Muse SparkMeta AIAI CodingAI AgentsNews

Muse Spark 1.2: What Changed in Meta's Coding Model

Meta describes Muse Spark 1.2 as a coding-focused update co-trained with Muse Code, with long-horizon task training, context compaction, and a published kernel optimization case study.

Muse Spark 1.2: What Changed in Meta's Coding Model
Analyzed Foundation Models (3)All Models Directory→

Meta announced Muse Spark 1.2 on August 5, 2026, alongside Muse Code beta. Meta describes it as a coding-focused update to Muse Spark 1.1, with more training compute and a wider range of coding environments. The company says it co-trained the model with Muse Code so the model works well with that harness.

What Meta says changed

Meta highlights code generation, complex debugging, codebase understanding, and end-to-end developer workflows. The training included Muse Code trajectories and optimizations for goals, context compaction, and subagents. Meta also says Spark 1.2 was trained on long-horizon coding tasks, including whole-repository generation and large end-to-end projects.

The release describes a kernel-optimization case study in Muse Code: more than 1,000 tool calls over up to 24 hours on KDA and MLA kernels for NVIDIA Hopper GPUs. This is a vendor demonstration under a specific harness and workload, not a general guarantee that the model can run unattended for a day or improve arbitrary GPU kernels.

The co-training is relevant because a coding model’s results depend partly on the harness around it. Meta says it included Muse Code toolset integration and sampled harness trajectories in training. This is intended to improve compatibility with that runtime; it should not be read as proof that the model will work equally well with every agent framework.

Evaluations

Meta links to its evaluation report and publishes comparisons for Terminal-Bench 2.1, DeepSWE 1.1, and an internal coding benchmark. Read the report for benchmark versions, harnesses, tools, and attempt budgets before comparing the scores. Results from an internal coding benchmark are not directly comparable with a public benchmark.

Do not confuse DeepSWE with SWE-bench Verified. They are distinct evaluations, and a result on one does not translate directly to the other.

For a practical trial, choose a small repository task with a clear expected result, save the starting commit, and record which model release, agent version, tools, and checks were used. Compare the resulting patch and runtime behavior with another model using the same conditions. Report whether the agent could run tests or browse the network, since these capabilities can materially change task outcomes.

Availability and practical use

At launch, Meta said Muse Spark 1.2 was available in Muse Code and the Meta Model API. The official announcement documents Muse Code for macOS and Linux and provides an installation command on its page. Check the current Meta developer documentation for account access, model availability, context limits, pricing, and commercial terms.

Muse Spark 1.2 is a version-specific example of model and coding harness co-training. Meta has since announced Muse Spark 1.3; its release describes improvements to longer-horizon tasks and instruction following. Evaluate the exact model and harness version you intend to use on representative repository tasks, inspect the generated diff, and run your own tests before relying on results. Read Meta’s Muse Code and Muse Spark 1.2 announcement and Muse Spark 1.3 update for release context.

Frequently Asked Questions

What is Muse Spark 1.2?

Meta describes Muse Spark 1.2 as a coding-focused update to Muse Spark 1.1, co-trained with the Muse Code terminal agent for coding workflows.

How much does Muse Spark 1.2 cost?

The launch announcement does not list per-token prices. Check Meta's current Model API documentation for access and pricing.

What is the context window of Muse Spark 1.2?

Meta's Muse Spark 1.2 announcement describes context compaction for long-horizon coding but does not specify a context-window size.

What benchmarks did Muse Spark 1.2 achieve?

Meta links to evaluations on Terminal-Bench 2.1, DeepSWE 1.1, and an internal coding benchmark. It also describes a kernel-optimization case study using more than 1,000 tool calls over up to 24 hours.

Sources & disclosures

Source links for data and claims discussed in this article are listed below. Refer to each source for its current figures and methodology.

Share Article

Recommended Reading

Related by topic and recency.

All articles→
GuidesDeepSeek

DeepSeek Harness: Architecture, Tool Execution, and Setup Guide

A practical overview of DeepSeek Harness, its plugin-based architecture, developer-preview status, official quickstart commands, and precautions for running an agent.

Aug 17, 2026 · 2 min read
ComparisonsDeepSeek V4 Pro

DeepSeek V4 Pro 0813 vs GLM-5.3: Context and API Pricing

Compare DeepSeek V4 Pro 0813 and GLM-5.3 context limits, current native API pricing, and a reproducible way to evaluate coding tasks.

Aug 17, 2026 · 2 min read
NewsGLM-5.3

GLM-5.3 in ZCode: Agentic Coding Integration and Evaluation

Z.ai describes GLM-5.3 as a post-training update to GLM-5.2 and says it rolled out to GLM Coding Plan users. Learn how to evaluate the model separately from the ZCode agent.

Aug 17, 2026 · 2 min read