Muse Spark 1.3 vs Gemini 3.8 Flash: Which Wins on Coding Tasks?
Benchmark testing across eight coding, 3D, and agentic tasks shows Gemini 3.8 Flash surging while Muse Spark 1.3 regresses from 1.2.
Gemini 3.8 Flash beats Muse Spark 1.3 on coding benchmarks, scoring 81.25% versus 71.25% on an eight-task test covering front-end, 3D, SVG, math, and agentic work. Gemini’s score marks a jump of more than 51 points over its predecessor, Gemini 3.5 Flash, while Muse Spark 1.3 actually dropped five points below Muse Spark 1.2, its own prior version.
TL;DR
- Gemini 3.8 Flash scored 65 out of 80 (81.25%) on an eight-task coding benchmark, landing it in the top five of the tested leaderboard, tied with Qwen 3.8 Max and above Opus 4.8.
- Muse Spark 1.3 scored 57 out of 80 (71.25%), a regression from Muse Spark 1.2’s 76.25%, despite being a point release expected to improve on its predecessor.
- The regression is uneven: Muse Spark 1.3 set a new record on a 3D wristwatch task and improved on an elevator simulation, but fell hard on SVG generation and a 3D folding table, the exact tests where 1.2 had excelled.
- Both models tied on a hard math permutation problem (both landed on the correct answer, 2460) and on a fully agentic task involving fine-tuning a Gemma 2B model and serving results through a local web UI.
- Both models share a file-overwriting problem: asking for a small change often triggers a full file rewrite instead of a targeted edit, which is risky in real codebases and forces developers to check every diff.
- Gemini 3.8 Flash is available through Google’s Antigravity platform, where its free tier and fast inference make it a strong pick for research and lighter coding work.
- Both models can be run side by side through Verdant, an agentic coding workspace that supports parallel isolated agents, plan mode, and review steps that help catch the overwriting problem before it causes damage.
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
How were the two models tested?
The comparison used KingBench 3, a fixed set of eight coding and reasoning tasks scored out of 10 points each, for a maximum of 80. The tasks span front-end simulation (an elevator system with queuing logic), two Three.js 3D builds (a contact lens case and a folding table), an SVG generation prompt (a panda eating a burger), a browser-based game (a bow and arrow simulator with a leaderboard), a hard math problem (permutation counting), a fully agentic pipeline (generate a dataset, fine-tune a Gemma 2B model, and serve results through a local UI with no setup), and a 3D wristwatch with live time, date, and dual time zone tracking.
This structure matters because it separates raw code generation from visual polish and from multi-step autonomous execution, three skills that don’t always move together. A model can nail the math and still fumble the SVG, which is exactly what happened here.
Why did Gemini 3.8 Flash improve so much?
The jump from Gemini 3.5 Flash (30% on the same benchmark) to Gemini 3.8 Flash (81.25%) is the standout result. That’s a gain of more than 51 percentage points in one generation, taking a model that had one of the lowest scores on the leaderboard to a top-five finish, tied with Qwen 3.8 Max and sitting just above Opus 4.8.
The gains showed up most clearly on tasks that punish sloppy execution. Gemini 3.8 Flash scored a full 10 on the bow and arrow game, where aiming, physics, and leaderboard logic all had to work together, and a 9 on the 3D contact lens case, where the earlier model class would typically produce janky animations. It also handled the SVG panda test well, producing a recognizable panda actually eating a recognizable burger, a detail that trips up many models.
Flash-tier models are generally optimized for speed and cost rather than peak capability, so a flash-class model reaching top-five territory on a coding benchmark is a meaningful signal that the gap between “fast and cheap” and “frontier-capable” is narrowing.
Why did Muse Spark 1.3 score worse than 1.2?
Muse Spark 1.2 had been a standout release, jumping from a 60% baseline (version 1.1) to 76.25%, driven largely by strong visual and front-end output. Its SVG panda test scored a perfect 10, and its 3D folding table scored an 8. Those two tasks were widely considered the model’s signature strength.
