough to paraphrase directly, and none of the three stories cleared the bar for a dedicated primary-source fetch beyond that.