Download

Small models.
Careful minds.
Your machine.

We build fine-tuned language models under 27 billion parameters that are safe, honest, and good enough to rely on — and that run quietly on the computer already sitting on your desk.

Read our safety charter →

~/ — zsh

$ curl -fsSL get.emjv.ai | sh && emjv run kestrel-8b
→ detected Apple M3 · 16 GB unified memory
→ pulling  kestrel-8b-q5 (5.4 GB) ████████████ 100%
→ verifying signature + model card ✓
→ safety eval 412/412 checks passed ✓

kestrel › Ready. Running fully offline at 38 tok/s.
RUNS ON macOSWindowsLinuxRaspberry Pi 5Steam Deckyour old gaming PC

01 — Why we exist

Powerful AI shouldn't require a data center, a subscription, or trust in a stranger's server. It should be small enough to own, and careful enough to deserve it.

Local by default

Every emjv model runs offline. Your prompts never leave your device unless you decide they should.

Small on purpose

We cap ourselves at 27B parameters. Constraints force us to train smarter, not just bigger.

Safe before shipped

No model leaves the lab without passing our public red-team suite and an independent ethics review.

02 — The flock

Four models, each sized for a machine you might own.

Wren

3B

Pocket-sized. Great for summaries, notes, and on-device assistants.

min memory
4 GB
context
32k
runs on
phones, Pi 5

Kestrel

8B

Our everyday workhorse. Writing, coding help, and careful reasoning.

min memory
8 GB
context
128k
runs on
most laptops
most popular

Heron

14B

Patient and precise. Built for long documents and research tasks.

min memory
12 GB
context
128k
runs on
gaming GPUs

Osprey

24B

Our most capable model. Frontier-adjacent quality at home.

min memory
20 GB
context
256k
runs on
workstations

03 — The emjv charter

Safety isn't a feature we bolt on. It's the product.

  1. 01

    Honest about uncertainty

    Our models are trained to say “I don't know” rather than invent an answer. Calibration is a release metric, not an afterthought.

  2. 02

    Open evaluations

    Every model ships with its full safety report, red-team results, and known failure modes — published before launch, not after.

  3. 03

    Ethically sourced data

    We license our training data, honor opt-outs, and pay the annotators who shape our models a living wage.

  4. 04

    Safeguards that travel

    Local doesn't mean unguarded. Alignment is trained into the weights themselves, so safety goes wherever the model goes.

0

bytes of your data sent to us

412

safety checks per release

94%

refusal-calibration score

<27B

parameters, always

04 — From the lab

Research & notes

all posts →

Intelligence you can hold in your hands.

Founded in 2025, emjv is a small, independent team of researchers and engineers. We're hiring people who care about getting this right.

Download Kestrel Join us →

or say hello at hello@emjv.dev