Version 1.3 lost ground on both. The SVG panda dropped to a 5, with rough shapes and a burger that barely reads as a burger. The folding table also dropped to a 5, with geometry breaking mid-animation. At the same time, 1.3 improved on tasks 1.2 struggled with: the elevator simulation went from a 5 to a 7, and the 3D wristwatch task, one of the hardest in the set, went from a 6 to an 8, the highest score recorded on that task across tested models.
One coffee. One working app.
You bring the idea. Remy manages the project.
The net effect is a trade rather than a clean decline. Meta’s model got better at some logic-heavy tasks and worse at visual generation, and the visual regression outweighed the logic gains in the final score. In hands-on use beyond the benchmark, Muse Spark 1.3 also showed weaker performance on long-running agentic tasks, drifting from the original plan and declaring work finished when it wasn’t, a problem that was less pronounced in 1.2.
Do both models still overwrite files?
Yes, and this is a shared weakness worth flagging for anyone using either model in a real codebase. Both Muse Spark 1.3 and Gemini 3.8 Flash tend to rewrite an entire file in response to a small, targeted change request instead of editing just the relevant lines. Muse Spark carried this issue over from version 1.2, but Gemini 3.8 Flash appears to have newly picked up the same habit, something not as noticeable in earlier Gemini Flash releases.
In practice, this means small commits and careful diff review are necessary safeguards. An unrelated function or comment can get quietly altered or removed during what should have been a one-line fix. Tools that wrap model output in a plan-then-verify workflow, generating a plan before execution and running a review pass afterward, help catch this kind of unintended change before it merges.
Is Gemini 3.8 Flash worth using for coding?
Based on the benchmark results, yes, particularly for cost-conscious or speed-sensitive workloads. A top-five score among tested models at flash-tier pricing and latency is a strong combination. Gemini 3.8 Flash is available through Google’s Antigravity platform, which offers a generous free tier and fast response times, making it a reasonable default for students, researchers, or anyone doing lighter coding work without wanting to pay frontier-model prices.
The caveat is the file-overwriting behavior. It doesn’t disqualify the model, but it does mean it shouldn’t be trusted to make surgical edits in a shared or production codebase without review.
Is Muse Spark 1.3 worth using for coding?
It depends on what you need. Muse Spark 1.3 can still do things few other models manage, the record-setting wristwatch result is real evidence of that, but the version-over-version regression is hard to ignore. Anyone specifically drawn to Muse Spark for its visual and front-end generation strength, the reason 1.2 stood out in the first place, will likely be disappointed by 1.3’s SVG and 3D animation output. Combined with weaker long-running task performance and the unresolved file-overwriting issue, 1.3 reads as a step sideways rather than forward. Whether it’s worth adopting depends on whether your use case leans toward the logic-heavy tasks it improved on or the visual tasks it regressed on.
Frequently Asked Questions
What is KingBench?
KingBench is a benchmark consisting of eight coding and reasoning tasks, each scored out of 10, covering front-end simulation, 3D graphics (via Three.js), SVG generation, math reasoning, and fully autonomous agentic workflows. It’s used to compare how different AI models handle both visual polish and functional correctness in code.
How much did Gemini 3.8 Flash improve over Gemini 3.5 Flash?
Gemini 3.5 Flash scored 30% on the benchmark, one of the lowest results recorded. Gemini 3.8 Flash scored 81.25%, a gain of more than 51 percentage points in a single generation.
Is Muse Spark 1.3 better or worse than Muse Spark 1.2?
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
Worse overall on this benchmark. Muse Spark 1.2 scored 76.25%, while Muse Spark 1.3 scored 71.25%. The newer version improved on some tasks, including setting a new record on a 3D wristwatch test, but regressed sharply on SVG generation and 3D animation, the areas that made 1.2 stand out.
Do Muse Spark 1.3 and Gemini 3.8 Flash have any shared weaknesses?
Yes. Both models tend to overwrite entire files when asked to make small, targeted code changes rather than editing only the relevant section. This makes diff review important when using either model in a real project.
Where can I use these models?
Gemini 3.8 Flash is available through Google’s Antigravity platform, which offers a free tier well suited to research and lighter coding tasks. Both Gemini 3.8 Flash and Muse Spark 1.3 are also accessible through Verdant, an agentic coding workspace with parallel agent orchestration, plan mode, and review agents.
