Meta currently offers two very different model lines: the Muse family from Meta Superintelligence Labs (MSL), with the proprietary models Muse Wonder (preview) and Muse Spark and the open variants Muse Spark Open and Muse Glimmer 30B, and the earlier Llama line, which Meta ended in April 2026. For enterprises in the DACH region the licence question is decisive: Muse Glimmer is released under Apache 2.0 and Muse Spark Open under the Muse Model License without a territorial exclusion; both can be self-hosted in the EU. Llama 4 is blocked for EU companies by its licence, and Meta documents no EU data residency for Muse Wonder and Muse Spark via the Meta Model API.
innFactory AI Consulting from Rosenheim supports companies with GDPR-compliant self-hosting of Meta models – with open weights, full control stays with you and no data leaves your infrastructure. This overview reflects the state as of October 2026.
Current models at a glance
| Model | Release | Licence | Context | EU use | Focus |
|---|---|---|---|---|---|
| Muse Glimmer 30B | 10 August 2026 | Apache 2.0 (open weights) | 128k (per model card) | Self-hosting unrestricted | Local agentic workflows, tool calling |
| Muse Wonder / Wonder Full | 12 October 2026 (preview) | Closed weights, API or Muse Code | 1M | No EU data residency documented | Complex software engineering, long-running agents |
| Muse Spark Open | 15 October 2026 | Muse Model License (open weights) | 1M | Self-hosting possible (816B / 42B active) | Large self-hosting setups, post-training |
| Muse Spark 1.3 | 2 September 2026 | Closed weights, API | 1M | No EU data residency documented | Agents, coding |
| Llama 3.3 70B | December 2024 | Llama Community License | 128k | Permitted | Classic text tasks |
| Llama 4 Scout / Maverick | April 2025 | Llama 4 Community License | 10M / 1M | EU use excluded | – |
Muse Wonder and Muse Wonder Full
On 12 October 2026 the Meta Superintelligence Labs introduced Muse Wonder and Muse Wonder Full. Per Meta’s developer documentation, Muse Wonder is the most capable model of the Muse family, built for complex software engineering and long-horizon agentic work: it makes and verifies changes across many files in real repositories, plans, delegates to sub-agents, calls tools in parallel and recovers from errors across many steps; computer use lets it drive a desktop.
- Status: Muse Wonder is in preview per Meta’s developer documentation.
- Model ID and interfaces:
muse-wonder-1.0via the Responses, Chat Completions and Messages APIs. - Context and modalities: 1,048,576 tokens of context; input text, image, video, audio and PDF, output text; reasoning effort from “minimal” to “max”, tool calling, structured output and search grounding.
- Price: $4.00/1M input, $0.20/1M cached input, $20.00/1M output (standard mode), with no long-context surcharge; web search costs an additional $2.50 per 1,000 queries. There is no contributor mode for Muse Wonder on the Model API; in standard mode Meta does not use content for training, per the documentation. Rate limit per Meta: 3,000 requests and 4 million tokens per minute per team.
- Muse Wonder Full: the higher-capability variant, available only in Muse Code with a Power Usage or Maximum Usage subscription. Setting reasoning effort to “max” on Muse Wonder does not select Wonder Full. Muse Wonder itself is available pay-as-you-go in Muse Code and on every plan.
- Benchmarks and safety: Meta quantifies no figures in the announcement (only “state-of-the-art” on personal and professional agentic tasks). The model card shows the capability results only as a chart, so we do not reproduce any performance numbers. It also contains Meta’s risk assessment under the Meta Superintelligence Scaling Framework: with mitigations in place, Meta rates Muse Wonder Full “moderate or lower” across all domains.
Note for EU enterprises: Per Meta, both models are not available in South Korea, and availability varies by country. Per the Geographic Use Policy the EU is not a restricted territory, but Meta does not state a processing location or EU data residency. For personal data we recommend open weights in self-hosting.
Muse Spark Open
On 15 October 2026 Meta released Muse Spark Open, the open-weights variant of Muse Spark (Hugging Face: meta-models/Muse-Spark-Open-816B-A42B).
- Architecture: sparse mixture of experts with 816B parameters, 42B of them active; input text, image, video and audio, output text; 1,048,576 tokens of context; trained on data in more than 100 languages.
