fifth August 2026 – Hyperlink Weblog
Introducing Muse Code and Muse Spark 1.2 (by way of) But extra proof that an important attribute of any mannequin lately is long-sequence agentic device calling. Meta shipped their very own coding agent as a part of getting that to work!
Muse Spark 1.2 is a coding-focused replace to Muse Spark 1.1, with enhancements in code technology, advanced debugging, codebase understanding, and end-to-end developer workflows. In Muse Spark 1.2, we considerably scaled up coaching compute on coding duties whereas increasing coaching surroundings range. The mannequin additionally maintains its energy in different key areas like common brokers. […]
We co-trained Muse Spark 1.2 with Muse Code to make sure the mannequin reveals its finest efficiency and coding usability when paired collectively. The coaching included rejection sampled harness trajectories and recipe optimizations for targets, compaction, and subagents, alongside the mixing of the Muse Code toolset to maximise harness compatibility. […]
Muse Spark 1.2 was extensively skilled on long-horizon coding duties, together with whole-repository technology, massive end-to-end initiatives, and auto-research.
This is a pelican driving a bicycle SVG produced by Muse Spark 1.2:
You’ll be able to see the Spark 1.1 pelican from ninth July right here. I believe the 1.2 pelican is a small however materials enchancment.
An attention-grabbing twist on pricing is that the mannequin is obtainable as two totally different mannequin IDs. muse-spark-1.2 is priced at $1.25/million enter and $4.25/million output – near Gemini 3.6 Flash ($1.50/$7.50) – however if you happen to comply with let Meta use your knowledge “to enhance our merchandise” you should use muse-spark-1.2-contributor which is $0.10/$0.20 – an enormous low cost, nearer to GPT-5.6 Luna ($0.20/$1.20) and Gemini 3.1 Flash-Lite ($0.25/$1.50).
I added these new costs to llm-prices.com.
