Llama 4: Full Spec Sheet (June 2026)

Llama 4 is the first Llama generation to use a mixture-of-experts (MoE) architecture. In a dense model (Llama 3, GPT-4, Claude 3), every token passes through every parameter — a 70B model uses 70B parameters for every token. In an MoE model, each token is routed through a small subset of 'experts'; Llama 4 Scout has 17B active parameters per token despite 109B total parameters across 16 experts.

The benefit: training and inference compute scale with active parameters, not total parameters. Scout's inference cost is closer to a dense 17B model than to a dense 109B model. The trade-off: memory footprint at inference time still requires loading all 109B parameters into GPU memory (or paging from system memory at a latency cost), which constrains self-hosting to higher-VRAM setups.

Scout and Maverick share the same 17B active parameter count but differ in expert count and routing. Scout's 16 experts route more concentrated; Maverick's 128 experts route more diverse, with each expert specializing more narrowly. Maverick generally outperforms Scout on complex tasks at slightly higher inference cost.

Behemoth is a different beast — ~288B active parameters across 16 experts, positioned as the teacher model for the Llama 4 distillation pipeline rather than for direct production use. As of June 2026, Behemoth remains in training; Meta has not announced a release timeline.

Llama 4 is open-weight, so pricing varies by where you run it. Hosted serverless APIs from Together AI, Groq, Fireworks, Cerebras, and others charge per-token rates that vary widely by provider, model variant, and deployment mode.

Together AI snapshot (2026-06-20): Scout serverless at $0.18/M input + $0.59/M output. Maverick serverless at $0.27/M input + $0.85/M output. A representative 1,000-in / 500-out call on Scout: `0.001 × $0.18 + 0.0005 × $0.59 = $0.00018 + $0.000295 = $0.000475`. About 0.05¢ per call — among the cheapest frontier-quality calls on any provider.

Groq snapshot (2026-06-20): Scout at ~$0.11/M input + $0.34/M output, with token-per-second output rates often 5-10× higher than other providers (Groq's LPU inference hardware is the differentiator). For latency-sensitive workloads, Groq's higher per-token price often nets cheaper end-to-end at scale because of throughput.

Self-hosting math: a single H100 80GB can serve Scout at modest QPS; an 8×H100 node serves Maverick comfortably. AWS on-demand cost for an 8×H100 p5.48xlarge is ~$98/hour, so per-token break-even vs Together hosting is at sustained ~50-100 calls/sec depending on context length. Most teams under 1M calls/day are economically better served by hosted APIs; above that, self-hosting starts to pencil. Worked $ across providers: Llama 4 cost calculator.

Llama 4 Scout's 10,000,000-token context window is the longest in any open-weight model and the longest in production from any provider. Gemini 2.5 Pro is 1M (closed-weight). Claude Sonnet 4.6 has 1M beta (closed-weight). GPT-5 caps at 400K. Scout fits an entire mid-sized codebase, multiple full-length books, or a year of meeting transcripts in a single call.

Recall across 10M tokens is meaningfully weaker than Gemini 2.5 Pro's recall across 1M — needle-in-haystack benchmarks show degradation past ~2M tokens on Scout. For the strict 'must answer from anywhere in the input' use case, Gemini 2.5 Pro at 1M is the more reliable pick.

Where Scout's 10M genuinely matters: workloads where you'd rather pass the whole document and let the model find the relevant section than build a RAG pipeline. Scout's price ($0.18/M input on Together) makes 10M-context calls economically tractable — a 5M-input call costs $0.90 vs ~$15 on Gemini 2.5 Pro's >200K tier.

Llama 4 is the first Llama generation with native vision built into the base architecture (Llama 3.2 added vision as a separate adapter; Llama 4 has it as a first-class input modality). Pass images alongside text in a multi-content message; the model reasons across both.

Image token accounting depends on the provider implementation but generally follows the standard pattern of ~256-1024 tokens per image at typical resolutions. Together AI and Fireworks document the exact tokenizer behavior on their model pages.

Output is text only. Llama 4 does not natively generate images, audio, or video. For multimodal generation, pair Llama 4 (vision input + text reasoning) with a dedicated generation model (Stable Diffusion / Flux for images, ElevenLabs / Cartesia for audio).

Llama 4 supports tool use natively — the base model is trained on tool-use traces and exposes a `<|tool_calls_section_begin|>` token sequence to indicate a tool call. The user-facing API surface depends on the hosted provider's implementation.

Together AI exposes OpenAI-compatible function calling (`tools` parameter with JSON Schema) for both Scout and Maverick. Fireworks AI does the same. Groq supports tool use on selected models with somewhat different parameter shape — check provider docs.

Structured outputs (JSON Schema-guaranteed output) are similarly provider-dependent. Together AI and Fireworks support `response_format: {type: 'json_object'}` and `{type: 'json_schema', json_schema: {...}}`. For maximum portability across providers, define a tool whose input schema matches your desired output schema and force-call it — that pattern works across every provider that supports tool use.

**Pick Llama 4 (hosted)** when: cost is the dominant constraint and you need frontier-tier quality, you need 10M context (Scout — no closed-weight model has this), you need portable compatibility (the same prompt should work across multiple providers without rewrite), or you want the option to self-host later as scale grows.

**Pick Llama 4 (self-hosted)** when: data residency / sovereignty / on-premise requirements rule out hosted APIs, you have GPU capacity to amortize, or you've measured >$50K/month hosted bills and the break-even math works.

**Pick frontier closed models (GPT-5, Claude Opus, Gemini 2.5 Pro)** over Llama 4 when: you need top-tier reasoning quality on hard tasks (Llama 4 is competitive on benchmarks but typically a half-generation behind the frontier closed models on the hardest tasks), you need the closed-model tooling ecosystems (Responses API, Anthropic's prompt caching mechanics, Gemini's built-in tools), or your team prefers a single-vendor support relationship over multi-provider portability.

Every number on this page was verified against Meta's announcement, Hugging Face model cards, and hosted-provider pricing pages on 2026-06-20. Sources: ai.meta.com/blog/llama-4-multimodal-intelligence for the family overview; Hugging Face model cards for context, modalities, and license; together.ai/pricing for Together AI hosting; groq.com/pricing for Groq.

Hosted-provider pricing moves frequently and varies by deployment (serverless vs dedicated, region, reserved capacity). The snapshot in this guide is a point-in-time reference; re-verify the live provider page before committing budget.

Methodology: when a number could not be cross-confirmed against an official source on the verification date, it was omitted from this card rather than guessed.