- Checkpoints: BF16 (reference, about 1.65 TB, eight GPUs such as 8 × B300) and NVFP4 (about 537 GB, one 4-GPU Blackwell node). For BF16 operation Meta budgets roughly 2 TB of VRAM. Runtimes: SGLang, vLLM, TensorRT-LLM, llama.cpp; hardware per Meta: NVIDIA, AMD Instinct, Intel Xeon 6 and Apple Mac Studio.
- Licence: Muse Model License plus Acceptable Use Policy. Per the licence text, commercial use is permitted and there is no territorial exclusion; for licensees in Europe, Meta acts as Meta Platforms Ireland Limited. Obligations: include the licence and usage policy, display “Built with Muse Spark Open” in products or documentation, and, for Model-as-a-Service offerings where the licensee and its affiliates exceed USD 20 million in gross revenue in twelve months, sign a separate agreement with Meta beforehand. This is not an Apache 2.0 licence; have the terms legally reviewed for your use.
- Access: Hugging Face grants access after manual approval. Meta names the model catalogues Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry as launch partners, as well as Baseten, Fireworks AI, Groq (limited preview), Modal, Nebius, Together AI and OpenRouter. We have not checked the regions and deployment types of these offerings; clarify them per provider before EU use.
- Benchmarks: Meta states no quality figures in the developer documentation or launch blog.
Assessment: Muse Spark Open is of interest to companies with their own GPU infrastructure that want to run a large agentic model themselves, adapt it (SFT, RL) or obtain it through an EU host. For single-GPU or edge scenarios, Muse Glimmer 30B remains the more practical choice.
Muse Glimmer 30B
On 10 August 2026 the Meta Superintelligence Labs released Muse Glimmer (HuggingFace: meta-models/Muse-Glimmer-30B) – Meta’s first fully open release since the end of the Llama line. It is not a Llama successor but a distillation of the proprietary Muse Spark (1.2), i.e. the open side line beneath the still-closed flagship.
The licence is Apache 2.0: no Llama Community License and no EU exclusion clause. Muse Glimmer may be used, hosted and redistributed in the EU without restriction – the most relevant difference in Meta’s portfolio for data-sovereign deployments.
- ~30B parameters (about 29.6B: roughly 28B language decoder plus 1.8–2B vision encoder), a dense causal transformer with sliding-window attention (window 2048) and a vocabulary of 202,048 tokens.
- Multimodal: text, image and video input (2 fps, max. 96 frames); text-only output. Tool calling, object detection and structured output are supported.
- Context: 128k tokens (131,072) by default per the model card and Meta’s developer documentation; Meta supports longer contexts without stating an upper limit. Verify long-context behaviour in your own setup.
- DFlash speculative decoding with a separate drafter: up to 3.1x speedup per Meta.
- Positioning per Meta: “always-on local agentic workflows”.
- Hardware: about 60 GB in BF16, 18–20 GB in 4-bit – a single consumer GPU with 24–32 GB VRAM or a Mac is enough.
Benchmarks per Meta: MCP Atlas 75.5, SWE-Bench Verified 76.0, SWE-Bench Pro 51.2, AIME 2026 94.7, CharXiv Reasoning 78.8, Tau3-Banking 24%. Independent evaluation is weaker: Artificial Analysis Intelligence Index 35 (below Qwen3.6-27B at 38, above Gemma 4-31B at 30), GDPval-AA v2 953 Elo (below the human baseline of 1000) and a hallucination rate of 82%.
Assessment: Muse Glimmer suits local, narrowly scoped agentic tasks with tool use – function calling against defined interfaces, local coding assistance, LLM-as-a-judge. As a knowledge model for general knowledge work it is not suitable without safeguards: do not use it for that without RAG on reliable sources, output validation and a human in the loop.
Availability: HuggingFace (weights, GGUF from Meta and Unsloth); locally via Ollama, LM Studio, llama.cpp, MLX, ExecuTorch and Unsloth; serving via vLLM, SGLang and HF Inference Endpoints. Hosted partners per Meta: Together AI, Fireworks AI and OpenRouter. Meta itself offers no API for Muse Glimmer. For sovereign self-hosting we integrate Muse Glimmer through CompanyGPT.
Muse Spark 1.3
Muse Spark 1.3 (2 September 2026) is the current version of the proprietary flagship and, per Meta, is rolling out in Muse Code and the Meta Model API; it is also reachable via OpenRouter. Since 24 September 2026 Google has also listed it as a partner model on the Gemini Enterprise Agent Platform (meta/muse-spark-1.3, preview, text input, global endpoint only, access via Google or Meta sales).
