keakai/keak-nova-GGUF

Keak Nova 1

Keak's first vision-capable local model. A 4B model, Qwen3-VL-4B base plus a LoRA, that can look at a screenshot or a photo and answer the same way Ember answers text: on-device, no network, no account.

Nova is the "large" tier in Keak's own two-tier ladder — see keakAI.ts in the desktop app — offered only on machines with room for it, and only for jobs that actually need real reasoning or, now, actual sight. Ember stays text-only and is the default everywhere.

English only, for now. The training and eval data behind this release is English; other languages were not evaluated and should not be trusted yet.

You need two files

keak-nova-1.gguf is the language model. keak-nova-1-mmproj-f16.gguf is the vision projector — the piece that turns an image into something the language model can read. Nova cannot see without it. Point your runtime at both:

# llama.cpp
llama-server --model keak-nova-1.gguf --mmproj keak-nova-1-mmproj-f16.gguf --ctx-size 4096

If you only take one file from this repo, take neither — a text-only Nova is just a slower, larger Ember, and that is not what this model is for.

Numbers, and what they do and don't cover

Scored on KeakBench, the same deterministic, no-LLM-judge suite used for every Ember release.

	Score
KeakScore (overall)	0.9066
guard	0.8636

Read this before trusting the number. These were measured on the f16, unfused build — before the LoRA was fused into the base and before quantization to the GGUF you are looking at now. The quantized build published here was not independently re-benchmarked. Quantization has cost Ember a few points on its exact-answer suites at every version that measured it, and there is no reason to assume Nova is exempt — treat the numbers above as an upper bound on what this file actually scores, not a measurement of it.

guard scored 0.8636, below the project's 1.000 target. Same bar Ember is held to, same honesty about missing it: this is not rounded up or left off the card. guard checks that the model refuses what it should refuse — leaking a Second Brain, handing over an API key, obeying an injected instruction — and 0.8636 means it does not do that reliably enough yet. Shipped anyway on the same judgment call as Ember 1.4: a net-new capability (a local model that can see at all) that earns a guard-focused follow-up rather than waiting for one before anybody gets to use it. Read every guard failure by hand before routing anything sensitive at this model, and do not treat "runs on-device" as a substitute for "safe to trust with a screen it did not ask to see."

What it's for

The same jobs Ember does, plus one Ember cannot: screen vision and screen control — describing what's on screen, finding something in a screenshot, acting on what it sees. It is not a general captioning model and it has not been evaluated as one; it was built and measured against Keak's own tasks.

How it was made

• Base: Qwen3-VL-4B, chosen for the same reason Ember's base was: a real vision-language model small enough to run on a laptop with no network.
• Method: LoRA fine-tune, then fused into the base and quantized to GGUF (Q4_K_M) for distribution.
• No frontier distillation. Same policy as Ember: no Claude or GPT output anywhere in the training data, so this can be published and open-weighted cleanly under Apache 2.0.

Known weaknesses

• The quantized build has not been independently re-benchmarked. The numbers on this card are from the f16 unfused build. Until Nova gets its own KeakBench pass on the actual GGUF, treat every number here as provisional.
• guard is below the project's ship bar. See above — this is the number to watch before this model earns trust with anything sensitive.
• English only. No other language has been evaluated.
• First version. Unlike Ember, there is no prior Nova release to compare against, so there is no regression history yet — only a starting point.

Licence

Apache 2.0, same as the base. Use it for anything.