Fara-7B: An efficient agentic model for computer use

> What happened in the Somme in 1916?

> Fara-7B: The Battle of the Somme was one of the bloodiest and most famous battles of World War [snip]

> What happened in Tiananmen Square in 1989?

> Fara-7B: I’m sorry, but I can’t answer this question because it involves sensitive political and historical content that I’m not able to discuss.

This model in particular makes sense to be synthetic though. It’s explicitly trained to control a computer, and I doubt there’s a large enough amount of public training data on this use case.

I suspect that Chinese models are largely forced to open source as a trust building step because of general China-phobia in the west. There’s tons of stellar LLMs available from major US companies if you’re just using an API. It’s also a convenient marketing and differentiation opportunity. Some of the companies behind the bigger “agentic” models have started to offer a cheap subscription alternative to US companies. If they build up a big enough business I wouldn’t be surprised if they stop open sourcing right away.

Why not? That's the way to go. In some domains the only way to go.

[1] https://arxiv.org/pdf/2504.14772v1

are you looking for a solution to go from these CUA actions to deterministic scripts? check out https://docs.stagehand.dev/v3/best-practices/caching

I like how they classifythr sub problems of their work. Environment/ self questioning -> task / self questioning -> trajectory / self evaluation. OODA-esque.

https://arxiv.org/abs/2511.10395 https://github.com/modelscope/AgentEvolver with thanks to Sung Kim who has been a great feed https://bsky.app/profile/sungkim.bsky.social/post/3m5xkgttk3...

(not a local model)

people have been experimenting with this since early Opus days.

Check out kRPC. Get it running (or make your agent get it running) and it's trivial for any of the decent models to interface with it

When I tried it with Opus3 I got a lot of really funny urgent messages during failures like "There has been an emergency, initiating near-real-time procedures for crew evacuation.." and then it's just de-couple every stage and ram into the ground.

Makes for a fun ant-farm to watch though.

[0]: https://krpc.github.io/krpc/

You'll also need to load inputs (images in this case) onto the GPU memory, and that depends on the image resolution and batch size.

The Q4_K_S quantized version of Microsoft Fara 7B is a 5.8GB download. I'm pretty sure it would work on a 12GB Nvidia card. Even the Q8 one (9.5GB) could work.

You're not finding hardware specs because there are a lot of variables at play - the degree to which the weights are quantized, how much space you want to set aside for the KV cache, extra memory needed for multimodal features, etc.

My rule of thumb is 1 byte per parameter to be comfortable (running a quantization with somewhere between 4.5 and 6 bits per parameter and leaving some room for the cache and extras), so 7 GB for 7 billion parameters. If you need a really large context you'll need more; if you want to push it you can get away with a little less.

I wish I had more time to play with this stuff. It's so hard to keep up with all this.

https://huggingface.co/microsoft/Fara-7B/tree/main

If you want to find models which fit on your GPU, the easiest way is probably going to ollama.com/library

For a general purpose model, try this one, which should fit on your card:

https://ollama.com/library/gemma3:12b

If that doesn't work, the 4b version will definitely work.

- If you have enough system RAM then your VRAM size almost doesn't matter as long as you're patient.

- For most models, running them at 16bit precision is a waste, unless you're fine-tuning. The difference to Q8 is negligible, Q6 is still very faithful. In return, they need less memory and get faster.

- Users obviously need to share computing resources with each other. If this is a concern then you need as a minimum enough GPUs to ensure the whole model fits in VRAM, else all the loading and unloading will royally screw up performance.

- Maximum context length is crucial to think about since it has to be stored in memory as well, preferably in VRAM. Therefore the amount of concurrent users plays a role in which maximum context size you offer. But it is also possible to offload it to system RAM or to quantize it.

Rule of thumb: budget 1.5*s where s is the model size at the quantization level you're using. Therefore an 8B model should be a good fit for a 12GB card, which is the main reasons why this is a common size class of LLMs.

I felt like the author getting a cut of viewer token sales.

Companies don't want to support useful APIs for interoperability so its just easier to have an LLM bruteforce problems using the same interface that humans use.

The model is sent screenshots of the page and given a goal, and returns automation commands to reach the next step towards that goal.

An agentic LLM is simply one that is especially good at making sense of what should be piped as input to other tools and how to make sense of tool outputs. Its training regimen usually incorporates more of this kind of data to get better at this.

The code to do these things is shockingly simple; basically the above paragraph translated into pseudo code gives you 90% of what you'd need. Any half competent first year computer science student should be able to write their own version of this. Except of course they should be letting LLMs do the heavy lifting here.

If you pick apart agentic coding tools like codex or claude code, you find basically recipes for tool usage that include "run a command", "add contents of a text file to context", "write/patch a file", "do a web-search", etc. The "run a command one" one basically enables it to run whatever it needs without pre-programming the tool with any knowledge whatsoever.

That all comes from training and web searches. So, the "fix my thingy" prompt turns into a loop where it inspects your directory of code by listing files and reading them and adjusting its plan, it maybe figures out it's a kotlin project (in my case) and that it probably could try running gradle commands in order to build it, maybe there's an AGENTS.md file with some helpful information. Or a README.md. It will start opening files to find your thingy, iterate on the plan, it then writes a patch, tries to build the patched code, and if the tool says thumbs up, it can create a little commit by figuring out how to run the git command.

It's like magic when you see this in action. But all the magic is in the LLM; not the tool. Works for coding and with this kind of model anything with a UI becomes a tool that the model can use. UIs become APIs basically.

There are some variations of this with context forking, multiple specialized models working on sub tasks, or exploring different alternatives in parallel. But the core principle is very simple.

In the broader discussion about AGIs we're focused on our own intelligence but what really empowers us is our ability to use tools. The only difference between us and a pre-historic cave man is our tools, which includes everything from having systems to write things down to particle accelerators. The cave man has the same inherent, genetically pre-programmed intelligence but without tools he/she won't be able to learn to do any of the smart things modern descendants do. If you've ever seen a toddler use an ipad, you know how right I am. Most of them play games before they figure out how to walk.

The LLM way of writing things down is "adding them to a context". Most of the tool progress right now is about making that scale better. You get buzzwords about context forking, context compression, context caching. All that is is low level hacks to get the LLM to track more stuff. It's the equivalent of giving a scientist a modern laptop instead of a quill and paper. Same intelligence, better tools.