Strengths
- A truly permissive Apache 2.0 license with no EU restrictions and no revenue cap, unlike Llama 4 and Muse Spark Open.
- Runs on a single consumer graphics card: at 4 bits the language model is under 20 GB. Meta says that on an RTX 5090 with DFlash speculative decoding it writes up to 233 tokens per second.
- In Meta's table it leads similar-sized models on tool use: MCP Atlas 75.5 versus 54.2 for Gemma 4 31B and 62.5 for Qwen3.6-27B.
- Its knowledge runs to January 2026, much later than Llama 4 (August 2024).
Weaknesses
- Even in Meta's own table it trails Qwen3.6-27B on several tests: OSWorld-Verified 65.9 versus 75.6, Terminal-Bench 2.1 51.7 versus 60.7.
- It does not understand audio, handles video only as individual frames, and its default 128K window is much smaller than the 1 million of Meta's hosted models.
- Full precision needs more than 55 GB of memory.
Best for
- a private assistant on your own computer, even offline
- a coding agent on a workstation
- fine-tuning for your own field
- EU companies that need a clean open license
Not ideal for
- understanding audio and video
- the hardest tasks that need a top-tier model
How to use it in Europe
You can download the weights for free from Hugging Face and run them with Ollama, LM Studio or llama.cpp on your own computer. The Apache 2.0 license has no geographic limits, so people and companies in the EU can use it freely. When it runs on your machine, your text goes nowhere. Without your own hardware, providers such as DeepInfra and Together host it.
Price compared with similar models
Benchmark results
| Benchmark | Score | Source |
|---|---|---|
| SWE-Bench Verified (high reasoning) | 76.0% | huggingface.co ↗, August 2026 |
| MCP Atlas (Public) | 75.5 | huggingface.co ↗, August 2026 |
| AIME 2026 | 94.7 | huggingface.co ↗, August 2026 |
| GPQA Diamond (AA) | 83.5 | huggingface.co ↗, August 2026 |
Benchmarks cover only a narrow slice of ability and vendors often report them at their best settings. Treat them as a rough guide.
More models from this vendor
- Muse Wonder: Meta's strongest model for long-running agents and coding, still released as a preview.
- Muse Spark 1.3: Meta's balanced model with a 1-million-token window and a fair price. The Muse Spark line also powers the Meta AI assistant.
- Muse Spark Open: Meta's large open model that a company can run on its own servers, so data never leaves its infrastructure.
- Llama 4 Maverick: The larger model in the newest Llama generation, from April 2025. Cheap through providers, but its license excludes people and companies in the EU.
- Llama 4 Scout: The smaller Llama 4 from April 2025, very cheap through providers. Its license grants no rights to people or companies in the EU.
Similar models from other vendors
- Gemma 4 31B (Google): Google's largest open model, which you can run on your own computer or server when data must not leave the company.
- Gemma 4 26B A4B (Google): Google's faster open model for your own server. Close to the larger Gemma 4 31B in quality, but quicker to respond.
- Ministral 3 14B (Mistral AI): Mistral's small open model under Apache 2.0 that runs on a single GPU, even without an internet connection.
- Nemotron 3.5 Bolt 74B-A9B (NVIDIA): NVIDIA's open model for AI agents on your own powerful PC, small enough to fit in 24 GB of graphics card memory.
- GPT-6 Luna (OpenAI): OpenAI's cheapest model, made for high volumes of simple tasks, sorting text and low-cost chatbots.
- gpt-oss-20b (OpenAI): OpenAI's smaller open model that runs on a computer with 16 GB of memory, even without internet access.