llama

109B parameters

Commercial OK

Reviewed June 2026

Llama 4 Scout

Meta's 2026 flagship MoE model. 109B total parameters with only 17B active per forward pass and a record 10-million-token context window — unmatched in production at any tier. Built for long-document workflows, RAG over entire codebases, and continuous-context agents.

License: Llama 4 Community License·Released Apr 5, 2026·Context: 10,000,000 tokens

BLK · VERDICT

Our verdict

OP · Fredoline Eruo|VERIFIED JUN 12, 2026

8.4/10

Positioning

Llama 4 Scout is Meta's "small flagship" of the new generation — natively multimodal, MoE architecture (109B total, 17B active), and the first Llama with a serious long-context story. It's the Llama 3.3 70B replacement for users with ~64 GB VRAM or unified memory.

Strengths

Native vision-language — single model handles image+text without a separate adapter, unlike Llama 3.2 11B Vision's bolted-on approach.
MoE active parameters (17B) keep tokens/sec respectable at flagship quality — ~30–38 tok/s on a 4090 at Q4 with offload.
Architectural long context that genuinely works further than 32K — recall stays competitive into 100K territory in practice.

Limitations

109B total params mean Q4 is ~62 GB — needs dual high-VRAM cards, an A6000-class workstation, or Apple Silicon with 96 GB+ unified memory.
License added new clauses vs Llama 3 — review the AUP if you ship at scale.
Vision quality is solid but not best-in-class — Pixtral and Qwen 2.5 VL still edge it on dense OCR and chart understanding.

Real-world performance on RTX 4090

Q4_K_M (62 GB) — heavy offload required: 8–14 tok/s, only practical with 64 GB+ system RAM
Q5_K_M (74 GB) — workstation only
Q8_0 (~110 GB) — Mac Studio territory

Should you run this locally?

Yes, for workstation rigs (dual 4090, A6000, RTX 6000 Ada) and high-RAM Mac Studios. Excellent native multimodal model. No, for single-card consumer setups — at Q4 you're CPU-offloaded; at lower quants, quality erodes faster than usual on MoE.

How it compares

vs Llama 3.3 70B → Scout is multimodal and has better architectural long context; Llama 3.3 70B is faster on a single 24 GB card. Pick Scout if you have the memory and want vision; otherwise stick with 3.3 70B.
vs Llama 4 Maverick → Maverick is the bigger sibling (400B/17B active). Same active compute but Maverick has a much larger expert pool — better quality if you can afford the disk + memory.
vs Qwen 2.5 VL 72B → Qwen 2.5 VL is stronger on dense visual reasoning; Scout is more usable as a general assistant. Different jobs.

Run this yourself

ollama pull llama4:scout
ollama run llama4:scout

Settings: Q4_K_M GGUF, 16384 ctx, --n-gpu-layers 30 of 49, RTX 4090 + 64 GB DDR5

›Why this rating

8.4/10 — the smallest Llama 4 is the model most local users will actually run, with native multimodality and a 10M-context architecture. Loses points only because real-world recall over the full advertised context is still imperfect.

Overview

Featured in this stack

The L3 execution stacks that pick this model as a recommended component, with the one-line note explaining the role it plays in each.

Stack · L3·Workstation tier·Role: Primary multimodal model (text + vision)
Build a local vision-model stack (May 2026)
Llama 4 Scout is the multimodal flagship in the open Llama 4 family. Strong image understanding combined with the same reasoning quality the text-only Llama 4 line delivers. The pick when you need image-grounded analysis at frontier-tier quality.

Execution notes

L1.25 enriched

Operator notes

Llama 4 Scout is the production multimodal flagship in the Llama 4 line for the workstation-cluster tier (datacenter-class deployment, not consumer-tier). Apache-equivalent Llama Community License. Strong on image-text reasoning at the workstation-cluster scale; the right pick when you need multimodal capability above the consumer Pixtral / Qwen-VL tier but don't need Maverick's frontier-cluster cost.

Deployment notes

The /stacks/local-vision-model recipe pairs this model with vLLM on RTX 4090 (tight, AWQ-INT4 fits with vision-token headroom carefully managed). Production-tier deployment is 2x A100 80GB or 1x H100 with comfortable headroom for multi-image queries.

For consumer-tier multimodal, drop to Pixtral 12B or Qwen 2.5-VL 7B — same multimodal capability shape, smaller hardware envelope.

For frontier-tier multimodal, Llama 4 Maverick is the parent — better image understanding at the multi-node cluster cost.

Runtime compatibility

vLLM ✓ excellent. Vision-language support landed in v0.7+; multi-image inputs handled cleanly.
SGLang ✓ good. Vision support younger than vLLM's — verify on your specific multi-image patterns.
Ollama ✓ partial. Vision-model support landed but vLLM has the production lead for multimodal workloads.
TensorRT-LLM ✗ partial. Multimodal support exists but the per-model recompile friction is severe for vision.
MLX-LM ✓ partial. Apple Silicon multimodal path is younger; Pixtral and Qwen 2.5-VL have stronger MLX integration.

Vision token economics

Vision-language models tokenize images as long sequences of vision tokens — Llama 4 Scout uses approximately 512 vision tokens per 1024×1024 image at default resolution. Multi-image queries (5 images at 512 tokens each) consume 2560 tokens of context just on images, before the user prompt or model response. Plan KV-cache budget accordingly.

Best use cases

Production multimodal serving at the workstation-cluster tier — image-Q&A, document understanding, visual reasoning.
Multi-image queries — the model handles 5-10 images per query cleanly.
Document layout + OCR + reasoning combined — strong on mixed-content workloads.
Llama-ecosystem migration path — drop-in for teams already on Llama 3 with multimodal needs.

When to use a different model

Consumer-tier multimodal (16-24GB VRAM): use Pixtral 12B or Qwen 2.5-VL 7B.
Frontier-tier multimodal: use Llama 4 Maverick — same family, larger.
Apple Silicon multimodal: Pixtral 12B has stronger MLX integration today.
OCR-first workloads: dedicated OCR models (Florence-2, MiniCPM-V) often beat general VLMs at text extraction.
Apache 2.0 license required: Pixtral 12B or Qwen 2.5-VL 72B — clean Apache 2.0.

Failure modes specific to this model

OOM on multi-image queries. A 5-image query blows past KV-cache budgets sized for text-only workloads. Lower `--max-num-seqs` to 4 or 2.
Image-format mismatch. Some vision models require RGB; obscure formats (TIFF, RAW) fail. Pre-convert client-side.
Resolution silently downsampled. Default vision encoder downsamples; high-detail tasks (small text OCR) need explicit higher-resolution model variants.
Mixed-modality tool-call format. Some agent harnesses can't handle text+image content blocks correctly.

Going deeper

/stacks/local-vision-model — the canonical deployment recipe
Llama 4 Maverick — the parent / larger family member
Pixtral 12B, Qwen 2.5-VL 7B — consumer-tier multimodal alternatives
vLLM operational review — multimodal serving runtime

Reviewed May 6, 2026 by Fredoline Eruo

Family & lineage

How this model relates to others in its lineage. Family members share architecture and training-data roots; parent / children edges record direct distillation or fine-tune relationships.

Parent / base model

Llama 4 Maverick400B

Frontier

Family siblings (llama-4)

Llama 4 70B70B

Datacenter

Llama 4 Scout109B

You are here