Meta Muse Spark 1.3
Meta's multimodal reasoning model for long-running agentic, multi-agent and coding workflows, with image input and a 1M-token context window.
Meta's Muse Spark 1.3 is a cloud text model. It is multimodal: alongside a text prompt it accepts images and video, then replies with generated text. Its context window handles up to 1.05M input tokens and up to 131K output tokens, and it supports adjustable reasoning effort for harder problems. Generation can be tuned with system instructions, temperature, and top-p sampling. It runs through OpenRouter using your own API key, from $1.25 per million input tokens.
- Released
- Sep 2026
- Pricing
- $1.25 / 1M in · $4.25 / 1M out
- Inputs
- Text, Images, Video
- Context window
- 1.05M in · 131K out
- Reasoning
- Adjustable effort
- Controls
- System prompt, Temperature, Top-p
- Text rank
- #10
- Code rank
- #25
- Vision rank
- #13
Samples
Generated with Muse Spark 1.3 via OpenRouter. The same three prompts run against every text model in the catalog, shown verbatim, so the only thing that changes between two models’ answers is the model.
Explain what an API is to someone who has never written code. Do not use a restaurant or waiter analogy. Plain prose only — no headings, bullet points, or markdown formatting. Maximum 50 words.
An API is a way for different software programs to talk to each other. It defines what requests one program can make, and what answers it will get back. For example, a weather app uses an API to ask a weather service for the current forecast.
A two-person startup is deciding whether to build their own user authentication or use a hosted provider. Give them a clear recommendation, then the single strongest argument against your own recommendation. Plain prose only — no headings, bullet points, or markdown formatting. Maximum 150 words.
Use a hosted provider like Auth0, Clerk or Supabase Auth and do not build your own. For a two person team, authentication is undifferentiated work that burns weeks on password hashing, session management, password resets, email verification, rate limiting, account recovery and OAuth edge cases while creating serious security and liability risk if you get it wrong. A provider gives you all of that in an afternoon and handles maintenance and compliance updates for you. The strongest argument against this is that you become dependent on a critical vendor for logins, pricing and uptime, so if they raise prices, suffer an outage or shut down, your users cannot access your product and migrating away under pressure is painful.
Write a briefing for a non-technical executive explaining why their company's spending on AI is hard to predict, and what they can do about it. Cover at least three distinct causes. Plain prose only — no headings, bullet points, or markdown formatting. Write exactly 205 words.
Your AI bill feels unpredictable because you are not buying software in the traditional sense, you are paying for usage that shifts with behavior, experimentation, and market prices. First, costs scale with activity rather than seats. Every question asked, document summarized, or image created consumes computing power, so a successful pilot that gets popular across teams can multiply expenses overnight without anyone changing a contract. Second, development is exploratory. Teams test many approaches, retry failed prompts, and run evaluations to improve quality, and much of that trial and error is invisible yet billable. A project that looks finished can require costly retuning when customer needs change. Third, prices and technology keep moving. Providers adjust rates, release new models, and retire old ones, meaning the efficient choice last quarter may be wasteful today, while data storage and safety checks add background costs. To regain control, treat AI like a utility budget with guardrails rather than a fixed purchase. Set spending limits per team, track usage by project so owners see their impact, require business approval before scaling pilots, and review vendors regularly to match each task to the most affordable capable model. With visibility, limits, and routine reviews, innovation can continue without surprise invoices undermining confidence.
Meta
Meta's FAIR and GenAI labs build the open-weight Llama family of language models — among the most widely used open models for local inference.
ai.meta.com ↗Meta's frontier reasoning model for complex personal and professional agentic tasks, with image input, tool calling and a 1M-token context window.
Meta's open-weight multimodal reasoning model (816B total, 42B active parameters) built for agentic coding, with image input and a 1M-token context window.
Meta's smallest and most efficient Muse model for the work agents repeat all day: tool calls, coding subtasks, routing, extraction and long document parsing, with image, video and audio input and a 1M-token context window.
Meta Superintelligence Labs' agentic image model, which plans a layout and calls search and coding tools before it renders, for prompt-faithful generation, multi-reference composition and precise local edits.
Meta's speech-to-text model for push-to-talk and speaker-aware transcription, which returns plain text.
Meta's 1B instruction-tuned model — the smallest Llama, runnable in-process via HuggingFace or through Ollama.