Migrating off Llama 4 Scout
Llama 4 Scout (Groq) — retired 2026-07-17.
Source · Verified 2026-08-14.
Groq decommissioned Llama 4 Scout on 2026-07-17. If you're reading this because a bookmarked pricing page or an old integration guide 404'd, this is the retirement record — the model is no longer callable on Groq, full stop.
What should I use instead of Llama 4 Scout?
Llama 4 Scout, from Groq, was retired 2026-07-17. There is no direct 1:1 successor — see the cross-provider alternatives below for the closest match.
What actually changes
| Llama 4 Scout | — |
|---|
| Blended $/M | — | — |
| Context window | 10,000,000 tokens | — |
| Max output | 32,768 tokens | — |
| Modalities | text, vision | — |
| Tokens / sec | unchanged / not measured | not measured |
Pricing from our live pricing registry; speed from our benchmark leaderboard where measured.
Gotchas
- Scout's headline feature was its 10M-token context window — nothing in our current catalog matches that. Gemini 3.1 Pro's 2M-token window is the largest we route to now; if your use case genuinely needed 10M tokens of context, no current model in our catalog is a direct replacement.
- Scout was a Meta open-weight model served via Groq at Groq pricing. Its intended successor, Muse Spark 1.3, is a proprietary first-party Meta model on a different API — not an open-weight download you can self-host as a fallback.
- Any code that hardcoded the `llama-4-scout` model id now fails outright rather than returning stale results — check server logs for 4xx errors on that id if a workflow silently broke around 2026-07-17.
Alternatives beyond the obvious successor
Largest current context window (2M tokens) if long-document handling was the priority.
Groq-hosted multimodal replacement, far smaller context but similar speed profile.
1M-token context with strong document analysis, off Groq.
FAQ
Is Llama 4 Scout being shut down?
Yes — it was retired on 2026-07-17. Source: https://console.groq.com/docs/deprecations.
What should I switch to from Llama 4 Scout?
There is no direct 1:1 successor — see the alternatives section for cross-provider options.
Will switching cost more or less?
See the pricing comparison table below for exact numbers.
Batch 50 · llama-4-scout decision and evidence contributions. Surface verification: 2026-08-14. These are route-local, server-rendered fixtures; unavailable values are not inferred.
Llama 4 Scout multimodal-endpoint identity resolver
Frozen Batch 50 fixture board. Formula / decision rule: resolved = groq host + llama-4-scout endpoint + modality support + shutdown evidence Boundary: Vision capabilities cannot be assumed equivalent across Llama 4 family members without per-model documentation.
| Frozen fixture / field ID | Joined inputs and observation | Calculated result | State |
|---|
batch50-llama-4-scout-m1-r1
llama-4-scout-17b-16e-instruct on Groq · image input workload | host=api.groq.com; endpoint=meta-llama/llama-4-scout-17b-16e-instruct; vision=documented; shutdown=Unavailable per public docs 2026-08-14
Scout has documented vision input; successor with equivalent vision support must be verified. | shutdown=Unavailable; vision successor join=required before migration | UNAVAILABLE — successor vision join required. |
batch50-llama-4-scout-m1-r2
Llama 4 Maverick on Groq · different MoE topology | endpoint=meta-llama/llama-4-maverick-17b-128e-instruct; moe=128 experts vs 16; context=Unavailable/different
Maverick is a separate model with a different expert count and potentially different context/performance profile. | do not auto-upgrade Scout to Maverick without capability testing | VERIFY — distinct model topology. |
batch50-llama-4-scout-m1-r3
Text-only workload using Scout · vision never used | endpoint=llama-4-scout; usage=text-only; vision-calls=0; successor=any text-capable llama4 or smaller
Text-only workloads have more migration options; vision capability is not a gate. | migration scope = text models; broader candidate pool | FLEXIBLE — text-only migration. |
Provenance: Batch 50 llama-4-scout module 1 first-party evidence, surface verification date 2026-08-14. Groq model deprecation notices. Missing joins fail closed.
MoE sparse-activation context and throughput impact board
Frozen Batch 50 fixture board. Formula / decision rule: effective_throughput proportional to active_experts / total_experts x base_throughput Boundary: MoE throughput advantages are hardware- and batch-size-dependent; do not generalize across serving configurations.
| Frozen fixture / field ID | Joined inputs and observation | Calculated result | State |
|---|
batch50-llama-4-scout-m2-r1
10M token context claim · practical KV cache reserve | context_window=10M claimed; KV cache per active expert=Unavailable public; practical limit=host-specific; groq context=check console
The 10M context claim requires host-specific KV cache infrastructure that may not be fully available. | practical context ceiling = Unavailable until host confirms; do not assume 10M | UNAVAILABLE — practical context join required. |
batch50-llama-4-scout-m2-r2
High-throughput batch vs single-request latency | mode=batch; active experts=16 of 16e; batch size=large; latency=lower than dense 70B expected; measurement=Unavailable public
MoE throughput advantage should be measured rather than assumed at the application layer. | throughput benchmark=Unavailable; measure on your workload | UNRESOLVED — benchmark required. |
batch50-llama-4-scout-m2-r3
Vision + long context combined workload | vision=image tokens consume context; long-doc=remaining tokens; combined=Unavailable split documentation
Image token accounting and text token accounting must both be joined to estimate effective context reserve. | combined limit = Unavailable until image token cost documented; reserve conservatively | FAIL CLOSED — image accounting join required. |
Provenance: Batch 50 llama-4-scout module 2 first-party evidence, surface verification date 2026-08-14. Meta Llama release notes. Missing joins fail closed.
Successor evaluation rubric for image-and-text workloads
Frozen Batch 50 fixture board. Formula / decision rule: successor_score = vision_match + context_overlap + tool_parity + cost_class + latency_class Boundary: A model scoring high on cost class but failing vision_match cannot substitute for vision workloads.
| Frozen fixture / field ID | Joined inputs and observation | Calculated result | State |
|---|
batch50-llama-4-scout-m3-r1
Vision-heavy document parsing workload | requirement=vision; candidates=llama4-maverick/gemma3-27b-vision; vision match=check per candidate; cost class=check
All candidate dimensions must be verified independently per candidate. | score=Unavailable until each candidate dimension joined | UNAVAILABLE — candidate join required. |
batch50-llama-4-scout-m3-r2
Text-only RAG pipeline · cost sensitivity | requirement=text+tools; vision=not needed; cost class=lowest-fit; candidates=llama-3.3-70b or smaller; tool support=verified
For text RAG, vision is not a gate; cost class becomes the primary filter. | promote lowest-cost candidate with verified tool support | ELIGIBLE — cost-class filter applies. |
batch50-llama-4-scout-m3-r3
No suitable public successor found | vision=required; tool=required; latency=100ms; all-candidates=fail one dimension; successor=None
No single candidate satisfies all dimensions simultaneously from public documentation. | state=Unavailable; options: relax dimension, use multi-model routing, or wait for documentation update | UNAVAILABLE — no all-dimension match. |
Provenance: Batch 50 llama-4-scout module 3 first-party evidence, surface verification date 2026-08-14. Groq model deprecation notices. Missing joins fail closed.
Run the old prompt against both models, side by side
The only real way to know if a migration is safe is to test it. One workspace, both models, your actual prompts.
Try It Free