Llama (Meta)
Overview
Meta's model family runs on two tracks: the open-weight Llama line and the closed Muse line from Meta Superintelligence Labs. The frontier reasoning and agent slot belongs to Muse Spark 1.1 (closed, $1.25/$4.25 per 1M tokens), local execution to Muse Glimmer 30B (released August 10, 2026 under Apache 2.0), and large open-weight MoE work to Llama 4.
Current lineup: Muse Spark 1.1 / Muse Glimmer 30B / Llama 4
| Model | Type | Notes | Price |
|---|---|---|---|
| Muse Spark 1.1 | Closed (API) | Multimodal reasoning built for agentic work: 1M-token context, stronger tool use, computer use, and coding | $1.25 / $4.25 per 1M tokens |
| Muse Glimmer 30B | Open weights (Apache 2.0) | 30B distilled from Muse Spark via logit distillation; aimed at function calling, local coding, long tool-use sessions | Free (self-hosted) |
| Llama 4 Scout | Open weights | 17B active / 109B total, 10M-token context | Free (self-hosted) |
| Llama 4 Maverick | Open weights | 17B active / 400B total, 1M-token context | Free (self-hosted) |
Key capabilities
1. Muse Glimmer — an agent model that fits one consumer GPU
At 30B parameters it runs on a Mac or PC with a single 24GB-class GPU, works with or without a network connection, and targets agentic workloads: function calling, local coding, long tool-use sessions, and LLM-as-a-judge evaluation. Weights are on Hugging Face with optimized integrations for llama.cpp, MLX, and ExecuTorch.
2. Back to open weights, under Apache 2.0
Muse Glimmer ships under Apache 2.0 rather than the Llama Community License, so there is no MAU threshold to reason about before using it commercially.
3. Industry-leading context (Llama 4 Scout)
Scout carries a 10M-token context window and runs on a single H100 when quantized to Int4.
4. Natively multimodal MoE
Llama 4 was the first open-weight family with a natively multimodal MoE architecture, handling text, images, code, and many languages in one model.
5. Meta Model API (paid access to Muse Spark)
Consumer access at meta.ai stays free, while developers get Muse Spark through the Meta Model API in public preview. At $1.25/$4.25 the rate is a fraction of the major frontier APIs.
6. Plenty of hosting options
- AWS Bedrock
- Azure AI Foundry
- Google Cloud Vertex AI
- Groq (very fast inference)
- Fireworks AI
- Together AI
Typical uses
- Local agents: run Muse Glimmer on a laptop and let it execute tools with no network dependency
- On-prem / private LLM: keep confidential data inside your own infrastructure
- Long-document work: use the 10M context for large documents and whole codebases
- Custom fine-tuning: build domain-specific models
- Cost control: self-host instead of paying per token
Strengths
- ✅ Muse Glimmer 30B is Apache 2.0 and runs agents on a single local GPU
- ✅ 10M-token context (Llama 4 Scout, among the longest available)
- ✅ Efficient MoE — 400B total quality from 17B active parameters
- ✅ Natively multimodal (a first for open models)
- ✅ Muse Spark's API rate of $1.25/$4.25 undercuts the major frontier APIs by a wide margin
- ✅ Many hosting options
Weaknesses
- ❌ Self-hosting required (operational cost and infrastructure know-how)
- ❌ Still short of GPT-5.6 and Claude Opus 5 at the very top
- ❌ The frontier tier, Muse Spark, is closed — no published weights
- ❌ Limited Japanese optimization (English and code first)
- ❌ Llama models carry license conditions for commercial use (a separate license above 700M MAU)
Official links
- Website: https://llama.com
- Meta AI: https://ai.meta.com
- Downloads: https://llama.com/llama-downloads
- Hugging Face: https://huggingface.co/meta-llama