Fara-7B
Microsoft · United States · MIT
Commercial use: Yes — free for commercial use
MIT. No conditions.
A small model that can look at a screen and click things, running on your own machine.
Why it should matter to you
The most common automation blocker we meet is a system with no API and no export. Traditionally that means screen-scraping scripts that break whenever the interface changes. A local computer-use model is the first credible alternative, and running it locally matters because the alternative is streaming your internal systems to someone else.
How it connects to our work
We treat these as constrained tools, never as unattended workers. A misclick in a records system is a data-integrity incident, so the design question is always blast radius and rollback before capability.
From our field notesThe 90-second integration feasibility test any business owner can run.
Advantages
- A small agentic model built for computer use and UI interaction rather than chat
- MIT, and small enough to run beside the application it is driving
- Runs locally, so the screen contents it reads never leave the machine
Disadvantages
- Narrow: it is a computer-use model, not a general assistant
- Early, with a limited track record
- Computer-use agents fail in ways that are hard to detect automatically — a wrong click looks like a successful step
Reach for it when
Automating a legacy desktop or web application that has no API, locally, without screenshotting your systems to a third party.
Where it falls down
Unattended operation. Computer-use agents need a constrained blast radius and human checkpoints; the failure mode is silent and expensive.
Running it
8 GB GPU or any 16 GB Mac. 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 4.5 GB, so 20 catalogued instances fit with room for context, of which the 8 best value are shown. Cheapest per token is Hetzner GEX45 (~$0.54 per million output tokens at an estimated 149 tok/s). Fastest on a single card is Azure NCads H100 v5 at roughly 744 tok/s for $2.60 per million tokens, which is the speed-versus-cost trade in one line.
| Instance | GPU | VRAM | Bandwidth | Est. tok/s | $/hr | $/M tokens |
|---|---|---|---|---|---|---|
| Hetzner GEX45 | RTX PRO 4000 Blackwell | 24 GB | 672 GB/s | ~149 | $0.290 | $0.54 |
| RunPod A100 PCIe | A100 80GB | 80 GB | 1935 GB/s | ~430 | $1.190 | $0.77 |
| Hetzner GEX131 | RTX PRO 6000 Blackwell Max-Q | 96 GB | 1792 GB/s | ~398 | $1.420 | $0.99 |
| RunPod L40S | L40S | 48 GB | 864 GB/s | ~192 | $0.790 | $1.14 |
| RunPod H100 PCIe | H100 PCIe | 80 GB | 2000 GB/s | ~444 | $1.990 | $1.24 |
| Lambda A100 40GB | A100 40GB | 40 GB | 1555 GB/s | ~346 | $1.990 | $1.60 |
| Lambda H100 PCIe | H100 PCIe | 80 GB | 2000 GB/s | ~444 | $3.290 | $2.06 |
| AWS g5.xlarge | A10G | 24 GB | 600 GB/s | ~133 | $1.006 | $2.10 |
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.
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.
Frequently asked
Can I use Fara-7B commercially?
MIT. No conditions.
What hardware do I need to run Fara-7B?
8 GB GPU or any 16 GB Mac. Weights alone are roughly 4.5 GB at Q4_K_M and 8 GB at Q8_0. Add KV cache on top of that, which grows with your context length.
What licence is Fara-7B released under?
MIT. Full commercial use. Shortest and least restrictive of the common licences; no explicit patent grant.
Where can I use Fara-7B for free?
Self-hosting the weights is the free route. 8 GB GPU or any 16 GB Mac.
Similar models
Gemma 4 31B
24 GB GPU at Q4_K_M, or a 32 GB Mac.
Llama 4 Scout 17B
80 GB GPU or 96 GB+ Mac at Q4_K_M.
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 Fara-7B 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.