Meta released Muse Spark 1.2 on August 5, 2026 - a coding-focused update shipped alongside Muse Code, its first terminal coding agent for macOS and Linux. The headline Terminal-Bench 2.1 score is 82.9%. That number will circulate for weeks. It deserves more than a glance before you act on it.
What Muse Spark 1.2 actually ships
Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1, released alongside Muse Code - a terminal tool built around persistent background agents and parallel sub-agents in isolated git worktrees.
Reasoning is mandatory across five effort levels - minimal, low, medium, high, and xhigh - with medium as the default; there is no configuration where the model skips deliberation.
The context window sits at 1 million tokens.
Muse Spark 1.2 reports two tracked benchmark scores at release by Meta: DeepSWE 1.1 at 59.3% and Terminal-Bench 2.1 at 82.9%. Those are the numbers you will see in most coverage. Here is what that coverage tends to omit.
Meta co-trained the model and the Muse Code agent together - the same co-training play that produced Claude Code and Codex. That is not cheating; it is how agentic evaluation now works. It does mean a Muse Spark 1.2 score obtained through some other scaffold is an open question rather than a safe assumption.
On Terminal-Bench v2.1, independent evaluation puts Muse Spark at 80% - compared to Meta's reported 82.9% - competitive but below Claude Opus 5 either way. And on SciCode, Spark 1.2 regressed versus 1.1, making scientific code generation its weaker flank.
The co-training problem is not specific to Meta
Every major coding agent release in the last six months has faced the same issue: the model's headline coding scores were produced inside the harness it was trained against. That is how agentic evaluation now works.
This matters because it changes what a benchmark score actually tells you. When a model and its agent scaffold are trained together, the benchmark score reflects a paired system, not a model you can swap freely into your stack. Muse Spark 1.2 runs on the Meta Model API and OpenRouter, but Meta co-trained the model and the agent together - the headline coding charts benchmark 1.2 inside Muse Code. If you already run a terminal AI coding tool, expect a different result.
The practical test is straightforward: Meta's benchmark results are company-reported and should be read with the evaluation setup in mind. Results for coding and tool-use models can change substantially with the harness, system prompt, tool definitions, sampling settings, and whether the measurement covers the model alone or the complete agent system. Run your own eval on your own scaffold with your own repo before changing anything.
Beagle in action#eng-platform, evaluating a model switchWhere Muse Spark 1.2 stands in the open-weight picture
The planned Muse Spark 1.2 open-weight release is the more important strategic signal. Muse Spark launched on April 8 as Meta Superintelligence Labs' first model and now powers Meta AI across the company's consumer products. Meta introduced Spark 1.1 in July with features for planning, connecting to applications, and completing multi-step tasks.
But the open-weight version does not exist yet. Meta has not published the Spark 1.2 weights, a final release date, a complete license, or the hardware specifications needed to run them. The announcement establishes a planned release, not a completed open-weight launch.
Chief AI Officer Alexandr Wang posted on August 10, 2026, that an open-weight version is coming "soon," alongside the Muse Glimmer 30B release under Apache 2.0.
Compare that to where the open-weight coding field currently sits:
| Model | SWE-Bench Pro | Open Weights | License |
|---|---|---|---|
| Muse Spark 1.2 | 59.3% (DeepSWE 1.1) | Not yet | TBD |
| Kimi K2.6 | 58.6% | Yes | Modified MIT |
| GLM-5.2 | - | Yes | MIT |
| DeepSeek V4 Flash | ~80.6% (V4 Pro) | Yes | Custom |
On SWE-Bench Pro, Kimi K2.6 scores 58.6 - ahead of GPT-5.4 (57.7%), Claude Opus 4.6 (53.4%), and Gemini 3.1 Pro (54.2%). It was the first open-weight model to credibly out-score those closed models on that benchmark.
The hallucination rate on AA-Omniscience fell from 65% on K2.5 to 39% on K2.6 - a calibration jump that matters more for production deployment than most top-line benchmark gains.
Muse Spark 1.2, whenever it does release its weights, enters a field that is no longer empty. The non-obvious thing here: Meta has moved 11 index points in four months and is now shipping faster than the evaluation ecosystem can keep up with - there is still no published per-benchmark breakdown for 1.2, and no Design Arena record for it at all, while 1.1 has both. Speed of release has outpaced independent verification.
Beagle in action#dev-tools, model evaluation threadTwo frontier releases in five weeks suggests the version you benchmark today may not be current for long. That is true of every model at this pace, but it is especially true of a model where the open-weight version has not shipped and the independent evaluation record is thin. Wait for the weights. Run it on your scaffold. Compare the gap.
Muse Spark 1.2 benchmarks: common questions
What score does Muse Spark 1.2 get on Terminal-Bench 2.1?
Meta reports 82.9% on Terminal-Bench 2.1; independent evaluation puts the figure at 80%. The difference reflects the co-training between Muse Spark 1.2 and the Muse Code scaffold used in Meta's benchmark run. If you use a different agent harness, plan for the lower number.
Is Muse Spark 1.2 an open-weight model?
Not yet. Meta has not published the Spark 1.2 weights, a final release date, a complete license, or the hardware specifications needed to run them. The announcement establishes a planned release, not a completed open-weight launch. It is available today only through Meta's API and Muse Code.
How does Muse Spark 1.2 compare to open-weight coding alternatives?
The closest open-weight competitors on agentic coding are Kimi K2.6 and GLM-5.2. Artificial Analysis ranks K2.6 number one among open-weight models. On SWE-Bench Pro it scores 58.6% - ahead of GPT-5.4 at 57.7%. Muse Spark 1.2 posts 59.3% on DeepSWE 1.1, a different benchmark, so a direct comparison requires running both on the same harness.
Did Muse Spark 1.2 improve over 1.1 across all benchmarks?
No. On SciCode, a coding benchmark, Muse Spark 1.2 went backwards compared to 1.1 , while the agentic index and GDPval Elo rose. The gains concentrate in terminal-agent tasks; scientific code generation regressed.
What is Muse Code and does it come with Muse Spark 1.2?
Muse Code is Meta's terminal coding agent, released in beta alongside Muse Spark 1.2, and it is optional. The catch is that the two were co-trained, so the headline benchmark scores reflect the pair. Using Muse Spark 1.2 without Muse Code - through the API or OpenRouter - will produce different results from what the release benchmarks report.