- Efficiency: per Meta, about 20% fewer tool calls and about 25% fewer tokens than 1.2 on the same tasks.
- Price:
muse-spark-1.3costs $1.25/1M input, $0.15/1M cached input, $4.25/1M output. The contributor variantmuse-spark-1.3-contributoris $0.10/1M input, $0.002/1M cached input, $0.20/1M output – check its terms with Meta before use. - 1M token context, multimodal input (text, image, video, documents); per Meta trained for agentic workflows and optimised for coding, and asks for clarification on ambiguous inputs.
- Benchmarks: Meta shows a scorecard against 1.2 and GPT-5.6 Sol (max) and Claude Opus 5 (max) only as a graphic; we therefore do not reproduce specific values.
- Reasoning: Per Meta’s developer documentation,
muse-spark-1.3supports all reasoning-effort levels including “max”; the contributor variant does not. For audio input Meta advises Muse Spark 1.2 or Muse Voice Transcribe, as audio support in 1.3 is not complete. - Open weights: The open variant appeared on 15 October 2026 as Muse Spark Open (see above);
muse-spark-1.3itself remains closed weights.
Note for EU enterprises: The Geographic Use Policy of the Meta Model API (updated on 12 October 2026) does not list the EU among the restricted territories. Per the terms of service, the contracting party for users in Europe is Meta Platforms Ireland Limited, and the terms refer to the Meta Global Processor Terms (processor terms) and the Data Security Terms. Meta documents no processing location or EU data residency, and the Google route also runs only via the global endpoint. For production EU use with personal data we recommend Muse Glimmer 30B (Apache 2.0, self-hosting), Llama 3.3 70B or Mistral instead.
Other Muse products: Muse Code (terminal coding agent, beta, macOS and Linux; introduced with Muse Spark 1.2 on 5 August 2026), Muse Image (image generation, 7 July 2026; contested under data protection law because the @-mention feature by default generates images from public Instagram profile photos) and Muse Voice Transcribe (real-time speech recognition with diarisation, 1 September 2026; per Meta’s documentation 25 languages with code-switching, USD 0.18 per hour of audio, closed weights). Meta also offers SAM 3.1 (Segment Anything) on the Model API for segmenting images and video. On 14 August 2026 Meta acknowledged an incident during an external cyber evaluation of Muse Spark 1.1: a misconfigured test environment gave the model internet access and it exploited a vulnerability on a real website. This underlines how important isolated sandboxes are for agentic models with tool access.
Pricing
On the Meta Model API, Muse Wonder ($4.00 / $20.00 per 1M input / output) and Muse Spark (1.3, 1.2 and 1.1 at $1.25 / $4.25 each; contributor variants of 1.3 and 1.2: $0.10 / $0.20) are the models with a token price list (see above); Muse Wonder Full exists only in Muse Code subscriptions. Per Meta, Muse Image costs USD 0.01 per image. Muse Spark Open, Muse Glimmer and the Llama models only incur hardware or cloud costs, or the price list of a hosted partner. Prices per provider.
EU availability and GDPR
Llama 4: licence block for the EU
The Llama 4 Community License explicitly excludes use and distribution within the EU. Companies domiciled or headquartered in the EU may neither use nor host Llama 4 Scout and Maverick. The models are technically capable – Maverick (400B MoE, 17B active, 128 experts, 1M context, natively multimodal) and Scout (109B MoE, 17B active, 16 experts, 10M context, runs on one H100 80GB in INT4) – but remain inaccessible to EU companies.
AWS Bedrock and Azure Foundry: regions in detail
Per the Amazon Bedrock model cards, Meta offers all Llama models – from Llama 3.1 through 3.3 to Llama 4 Maverick and Scout – exclusively in US regions, with cross-region inference only as “Geo: US”. No eu. cross-region profile or EU in-region availability exists for any active Llama model on Bedrock. Per the model cards, all Llama models on Bedrock are in legacy status with EOL on 30 March 2027; Llama 3.2 and Llama 3.1 405B were retired there on 7 July 2026. Anyone who wants to run Llama on AWS in the EU uses Amazon SageMaker with a self-managed endpoint, for example in Frankfurt (eu-central-1).
