Muse Spark Open, or MW SPK: price, speed, and test results for games, scenes, and websites

Muse Spark Open, also identified in the materials as MW SPK, is Meta's multimodal model for code generation and agentic tasks. The available description from Meta specifies open weights, a mixture of experts, approximately 816 billion parameters, about 42 billion active parameters per token, and a context of up to 1 million tokens. The stated specifications and the results of the author's practical test are examined separately below.

What is known about Muse Spark Open and the name MW SPK

Meta's official description classifies Muse Spark Open as a model for programming, working with large codebases, and multistep tasks. The model accepts text, images, audio, and video. The available model materials also list inference stacks, including Transformers, llama.cpp, SGLang, vLLM, and MLX; this information is provided in the Hugging Face instructions rather than entirely in Meta's official description. File support in a specific interface depends on the provider, request format, and method used to pass content to the model.

The original materials use the designation MW SPK, as well as an inaccurate Russian spelling of the name. For searching for and discussing the model, it is reasonable to use the primary name Muse Spark Open and treat MW SPK as an alternative label for the tested version. The statement that the model was released only a few days before the test remains the author's claim and does not confirm the release date. The date 15 October 2026 is given in the available Hugging Face material and in the review author's post, but it is not confirmed as the release date in Meta's official description. The current availability of the specific version through OpenRouter should be checked separately.

Muse Spark Open pricing and the cost of the practical tests

In the test material, the author stated a price of 1,25 dollars per 1 million input tokens and 4,25 dollars per 1 million output tokens. These rates are not separately confirmed in the available evidence from the author, so they should be regarded as figures cited in the original review rather than verified current pricing. Even if such a rate applied, it would refer to the API access used: providers change prices, add their own markups, and count reasoning-related service tokens differently.

The amounts in the table show the cost of a specific run, not a fixed cost for an equivalent task. The final amount depends on the length of the request, the number of iterations, the reasoning level, repeated code generation, and the size of the response. When the model fixes errors in the project after checking it, the actual budget is higher than the initial amount, even if the first generation appears inexpensive.

How Muse Spark Open was tested

The author conducted the test through OpenRouter with the maximum reasoning level and used four scenarios: an HTML scene with an orc, an aquarium, a Mario-style game, and an interactive website design. This set tests not only the ability to write HTML, CSS, and JavaScript, but also the model's stability with visual objects, game mechanics, animation, interface structure, and data consistency.

The comparison is not a standardized benchmark. The result is affected by the prompt wording, the selected provider, launch parameters, waiting time, the local server, and the evaluation criteria. The author himself also expressed doubts about the reliability of the numerical figures, so the time and cost should be viewed as observations from a single test scenario rather than universal indicators of Muse Spark Open.

Muse Spark Open test on an HTML scene with an orc

The orc scene was the most successful episode of the test. According to the author's observation, Muse Spark Open prepared the HTML page in approximately 1 minute 22 seconds, at a cost of about 9 cents. The character was recognizable, and the animation looked convincing for a quick prototype, so the author rated the result as very good given the price.

Visual accuracy was still limited. The scene contained errors in the hands, the number of fingers, and certain character details, which is typical of generating complex graphics through code without a complete set of prepared assets. This result is suitable for a demonstration screen, a rough prototype, or a basis for further editing, but it does not replace manual checking of geometry, animations, and element behavior.

Aquarium: object logic is present, but visual accuracy is lacking

The model assembled the aquarium in 5 minutes 22 seconds. The author noted that the crab moved normally and the snails looked acceptable. The fish and some objects retained the specified behavior logic, but the plants and castle were poorly developed, while the fish resembled simple builder-style objects.

There were noticeable distortions on the glass or water that impaired the perception of the scene. The author rated the aquarium as significantly inferior to top models and specifically noted that the result was worse than the Gemini version in his comparison. This conclusion applies to the specific test, not to a universal model ranking: a different prompt, Gemini version, or provider could have changed the balance.

Mario-style game: basic mechanics did not become a finished prototype

The game scenario took 5 minutes 49 seconds and cost approximately 52 cents. The generated project ran on a local server and included a character, a mushroom, enemies, fireballs, and a secret level. The presence of these elements shows that in the presented test, the model quickly assembled a game framework and connected several mechanics in a single JavaScript project.

