Tasks and inputs
Muse Glimmer 30B is listed for Image-text-to-text, Text generation, Coding & Code Generation. It accepts text, image and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
SyncDev model directory · Profile updated Oct 13, 2026
Muse Glimmer is Meta’s 30B dense multimodal model for local agents, with Apache 2.0 weights, image understanding and a default 128K context.
Pricing conditions: The recorded OpenRouter / DeepInfra route lists $0.30/M input, $0.04/M cached input and $1.20/M output; other partner routes differ. No Meta Model API rate applies. Partner endpoints impose their own output limits. Self-hosting uses your own hardware.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Muse Glimmer is a 30-billion-parameter dense model distilled from Muse Spark. It accepts text and images and generates text, making it a candidate for local assistants that need to inspect screenshots or other visual material. The default context is 128K tokens. Longer-context support is runtime-dependent and should be tested on the deployment you intend to use.
The downloadable weights use Apache License 2.0. Supported deployment options include vLLM, SGLang, llama.cpp and ExecuTorch. Check the exact format, quantization and vision support in the selected runtime. A compatible text server does not by itself establish that the image encoder and chat template are configured correctly.
Glimmer is not served through Meta Model API. Its partner ecosystem includes inference hosts and local tooling, but a partner route has separate access conditions and billing. Open weights remove a download fee; they do not remove memory, hardware or operating costs. Compare a complete local workload before committing to a serving configuration.
Test a representative image-and-text task with the exact weight format and runtime you plan to deploy. Check fine text, multiple images and ambiguous screenshots. Monitor memory during the longest expected conversation and verify that tool arguments follow the required schema. Keep the prompt template and quantization fixed when comparing deployment choices.
Record the exact model identifier, deployment route and request settings with each result. Count complete runs, including rejected outputs, retries and tool calls. When the workflow creates files or code, open the files or execute the existing checks before accepting the result. A plausible first response can hide an incomplete final deliverable.
The specifications describe supported interfaces and request capacity. They do not establish a measured quality ranking or a latency guarantee for your account. SyncDev has not independently benchmarked this model. Choose a deployment after testing the languages, input formats and operating conditions that matter to your users.
Explore meta models, browse model tasks, or compare two models with the same acceptance criteria.
Compare the model on the work you need to complete.
Muse Glimmer 30B is listed for Image-text-to-text, Text generation, Coding & Code Generation. It accepts text, image and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Documented limits: 131.07K context. The listed route has starting rates of $0.30 input and $1.20 output per million tokens. Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
Gemini 4 Argon handles coding, reasoning and multimodal work with a shared 1M-token window and background long decode.
Meta
Meta’s Muse Wonder handles coding, computer use and sustained agent tasks through Meta Model API and Muse Code, with a 1M-token context.
Meta
Muse Wonder Full is Meta’s higher-capability Wonder variant for sustained coding and agent work, available through eligible Muse Code subscriptions.
Xiaomi
MiMo-V2.6-Distill-Qwen-9B is Xiaomi’s smaller downloadable checkpoint for agent research and reinforcement-learning experiments.
Sakana AI
Fugu Max from Sakana AI: Multi-agent model for routing expert agents across complex analytical tasks.
Sakana AI
Fugu Ultra v2 from Sakana AI: Quality-first multi-agent model for hard research, analysis, and competitions.