Meta says its model scored 30/30 on Asian Physics Olympiad theory exam

• Meta said a model scored 30/30 on APhO 2026's theory exam, with AIatMeta's announcement saying the score tied the top three student contestants.
• The attached committee graphic says the model took the May 19 exam under authentic conditions and scored 10/10 on all three questions; AIatMeta's announcement carries that document.
• The evaluated checkpoint is not cleanly named: alexandr_wang's post called it Muse Spark, shuchaobi's post called it Avocado, and shengjia_zhao's post called it an internal Muse Spark.
• Team posts describe a competition pipeline, with shuchaobi's post listing data curation, training, debugging, and live participation.
• A public developer surface exists separately, with openclaw's post saying Muse Spark 1.1 is live in OpenClaw for agentic coding, tool use, and computer-use workflows.

APhO's statutes say the theory contest normally means three problems in a five-hour session, while Busan's event notice framed the 2026 event as 27 countries competing in theoretical and experimental physics exams. The score graphic attached to AIatMeta's announcement says this was the first AI model in the APhO Theoretical Examination. The messy part is identity, because the public posts bounce between alexandr_wang's “Muse Spark” wording, shuchaobi's “Avocado” wording, and shengjia_zhao's “internal muse spark” wording.

The score graphic attached to AIatMeta's announcement gives the useful details:

• Exam date: May 19, 2026.
• Conditions: “authentic conditions,” according to the APhO 2026 Organizing Committee graphic.
• Total score: 30.00 / 30.00.
• Question scores: Q1 10.00, Q2 10.00, Q3 10.00.
• Ranking: tied first with three student contestants.
• Committee framing: first AI model participation in the APhO Theoretical Examination.

The score is narrower than a full Olympiad result. The public evidence covers the theoretical exam, and the screenshot attached to AIatMeta's announcement reports no experimental score.

The naming trail is the main caveat. AIatMeta used the generic phrase “a model” in its announcement, while alexandr_wang's post called the system Muse Spark.

Other team posts used different wording:

• TrapitBansal's post called it “our model.”
• shuchaobi's post called it the “Avocado model.”
• shengjia_zhao's post called it an “internal muse spark.”
• eliebakouch's question asked whether the result came from Muse Spark 1.1 or an internal model.

Meta's Muse Spark developer post separately says Muse Spark is available in public preview on Meta Model API for US developers, with self-serve access, OpenAI SDK compatibility, and $20 in free credits. The public APhO posts do not settle whether the Olympiad checkpoint was the same public Muse Spark 1.1 build, a private variant, or a differently named internal system.

shuchaobi described the APhO run as a full competition effort, “from data curation, model training, debugging to live participation.” That is the most useful sentence in the thread for engineers trying to interpret the score.

TrapitBansal framed APhO 2026 as an eval of multimodal and reasoning capabilities “as we continue to scale.” shuchaobi went further, calling the result an “independent and un-hackable demonstration” of the model's reasoning and multimodal capabilities in the same post.

According to APhO's statutes, the competition runs theory and experiment on separate days, with at least one rest day between them. The statutes say problem-solving time should normally be five hours, with three theoretical problems and one or two experimental problems.

The rules also constrain the tool environment. Contestants may use logarithm tables, physical constants, non-programmable calculators, and drawing materials, but not collections of math or physics formulae, according to the same APhO statutes.

Busan's event notice said the 2026 Busan contest ran May 17 to 25 and brought students from 27 Asian countries to compete in advanced theoretical and experimental physics examinations.

openclaw said Muse Spark 1.1 is live in OpenClaw v2026.7.1. The post describes the public model surface as aimed at agentic coding, tool use, and computer-use workflows.

That sits next to Meta's official developer framing, which says most agent CLIs are OpenAI-compatible and can connect to Meta Model API quickly. The APhO score is the shiny science eval; the immediate engineering surface is a reasoning model being wired into agent tooling.