Llama 4 Scout 17B
Meta · United States · Llama Community License
Commercial use: Conditional — commercial use with a trigger
Llama 4 Community License. Commercial use permitted below 700 million monthly active users; derivatives must carry "Llama" in the name and display "Built with Llama". Not an OSI open-source licence, whatever it is called elsewhere.
Note that Llama Community License is not an OSI-approved open-source licence. The weights are downloadable, which is not the same thing.
The small model with the widest support anywhere. If a tool runs one local model, it runs this one.
Why it should matter to you
A reasonable first thing to install if you want to find out whether local AI is viable for you at all, mostly because everything supports it. It is no longer the best model at its size — Qwen3 and Gemma 3 beat it — so treat it as the compatibility baseline rather than the destination.
How it connects to our work
Note that the Llama licence is not open source, despite universally being called that. Commercial use is permitted below 700 million monthly active users, derivatives must carry 'Llama' in the name, and the terms must flow downstream. Almost nobody hits the threshold; everybody should still know the clause exists.
Advantages
- Very long advertised context
- Vision capable
- Still the widest tooling support of any family
Disadvantages
- Meta has published no newer open-weight model since mid-2025, so this line is ageing while Chinese and European labs ship quarterly
- The 700M MAU clause and naming requirement rule it out of OSI open source
- Advertised context far exceeds the length at which quality actually holds
- Gated on the Hub: access must be requested and manually approved, which breaks any pipeline that re-fetches weights unattended
Reach for it when
Teams already standardised on Llama tooling, and cases where a US-origin model is a procurement requirement.
Where it falls down
Capability-per-gigabyte against current Qwen and Ornith releases, and long-context work at anything approaching the headline number. Benchmark before you rely on that figure. The manual access gate is also a real operational constraint — automated rebuilds cannot pull it without a cached copy.
Running it
80 GB GPU or 96 GB+ Mac at Q4_K_M. See the hardware sizing tables for how that maps to specific chips and cards, and the quantisation guide for what you give up at each bit width.
Which cloud instance actually runs this, and what it costs
At Q4_K_M the weights are about 61 GB, so 10 catalogued instances fit with room for context, of which the 8 best value are shown. It is a mixture-of-experts model: all 61 GB must sit in memory, but each token reads only about 9.5 GB of active experts. The speeds below assume every weight is read, so treat them as a conservative floor and the costs as a ceiling. Measured single-user speeds for this kind of model run several times higher: about 100 tok/s for gpt-oss-120b on one A100 under vLLM, and around 240 tok/s for Qwen3.6 35B-A3B with multi-token prediction on an RTX 6000 GPU in Unsloth's tests. Cheapest per token is RunPod A100 PCIe (~$10.42 per million output tokens at an estimated 32 tok/s). Fastest on a single card is Azure NCads H100 v5 at roughly 55 tok/s for $35.31 per million tokens, which is the speed-versus-cost trade in one line.
| Instance | GPU | VRAM | Bandwidth | Est. tok/s | $/hr | $/M tokens |
|---|---|---|---|---|---|---|
| RunPod A100 PCIe | A100 80GB | 80 GB | 1935 GB/s | ~32 | $1.190 | $10.42 |
| Hetzner GEX131 | RTX PRO 6000 Blackwell Max-Q | 96 GB | 1792 GB/s | ~29 | $1.420 | $13.43 |
| RunPod H100 PCIe | H100 PCIe | 80 GB | 2000 GB/s | ~33 | $1.990 | $16.86 |
| Lambda H100 PCIe | H100 PCIe | 80 GB | 2000 GB/s | ~33 | $3.290 | $27.87 |
| Azure NC24ads A100 v4 | A100 80GB | 80 GB | 2039 GB/s | ~33 | $3.673 | $30.52 |
| Azure NCads H100 v5 | H100 | 94 GB | 3350 GB/s | ~55 | $6.980 | $35.31 |
| AWS g6e.12xlarge | 4x L40S | 192 GB | 864 GB/s | ~14 | $10.490 | $205.73 |
| AWS p5.48xlarge | 8x H100 | 640 GB | 3350 GB/s | ~55 | $55.040 | $278.39 |
Token generation on a single request is memory-bandwidth-bound, not compute-bound: the GPU must read every active weight from memory for each token it emits. So the ceiling is roughly GPU memory bandwidth divided by the size of the active weights. A 16 GB model on a 300 GB/s L4 tops out near 19 tokens/sec; the same model on a 2039 GB/s A100 tops out near 127. Figures on model pages are that arithmetic, computed from the bandwidth and weight-size columns here. They are a ceiling for one request at batch size 1, before serving overhead, so treat them as an upper bound for comparing instances rather than a benchmark. Batching raises total throughput well above this and does not raise per-request speed.