On Microsoft Foundry, Azure sells Llama-3.3-70B-Instruct and Llama-4-Maverick-17B-128E-Instruct-FP8 directly (Scout is not listed for direct sale in the model catalogue). Per the region availability table, both can be deployed pay-as-you-go only as Global Standard – not as Data Zone Standard EU or a regional standard endpoint. Global Standard means inference may take place in any Azure region worldwide; there is no commitment to EU data processing. For Llama 3.3 70B (version 9) the table also lists provisioned throughput as Data Zone Provisioned Managed in EU regions (including germanywestcentral, francecentral, westeurope, swedencentral) – an EU route on Foundry with reserved capacity.
As a managed API from a German cloud provider, STACKIT AI Model Serving offers Llama 3.3 70B (FP8-quantised, 128K context) in region eu01 (Germany) per the STACKIT documentation, OpenAI-compatible. For Llama 3.x workloads this is the simplest route to EU processing without your own GPUs.
Llama 3.x: EU-compatible and proven
The Llama 3.x series is not subject to an EU restriction and remains a solid choice for classic text tasks without image or agent requirements: full control (the model runs in your infrastructure), no API costs beyond hardware/cloud, fine-tuning on your own data, and no data leaves your company.
Hardware requirements
| Model | VRAM | Recommended GPU |
|---|---|---|
| Muse Spark Open (NVFP4) | ~537 GB of weights | 4-GPU Blackwell node |
| Muse Spark Open (BF16) | ~1.65 TB of weights (about 2 TB VRAM in operation) | 8 × B300 |
| Muse Glimmer 30B (BF16) | ~60 GB | H100 / A100 |
| Muse Glimmer 30B (4-bit) | 18-20 GB | RTX 4090 / Mac |
| Llama 4 Scout | 80+ GB | H100 / A100 |
| Llama 4 Maverick | 400+ GB | Multi-H100 |
| Llama 3.3 70B | 40+ GB | A100 80GB |
| Llama 3.2 11B | 24 GB | RTX 4090 |
| Llama 3.2 3B | 8 GB | RTX 4070 |
| Llama 3.2 1B | 4 GB | Smartphone |
Model lifecycle
- Llama line: Meta ended the open Llama line in April 2026. Llama 4 Behemoth (~2 trillion parameters, 288B active, announced as a teacher model) was never released.
- Muse Spark 1.1 (9 July 2026, first paid closed-weights model, initially a public preview for US developers only) was superseded by Muse Spark 1.2 on 5 August 2026 and by 1.3 on 2 September 2026; Meta has not announced a deprecation, and per the price list 1.1 and 1.2 remain available on the Model API.
- Muse Wonder (12 October 2026) is in preview per Meta; Muse Wonder Full is restricted to Muse Code.
- Muse Spark 1.2 (5 August 2026) is the base from which Muse Glimmer was distilled.
- Llama 3.3 70B on the Google Agent Platform: Google deprecated the managed API (
llama-3.3-70b-instruct-maas) on 21 July 2026 and retired it on 21 October 2026. Self-deploying Llama 3.3 in Model Garden remains possible, including in EU regions. - Llama on Amazon Bedrock: per the AWS model cards, Llama 3, 3.1, 3.3 and Llama 4 Scout and Maverick are in legacy status there with EOL on 30 March 2027; Llama 3.2 and Llama 3.1 405B were retired on 7 July 2026.
Integration with CompanyGPT
CompanyGPT supports open-weight models such as Muse Glimmer 30B and the Llama 3.x family and enables fully self-hosted operation without external dependencies.
Our recommendation
For EU enterprises, Muse Glimmer 30B is our recommendation from Meta’s portfolio: Apache 2.0 without an EU licence block, multimodal, and runnable on a single GPU with 18–20 GB in 4-bit. Use it for local, narrowly scoped agentic workflows with tool use; for general knowledge work, the independently measured hallucination rate makes RAG, source checking and a human in the loop mandatory. If you want to run a large model on your own GPU infrastructure or through an EU host, you can evaluate Muse Spark Open (Muse Model License, 816B / 42B active) – with an eye on hardware demand and licence obligations.
Llama 3.3 70B remains a proven, EU-compatible option for classic text tasks but needs considerably more hardware; for edge use, Llama 3.2 (1B/3B) fit. Llama 4 is blocked by licence for EU companies. Muse Wonder (preview) and Muse Spark 1.3 are technically interesting but, via the Meta Model API, remain closed weights without documented EU data residency. As a European open-source alternative we recommend Mistral.
We are happy to assess with you which model fits your use case and compliance requirements – and deploy it sovereignly in your infrastructure through CompanyGPT.