The last two weeks of this desk were previews: the MCP spec would finalize, Kimi K3's weights would drop, the benchmarks were coming. This week the placeholders resolve. The spec is final today. The weights are on Hugging Face. The numbers are public. And in all three cases, the honest story is the paragraph the launch posts left out.
Here's the whole week in one line: the headlines are true, and each one has a piece of fine print that changes what you actually do about it.
1. MCP is final today — and nothing breaks#
The 2026-07-28 revision publishes today, turning the release candidate into the standard. The change everyone named is real: MCP is stateless. The initialize/initialized handshake is gone (SEP-2575), the protocol-level session and Mcp-Session-Id header are gone (SEP-2567), and every request now carries its own identity. The payoff is operational, not conceptual: a server that needed sticky sessions and a shared session store can run behind a plain round-robin load balancer, routing on the new required Mcp-Method header.
The fine print is the part that should relax you. Deprecated features keep working for a minimum of 12 months. Roots, Sampling, and Logging (SEP-2577) don't switch off today, and older servers still interoperate. So the correct response to "MCP finalized" is not a weekend fire drill — it's putting the migration on a calendar. The one silent trap when you do migrate: the missing-resource error moved from -32002 to -32602, so any client that branches on the old code stops recognizing the error with no exception thrown. We walked the full break list here.
2. Kimi K3's benchmark is real — read the harness#
Kimi K3's numbers landed with the weights, and they're strong: independent aggregators put it around #4 on the Artificial Analysis intelligence view, behind only Claude Fable 5 and the GPT-5.6 Sol tiers and ahead of Claude Opus 4.8, with roughly 76.8% on SWE-bench Verified. For an open model you can download, that is genuinely near-frontier, and it's why the "shot across the bow" framing isn't hype.
One benchmark number, quoted without its harness, is marketing. The same model can move 10–26 points on coding tasks depending only on the scaffold you run it in.
The fine print: Moonshot's own launch table mixes evaluation harnesses — KimiCode, Claude Code, Codex, mini-SWE-agent — and harness choice alone swings coding scores by 10 to 26 points. That's not an accusation; it's how agentic benchmarks work. But it means the honest comparison for your stack is K3 and your current model run through the same harness, not K3's best-harness row against your model's default. Before you re-point a production coding agent, run the bake-off yourself. (We keep a standing note on why one benchmark number lies.)
3. The self-host math got smaller — literally#
Every preview, including ours, framed self-hosting a 2.8-trillion-parameter model as a cluster problem — "~18 H100s," a five-figure monthly floor. The public weights correct that. K3 was trained natively in MXFP4, not quantized after the fact, so the full weights are about 1.4TB instead of the ~5.6TB an FP16 model this size would demand. That fits a single 8×H200 node — or a single 8-GPU Blackwell/MI400 node at interactive latency, where the FP4 support is native.
That reframes the decision. Self-hosting K3 is no longer "stand up a cluster"; it's "provision one dense node." Full-precision serving still wants ~64 accelerators, but almost nobody needs that. The catch that survives: a dedicated node bills whether or not it's busy, and day-0 hosting on Together AI and Modal plus Moonshot's API means the elastic option is one API key away. Self-host for data residency, fine-tuning, or sustained heavy throughput — not for a demo. Our rent-vs-own breakdown still holds; the GPU count is just lower than we said.
4. "Open" weights, a not-open license#
This is the one most founders will skip and some will regret. The weights are free to download, but the terms are not MIT — they're a bespoke Kimi K3 License. You can use, modify, distribute, and sell what you build. But if you run a Model-as-a-Service — reselling inference or fine-tuning access with control over inputs and parameters — and your revenue crosses $20M over any consecutive 12 months, you owe Moonshot a separate agreement. And any product past 100M monthly active users or $20M in monthly revenue must display "Kimi K3" in its interface. For a solo shop that's a non-event; for anything with a real growth curve, it's a clause to read before it's a clause to litigate. We pulled it apart in the $20M line founders miss.
The through-line#
Three genuine milestones, three pieces of fine print, one lesson: the agent stack has matured past the point where the headline is the whole story. A finalized protocol comes with a deprecation contract. A frontier-adjacent open model comes with a harness and a license. The founders who win the next quarter aren't the ones who read the announcement fastest — they're the ones who read the paragraph after it.