For identical silicon, the hyperscalers charge roughly 2 to 5 times what the specialist GPU clouds charge. An A100 80GB is about $3.43/GPU-hr on AWS, $3.67 on Azure and $5.03 on Google Cloud, against $1.19-1.39 on RunPod and $1.99 on Lambda. That spread is the single largest cost lever in a self-hosted inference budget, and it is larger than any saving from picking a smaller model.
Three costs that are absent from every headline rate: egress, storage for the weights, and idle time. A 70B model at Q4 is tens of gigabytes that must be pulled to the instance on every cold start, and an endpoint billed hourly is billed while it sits idle waiting for a request. Rates read 2026-09-06; check each provider's live pricing on the directory hub.
Free tiers carrying this model
| Provider | Type | The catch | Live limits |
|---|---|---|---|
| OpenRouter | Aggregator | Free-pool membership changes without notice — a model you built on can stop being free, and the list above will drift. Data passes through OpenRouter and then the upstream provider, so check the privacy and routing settings before sending anything sensitive. | check → |
| NVIDIA NIM | Inference host | It is a credit grant, not a permanent free tier. When the credits run out you are on a paid plan or self-hosting the container. | check → |
| Cloudflare Workers AI | Edge inference | The Neuron accounting makes it hard to predict how many actual requests you get, because different models consume at very different rates. | check → |
| GitHub Models | Aggregator | Explicitly positioned for experimentation, not production. Rate limits are low enough that they will stop a real workload. | check → |
Jurisdiction: United States
Best raw capability and the deepest tooling ecosystem. For EU personal data you are relying on a transfer framework rather than on data never leaving the bloc, so check whether your DPA and your customers accept that.
Watch for: Enterprise API tiers usually promise no training on your data; consumer tiers and free tiers frequently do not. The free tier is where this bites.
This model has a published technical report: arXiv:2407.21783 It is indexed in our research library alongside the work it builds on.
Frequently asked
Can I use Llama 4 Scout 17B commercially?
Llama 4 Community License. Commercial use permitted below 700 million monthly active users; derivatives must carry "Llama" in the name and display "Built with Llama". Not an OSI open-source licence, whatever it is called elsewhere.
What hardware do I need to run Llama 4 Scout 17B?
80 GB GPU or 96 GB+ Mac at Q4_K_M. Weights alone are roughly 61 GB at Q4_K_M and 115 GB at Q8_0. Add KV cache on top of that, which grows with your context length.
What licence is Llama 4 Scout 17B released under?
Llama Community License. Commercial use permitted, but a separate licence from Meta is required if the product had more than 700 million monthly active users in the month before release. Derivative models must carry 'Llama' in the name and an attribution notice. This is not an OSI-approved open-source licence, whatever the marketing says.
Where can I use Llama 4 Scout 17B for free?
Free tiers carrying it include OpenRouter, NVIDIA NIM, Cloudflare Workers AI, GitHub Models. Limits differ per provider and change often, so check each provider's own limits page. You can also self-host the weights, which has no rate limit at all.
Similar models
Gemma 4 31B
24 GB GPU at Q4_K_M, or a 32 GB Mac.
Fara-7B
8 GB GPU or any 16 GB Mac.
gpt-oss-120b
80 GB GPU, or 96 GB+ unified memory.
gpt-oss-20b
16 GB GPU or Mac. A strong default for local agents.
Wiring Llama 4 Scout 17B into something real?
We build the evaluation harness, the failover and the cost ceilings around a model like this, so it survives contact with production.
Maps to AI agent development, AI automation development and AI strategy and consulting. Or see it working: our case studies.