The author rated the implementation quality as very low. The character and locations were poorly developed, enemies and other objects often looked like squares, and animations and collisions worked inconsistently. After being hit, an object disappeared instead of showing the expected reaction, crouching did not work, the character fell through textures, and the turtle could only be kicked from above. The level logic was partially present, but the graphics, physics, and consistency of the controls did not reach the level of a finished game prototype.

For a developer, this is a significant limitation. In this scenario, Muse Spark Open saves time when creating the initial project structure, but it does not eliminate the need to debug collisions, check character states, and manually adjust visual assets. If the goal is to quickly obtain a project foundation, the result is useful. When a demonstration product is needed, it will require substantial further development.

Interactive website: fast assembly, weak data validation

The website design appeared in approximately 1 minute 33 seconds at a cost of about 10 cents. The project looked like an interactive dashboard, and the author liked the overall concept, fonts, and color scheme. The interface included a site map that visually resembled a radar, and a responsive version was provided for.

Content and functional testing revealed several problems. The team description listed five people and one dog, but the author counted only three people and did not find the dog. The sound did not work, responsiveness remained primitive, and the purpose of some elements was unclear. The interface looked overloaded, so the appealing concept did not compensate for the weak content and discrepancies between the stated and displayed data.

The website demonstrates a strength of the model within this test: Muse Spark Open quickly created a visual shell and connected the elements into an interactive screen. The test also highlights the need to manually check content, numbers, team composition, multimedia availability, and the mobile version. For an SEO specialist, this means that a generated dashboard cannot be published without checking the text, page structure, semantic markup, and factual data.

Muse Spark Open's strengths and weaknesses based on the test results

• Speed. The orc and website were assembled in about a minute and a half, while the aquarium and game appeared in less than six minutes.
• The cost of a short run. Simple HTML scenarios in the test cost about 9–10 cents, although the final amount increases with repeated fixes.
• Multimodality. Meta's official description specifies support for text, images, audio, and video.
• Fast edits. According to the author's assessment, the model is convenient for making simple changes to an existing project.
• Unstable visual detail. The aquarium and game scene showed weak development of objects, animations, and environments.
• Content validation errors. The website contained discrepancies between the stated team composition and what was displayed in the interface.

Overall assessment of Muse Spark Open

In the comparison conducted by the author, Muse Spark Open performed well on the orc task, handled the aquarium moderately, performed poorly on the game scenario, and received a neutral assessment for the website. The author considers processing speed and the ability to quickly make simple code changes to be the model's main advantages. The model's practical value is higher when a rough HTML interface or a starting version of a project is needed, rather than finished graphics and flawless game mechanics.

The author also believes that the smaller Haiku and Luna models might handle the same tasks better, while Muse Spark Open appears expensive compared with newer alternatives. This opinion is based on a specific comparison and does not replace a repeat test with identical prompts, settings, and providers. The claimed level, comparable to K3 and several Chinese models, should also be regarded as the author's assessment rather than official confirmation of equal capabilities.

What is confirmed and what remains an assessment

FAQ about Muse Spark Open

Is Muse Spark Open suitable for quickly generating websites?

Yes. According to the test results, the model quickly creates a visual shell and an interactive dashboard. The result cannot be published without checking it: the author found data discrepancies, nonworking sound, weak responsiveness, and unclear interface elements.

Can Muse Spark Open be used for games?

The model can assemble a game framework with a character, enemies, items, and separate levels. The tested scenario revealed problems with collisions, animations, controls, and graphics, so it is more reasonable to view Muse Spark Open as a tool for rough assembly rather than a means of releasing a finished game.

Is Muse Spark Open's price justified?

The answer depends on the task. For short HTML prototypes, the cited costs appear moderate, but the game in the test cost noticeably more and would still have required substantial further development. The cost of one response does not reflect the cost of a complete project, because fixes and repeated runs increase token consumption.

How objective are the comparison results?

This is the author's independent observation, not a standardized study. Time, cost, and scores depend on the prompt, OpenRouter provider, reasoning level, model version, and evaluation criteria, so the results are useful for preliminary selection but do not provide a universal ranking of Muse Spark Open.