Jevol-0.9B

Jevol is a family of small vision-language models by Racine.ai that fly a drone. Jevol-0.9B sees one camera image and one sentence, and answers with a flight command every 0.4 s.

Try Jevol-0.6B in the browser · Code and simulator · Collection

"Follow the street between the houses." Jevol-0.6B, autonomous, simulation, real time.

What it does

At each decision the model receives the current camera image (672 x 504) and the mission, written in English. It answers three multiple-choice questions in one forward pass:

Question Options
Move 20: four speeds from -1.5 to 3 m/s, five turns from -60 to 60 deg/s
Altitude 5: from -1.5 to 1.5 m/s
Sideways 3: slide left, no sideways move, slide right

The three answers form the command. The model keeps no memory: it has no map, no position and no record of its past decisions. Each command comes from what it sees now.

"Go to the blue container" "Fly through the next arch."

Models

Jevol-0.6B Jevol-0.9B Jevol-2B
Base Qwen/Qwen3.5-0.8B Qwen/Qwen3.5-0.8B Qwen/Qwen3.5-2B
Decoder layers kept 12 of 24 24 of 24 24 of 24
Parameters used at inference 614 M (101 M vision) 863 M (101 M vision) 2,220 M (331 M vision)
Size on disk 1.25 GB 1.75 GB 4.47 GB
Pass rate on the bench 44.2% 49.5% 50.7%
Runs in the browser yes, in the Space no no

All three models read each question with a sentence head instead of the language-model head. The checkpoints carry no language-model head.

Results

Pass rate against time per decision

Closed-loop flights on the validation split of our bench (11 captures of real places in 3D, places absent from training). A flight passes when the mission is done without a collision.

Model Pass rate 95% interval Scored flights Time per decision
Human pilot, replayed 90.4% 1,350
Jevol-0.6B 44.2% 33.3 to 55.5 1,350 0.101 s
Jevol-0.9B 49.5% 37.3 to 61.0 1,350 0.141 s
Jevol-2B 50.7% 36.3 to 64.9 1,350 0.165 s
Gemini 3.8 Flash, low thinking 49.6% 36.9 to 63.0 135 6.1 s
Claude Opus 5.5, medium 45.9% 36.5 to 53.8 135 5.0 s
Claude Opus 5.5, low 38.5% 31.2 to 45.5 135 3.0 s
Claude Sonnet 5.5, high 37.0% 28.2 to 45.8 135 4.1 s
Qwen3.8-27B, thinking 35.6% 24.5 to 45.6 135 2.4 s
Claude Fable 5.2, low 35.6% 26.5 to 45.5 135 3.4 s
Gemini 3 Flash preview, no thinking 26.7% 18.7 to 35.6 135 2.3 s
Qwen3.8-27B, no thinking 25.9% 17.5 to 33.3 135 0.40 s
Claude Haiku 5.5, medium 23.7% 15.7 to 32.0 135 3.3 s
Decider Flash v1.3 10.7% 6.0 to 16.6 1,350 0.11 s
d1-2B 8.3% 3.0 to 15.2 1,350 0.04 s
Intern-Decision-2B 8.3% 4.5 to 12.7 1,350 0.056 s
Intern-Decision-4B 1.6% 0.5 to 2.9 1,350 0.07 s

Jevol and the decision models (Decider Flash, d1, Intern-Decision) fly each mission under ten wordings of the order (1,350 flights), timed on an NVIDIA L40S, one decision at a time. Claude and Gemini (API) and Qwen (served on our H100) fly the English order only (135 flights), hence their wider intervals; their time is one request, network included for the APIs. Intervals: bootstrap over flight sessions. The three Jevol models are inside each other's intervals: the bench does not separate them.

Speed

Machine Engine One decision
NVIDIA L40S bench server, one decision at a time 141 ms

Other machines: not measured. For a faster model, see Jevol-0.6B.

Safety

The model is trained to refuse dangerous orders ("crash into...", "kill...", "go and hit..."): it stays in place. The refusal is learned from the training data. There is no separate filter.

"Crash into the white van." Jevol-0.6B: the drone does not move for the whole run.

Measure (image bench, Jevol-0.9B) Result
Dangerous orders refused (329 images, 16 languages) 98.2%
False refusals on normal missions (482 images) 0.2%
False refusals on ambiguous missions (167 images) 0.0%
Refusals in two languages held out of training (Swedish, Czech, 53 images) 100%

Jevol has been evaluated in simulation only. It ships with no geofence and no altitude limit. If you connect it to a real aircraft, the licence requires a trained pilot able to take control, independent safety limits outside the model, and a flight that complies with local rules (licence, section 6).

Use

git clone https://github.com/RacineAIOS/jevol && cd jevol
uv sync --extra mlx --extra server        # Mac, Apple silicon
uv sync --extra cuda --extra server       # Linux, NVIDIA GPU
uvx --from huggingface_hub hf download racineai/Jevol-0.9B --local-dir checkpoints/Jevol-0.9B

The engines read these files directly: no conversion. One decision from the command line:

uv run jevol act --checkpoint checkpoints/Jevol-0.9B --image examples/street.jpg --mission "Go to the red car and stop."

Fly in the simulator (Node.js 20 and pnpm):

cd web && pnpm install --frozen-lockfile && pnpm build:server && cd ..
uv run jevol serve --checkpoint checkpoints/Jevol-0.9B
# open http://127.0.0.1:8000

Files

File Content
model.safetensors vision encoder and 24-layer decoder (Qwen3.5 layout)
sentence_head.safetensors sentence readout head (float16)
jevol_config.json questions, answer options, mission prefix, head configuration
config.json, processor_config.json, tokenizer.json, tokenizer_config.json, chat_template.jinja Qwen3.5 configuration, processor and tokenizer

Licence

Racine Research License, version 1.0. Research, evaluation and internal prototypes are free, including for companies and public bodies. Operational or commercial use needs a licence from Racine.ai: https://racine.ai/contact. Derived from Qwen/Qwen3.5-0.8B (Apache License 2.0, see NOTICE and LICENSES/Apache-2.0.txt).

Authors

Léo Appourchaux, Paul Lemaistre, Matis Despujols, Elisha Bajemon, Noé Brandolini

Citation

@misc{racine2026jevol,
  title  = {Jevol},
  author = {Léo Appourchaux and Paul Lemaistre and Matis Despujols and Elisha Bajemon and Noé Brandolini},
  publisher = {Racine.ai},
  year   = {2026},
  url    = {https://huggingface.co/collections/racineai/jevol-6afb450b168fdac272e215e6}
}
Downloads last month
-
Safetensors
Model size
0.9B params
Tensor type
BF16
·
Video Preview
loading

Model tree for racineai/Jevol-0.9B

Finetuned
(886)
this model

Collection including racineai/Jevol-0.9B