0
MODEL SIGNAL · META

Llama 4 Maverick

Meta’s open‑weight natively multimodal Mixture‑of‑Experts model with ~400B total parameters and 17B active parameters using 128 experts, designed for high‑capacity text‑and‑image reasoning and cost‑efficient performance.

CATEGORYMultimodal
CONTEXT1M
RELEASEDApril 5, 2025
Key Features
  • Mixture-of-Experts architecture with ~400B total parameters and 17B active parameters
  • 128 experts per token for sparse activation
  • Natively multimodal (text + image) with early fusion
  • Open-weight release under the Llama 4 community license
  • Large context window (Maverick positioned as cost-efficient mid-context member of the Llama 4 family)
  • Optimized for multimodal reasoning, multilingual tasks, coding, tool calling, and agentic systems

Provider announcement →

Read the Model Signal report →

MODEL SIGNAL

Llama 4 Maverick

Meta’s massive open-weight MoE brings early-fusion multimodal reasoning to 17B-class inference costs.

Bottom line

Released by Meta on April 5, 2025, Llama 4 Maverick is an open-weight, natively multimodal Mixture-of-Experts (MoE) model designed to balance frontier capacity with operator efficiency. By packing approximately 400B total parameters while only activating 17B per token, Meta is delivering high-capacity text and image reasoning under the Llama 4 community license.

Signal

The core signal is the architecture: an early-fusion multimodal model paired with a highly sparse MoE setup. Using 128 experts per token, Maverick achieves a massive routing surface for specialized tasks—specifically optimized for coding, multilingual processing, tool calling, and agentic systems. For operators, this means gaining the representational capacity of a 400B-parameter model with the inference compute profile of a 17B-parameter model. The native text-and-image early fusion indicates that vision is foundational to the model's state, rather than a bolted-on afterthought.

Noise

Claims regarding the exact boundaries of the context window remain unresolved. While Meta positions Maverick as the cost-efficient "mid-context" member of the Llama 4 family, community chatter suggesting a specific 1M token context window is not verified by primary launch artifacts. Operators should treat the maximum context limit as an open question until serving frameworks and official technical specifications confirm the exact bounds.

Model profile

Provider: Meta
Release date: April 5, 2025
Modality: Natively multimodal (Text + Image)
Architecture: Mixture-of-Experts (~400B total parameters, 17B active parameters, 128 experts)
License: Llama 4 community license

Assessment

The operator read is that Meta is aggressively pushing the frontier of deployable AI by leaning into extreme sparsity. A 128-expert routing system is complex, but the payoff is maintaining high inference speeds for heavy agentic workloads. However, the directional implication of a 400B total parameter count is that memory bandwidth and VRAM footprint will be the true bottlenecks. Even if the active compute is only 17B, the weights still have to reside in memory, which will fundamentally dictate deployment infrastructure requirements.

Where it fits

Maverick is engineered for heavy, complex reasoning pipelines that require open-weight control but cannot afford the latency or compute costs of dense frontier models. It fits best in high-throughput agentic systems, multimodal pipelines where visual and textual data must be processed simultaneously (like visual UI testing or complex document intelligence), and enterprise environments leveraging tool calling for automated workflows.

Model Signal · Signal + Noise · Isaiah Steinfeld