Open source / self-host / multimodal

Llama (Meta)

  • Llama 4 Scout / Maverick from 04/05/2025 weights

    The last Llama generation, with no successor. Scout has 10M context, Maverick 1M.

  • Muse Spark 1.3 from 09/02/2026 $1.25/$4.25

    Current Meta model for code and agents. Max reasoning is in Muse Code and the Meta Model API. 1M context, closed weights. The Llama API shut down on 07/06/2026.

  • Muse Glimmer 30B from 08/10/2026 Apache 2.0

    Open weights, one GPU. 131k context.

The Llama 4 license has a 700 million monthly active user threshold. Meta has not signed the EU GPAI Code.

  • Verified2026-10-03
  • Models covered6
  • Sources8
  • Deployment channels4

Purchase decision (when to choose / when to avoid)

Choose if...

  • You want open-weight and control (self-host) + a wide tool ecosystem on a mature model line.
  • You have high volume and want to optimize inference cost (quantizations, vLLM, etc.).
  • You're building on-prem / edge solutions and can maintain models; Muse Glimmer 30B gives Apache 2.0 on a single GPU.

Avoid if...

  • You don't accept Community License / AUP restrictions - check details before deployment.
  • You want the 'simplest possible' deployment without MLOps - a SaaS API will be faster (Meta's Llama API is gone since 07/06/2026).
  • You need a frontier model from Meta - Muse Spark is closed and only available through the Meta Model API.

Cost in practice (scenarios)

Pilot

Hosted at a provider is fastest (Together: Scout $0.18/$0.59, Maverick $0.27/$0.85 per 1M tokens); self-host requires preparation.

  • small team
  • quality testing
Scale

Self-host often wins on cost at steady volume, but maintenance is added; Muse Glimmer 30B runs on one 24 GB GPU.

  • GPU, monitoring, guardrails
These are estimates/scenarios (not an invoice). Actual cost depends on context length, number of users, limits and retention policies.

Deployment / data / enterprise

Deployment channels

  • Self-host (vLLM/llama.cpp/Ollama)
  • Hosting providers (Together, Bedrock; Groq only in the Enterprise tier)
  • Integrations in custom applications
  • Meta Model API serves only Muse (not Llama); Llama API shut down 07/06/2026

Data policy

Training on data
Self-host: on your side; Meta Model API Contributor tier uses prompts and responses for training (applies to Muse).
Retention
Self-host: on your side; hosted: depends on provider.
Data residency
Depends on hosting location.
Key are license terms (Community License) and AUP.

Enterprise readiness

Admin
Self-host: on your side; hosted: depends on provider.
SSO/SCIM
Depends on the platform you deploy on.
Audit
Depends on platform.
DPA
Depends on provider/agreement.
Certifications
Depends on provider/agreement.
Best when you have MLOps and want to control cost at scale.

Best use cases

  • companies wanting to host models on-prem/cloud/edge (privacy, long-term costs) on the line with the largest tool ecosystem
  • fine-tuning and low-latency applications; very long context (Scout 10M on paper, 1M at hosting providers)
  • deployments where license compliance and AUP are acceptable ('Built with Llama' attribution required), or Muse Glimmer 30B under Apache 2.0.

Strengths

  • Huge ecosystem (HF, llama.cpp, vLLM, Ollama, quantizations) and broad availability at hosts (Together, Bedrock).
  • Llama 4 MoE: Scout (10M ctx, 17B/16E), Maverick (1M ctx, 17B/128E), natively multimodal; Llama 3.3 70B still the most widely hosted.
  • Muse Glimmer 30B (Apache 2.0, 08/10/2026): Meta's new open-weight model, multimodal, fits one 24 GB GPU at 4-bit.
  • Muse Spark 1.3 (09/02/2026): fewer tool calls and tokens than 1.2 on agentic tasks (Meta); same API pricing.

Weaknesses / risks

  • A line without a successor: Behemoth never shipped, Meta announces no Llama 5, and the frontier (Muse Spark) is closed-weight.
  • Community License is not open-source in the OSI sense; license compliance and AUP are on the deployer.
  • Meta's Llama API shut down on 07/06/2026 - access only via self-host or third-party providers; GPAI Code refusal - interpretive risk in the EU.

Current models (examples)

  • Llama 4 Scout: MoE 17B active / 16 experts (109B), 10M context, text+image; Llama 4 Maverick: 17B / 128 experts (400B), 1M context. Both 04/05/2025, Llama 4 Community License.
  • Llama 3.3 70B (128K) and Llama 3.1 8B/70B/405B: dense models, still the most widely hosted (Together, Groq, Bedrock).
  • Muse Glimmer 30B (08/10/2026): dense multimodal model from Meta Superintelligence Labs, Apache 2.0, 131K context, one 24 GB GPU at 4-bit; Meta's new open-weight line outside the Llama brand.
  • Muse Spark 1.3 (09/02/2026): frontier agentic coding, 1M ctx, text+image+video; xhigh is generally available, max reasoning is in Muse Code and the Meta Model API; closed weights.
  • Muse Spark 1.2/1.1/1.0: earlier frontier versions; 1.2 weights promised without a date.
  • Llama 4 Behemoth (2T, 288B active) never shipped.

Alternatives (if this model doesn't fit)

MI
Mistral
often simpler model selection + EU
DE
DeepSeek
MIT + cheap reasoning/code
OP
OpenAI
no MLOps needed - faster via API