Muse Spark 1.1 beats GPT-5.6 Sol on RadLE

// 4h agoBENCHMARK RESULT

Muse Spark 1.1 beats GPT-5.6 Sol on RadLE

Meta's closed-source multimodal reasoning model, Muse Spark 1.1, has surpassed GPT-5.6 Sol and Gemini 3.1 on the "Radiology's Last Exam" (RadLE) benchmark. The benchmark evaluates frontier models against board-certified radiologists on complex diagnostic imaging cases, and while Muse Spark 1.1 has narrowed the gap to human expert baseline performance, it still trails human radiologists and the Fable model.

// ANALYSIS

Meta's focus on closed-source models with Muse Spark 1.1 and the Meta Model API highlights their intent to monetize specialized domain intelligence directly against OpenAI and Anthropic.

* Beating leading models like GPT-5.6 Sol and Gemini 3.1 in radiology reasoning illustrates the fast-paced development of Meta's multimodal architectures.

* The fact that human experts and Fable still lead suggests that high-stakes medical diagnosis remains a difficult frontier for general reasoning agents.

* Deep domain benchmarks are becoming the key battleground for evaluating true agentic capabilities beyond standard academic test sets.

// TAGS

muse-spark-1.1metabenchmarkradiologymedicinemultimodalreasoning

DISCOVERED

4h ago

2026-07-13

PUBLISHED

4h ago

2026-07-13

RELEVANCE

8/ 10

AUTHOR

_jasonwei

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

RESEARCH36m ago

A new study demonstrates that visual pretraining on raw visual documents bypasses OCR/parsing limitations and consistently outperforms traditional text-only pretraining for building language intelligence.

Web search paper title}

TUTORIAL1h ago

Tutorial runs MiniMax M3 inside Claude Code

A recent YouTube video explores how developers can integrate the MiniMax M3 model into Claude Code. MiniMax M3 is an open-weight mixture-of-experts (MoE) model that boasts a massive 1-million-token context window and strong performance on coding benchmarks, making it a viable alternative to Claude's native models for users hitting usage constraints.

NEWS2h ago

Tiny Army, Eyas win Build Small hackathon

Cohere co-sponsored Hugging Face's 'Build Small' hackathon, which challenged developers to create useful, whimsical, or cool applications using smaller, more efficient AI models. Two projects powered by Cohere's models received awards: 'Tiny Army,' an interactive game by @polats where players describe and create their own heroes, won second place on the Thousand-Token Wood track; and 'Eyas,' a security camera agent built by Hanhee Lee, Javier Huang, and Joe Lee to solve real-world security needs for a family convenience store, won the Best Agent award.