Article · 2026-08-05 · 5 moments

Introducing Muse Code and Muse Spark 1.2

Introducing Muse Code and Muse Spark 1.2 Yet more evidence that the most important characteristic of any model these days is long-sequence agentic tool calling. Meta shipped their own coding agent as part of getting that to work! Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1, with improvements in code generation, complex debugging, codebase understanding, and end-to-end developer workflows. In Muse Spark 1.2, we significantly scaled up training compute on coding tasks while expan ✦ AI generated

Read original ↗

GUEST the author HOST Simon Willison

Fact◆

Muse Spark 1.2 is a coding-focused update with improvements in code generation, complex debugging, codebase understanding, and end-to-end developer workflows, achieved by scaling up training compute on coding tasks and expanding training environment diversity.

The author describes Muse Spark 1.2 as a coding-focused update that scales training compute on coding while maintaining strengths in general agents.

transcript

The author: Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1, with improvements in code generation, complex debugging, codebase understanding, and end-to-end developer workflows. In Muse Spark 1.2, we significantly scaled up training compute on coding tasks while expanding training environment diversity. The model also maintains its strength in other key areas like general agents.

Mechanism

Muse Spark 1.2 was co-trained with Muse Code to ensure best performance and coding usability when paired together.

The author explains that co-training Muse Spark 1.2 with Muse Code, including rejection-sampled harness trajectories and harness toolset integration, maximizes combined performance.

transcript

The author: We co-trained Muse Spark 1.2 with Muse Code to ensure the model exhibits its best performance and coding usability when paired together. The training included rejection sampled harness trajectories and recipe optimizations for goals, compaction, and subagents, alongside the integration of the Muse Code toolset to maximize harness compatibility.

Fact

Muse Spark 1.2 was extensively trained on long-horizon coding tasks including whole-repository generation, large end-to-end projects, and auto-research.

The author notes that Muse Spark 1.2's training included long-horizon coding tasks spanning whole-repository generation and auto-research.

transcript

The author: Muse Spark 1.2 was extensively trained on long-horizon coding tasks, including whole-repository generation, large end-to-end projects, and auto-research.

Claim◆

The most important characteristic of any model these days is long-sequence agentic tool calling.

The author asserts that long-sequence agentic tool calling is now the defining capability for models, citing Meta's own coding agent as evidence.

transcript

The author: Yet more evidence that the most important characteristic of any model these days is long-sequence agentic tool calling. Meta shipped their own coding agent as part of getting that to work!

gives example · 1

Example

The Muse Spark 1.2 pelican is a small but material improvement over the 1.1 pelican.

Comparing the SVG pelican outputs, the author judges the 1.2 pelican a small but material improvement over the 9th July 1.1 version.

transcript

The author: You can see the Spark 1.1 pelican from 9th July here. I think the 1.2 pelican is a small but material improvement.

Highlight slides

Long-Sequence Agentic Tool Calling Is the Defining Capability✦ from: The most important characteristic of any model these days is long-sequence agentic tool calling.Muse Spark 1.2: Coding-Focused Update✦ from: Muse Spark 1.2 is a coding-focused update with improvements in code generation, complex debugging, codebase understanding, and end-to-end developer workflows, achieved by scaling up training compute on coding tasks and expanding training environment diversity.Unchanged Strengths in General Agents✦ from: Muse Spark 1.2 is a coding-focused update with improvements in code generation, complex debugging, codebase understanding, and end-to-end developer workflows, achieved by scaling up training compute on coding tasks and expanding training environment diversity.

Related episodes

Video · a16z Podcast

Muse Spark 1.2 is a coding-focused update with improvements in code generation, complex debugging, codebase understanding, and end-to-end developer workflows, achieved by scaling up training compute on coding tasks and expanding training environment diversity.

Muse Spark 1.2 was co-trained with Muse Code to ensure best performance and coding usability when paired together.

Muse Spark 1.2 was extensively trained on long-horizon coding tasks including whole-repository generation, large end-to-end projects, and auto-research.

The most important characteristic of any model these days is long-sequence agentic tool calling.

The Muse Spark 1.2 pelican is a small but material improvement over the 1.1 pelican.

Is Defense the Next Trillion-Dollar Category? | a16z American Dynamism Summit

I Built an LLM From Scratch

Exploring the J-Space