Breaking

Meta Releases Muse Spark 1.3 for Leaner Long-Horizon Coding and Agents

September 2, 2026 at 17:25 EDT

Foundation Models
Software Dev & Coding
AI Agents

Meta Superintelligence Labs released Muse Spark 1.3 on September 2, 2026, calling it its largest gain yet on coding and agentic work. The model is live the same day in Muse Code and the Meta Model API, and it has also been added to Design Arena, a public comparison arena for real-world generative tasks.

September 2, 2026 · Meta

Muse Spark 1.3, leaner on long jobs

Meta Superintelligence Labs released a coding and agent update that, versus Muse Spark 1.2, uses about 20% fewer tool calls and 25% fewer tokens in internal tests.

~20%

Fewer tool calls vs 1.2

~25%

Fewer tokens vs 1.2

Intelligence Index, xhigh

Token context window

Relative usage vs Muse Spark 1.2

Shorter columns mean less work to finish the same internal comparisons. Vertical stacks, same 100 baseline.

100

1.2 tools

1.3 tools

100

1.2 tokens

1.3 tokens

Takeaway: Meta’s internal coding comparisons show the new model reaching for tools less often and emitting fewer tokens.

DeepSWE v1.1 (Meta scorecard, max)

Long-horizon agentic coding. Column height is proportional to the published score.

55.0

Muse Spark 1.2

75.4

Muse Spark 1.3 (max)

What developers like

Unchanged aggressive API rates, contributor-tier discounts, and a model that some 1.2 users said followed repo patterns instead of over-editing.

What remains contested

Headline benches use max reasoning, which is not generally live. Some testers still saw loops on bugs and asked how multi-workflow threads stay isolated.

Where it runs — and what is still gated

Live: Muse Code, Meta Model API, Design Arena; prior reasoning modes
Held: max reasoning, pending more safety testing (limited partner preview)
Next: bigger models and Muse Spark open weights, undated

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.