Perceptron

Today we’re releasing two new models in the Isaac family: hybrid-reasoning 2B-parameter and a 1B-parameter vision-language model.

Both introduce major upgrades —optional reasoning traces, perceptive tool calling (including our new Focus system), better OCR, better desktop use, and improved structured output — while remaining fast, compact, and deployable.

Extending the efficient frontier of perception

Isaac 0.2 extends what we started with Isaac 0.1: small models that outperform systems 10× larger on visual reasoning and perception tasks, all running on commodity GPUs or edge devices.

From robotics to media search to industrial inspection, Isaac 0.2 delivers high-accuracy perception without the heavy compute footprint.

What’s new in Isaac 0.2

Reasoning via Thinking Traces

Short, structured reasoning traces improve multi-step decisions, small-object understanding, and ambiguous spatial tasks.

Tool Calling + Focus (Zoom & Crop)

Isaac 0.2 can trigger tool calls to focus (zoom + crop) and re-query the model on a smaller region — improving fine-grained perception.

Structured Outputs

More reliable structured output generation for consistent JSON and predictable downstream integration.

Complex OCR

Improved text recognition across cluttered, low-resolution, or distorted regions — enabling accurate extraction from documents, diagrams, labels, screens, and dense real-world scenes.

Desktop Use

Better performance on everyday desktop and mobile workflows such as UI understanding and navigation, making Isaac faster and more capable for agentic use cases.

Performance

Isaac 0.2 sits at the efficient frontier, outperforming perceptive models of similar size.

Isaac Model Benchmark Comparison

Performance across 23 vision-language benchmarks. Bold indicates best score per row.

	Isaac 0.1 2B params	Isaac 0.2-1B 1B params	Isaac 0.2-2B 2B params	Gemini 2.5 Flash Closed-weight	Gemma 3-27B-IT 27B params	LFM2 1.6B 1.6B params	SmolVLM2-2.2B 2.2B params	Qwen3-VL-2B 2B params
VISION-LANGUAGE UNDERSTANDING
VQA v2	82.12	80.00	80.80	75.90*	71.00	67.71*	65.54*	80.10*
DocVQA	91.22	91.10	92.10	93.70*	86.60	77.18*	79.98	90.10
RealWorldQA	74.87	75.70	77.90	64.20	53.90	65.23	57.50	62.90
TextVQA	79.39	76.60	78.10	79.80*	65.10	65.45*	73.21	79.70
OCRBench	562	870	859	835*	701*	742	729	794
MME	1,293	1,780	2,092	1,883*	1,582*	1,753	1,793	1,872
SEEDBench IMG	74.60	74.20	75.20	65.90*	55.36*	71.97	69.67*	74.40
BLINK	38.50	51.20	55.00	49.50*	39.60	44.40	42.30	53.80
VISUAL PERCEPTION & COUNTING
Pixmo Counting	62.95	61.40	66.70	61.00*	51.24*	55.89*	54.10*	59.80*
CVBench	74.50	74.80	72.50	84.80*	71.23*	60.75*	58.81*	65.00*
CountBench	86.68	81.80	89.70	84.80*	68.00	75.17*	72.76*	88.40
VSR Zeroshot	74.80*	75.40	78.60	70.74*	59.42*	65.87*	63.76*	70.48*
AERIAL PERCEPTION (PERCEPTRON)
Aerial-Understanding	56.00	55.40	59.70	54.00*	40.82*	45.03*	43.59*	48.18*
Aerial-Grounding	71.81	68.90	73.11	56.60*	42.79*	55.14*	53.38*	59.00*
PERCEPTRON GROUNDING
Perceptron-Grounding	53.30	52.10	58.90	25.10*	18.97*	44.42*	43.00*	47.54*
Perceptron-Grounding-Fine	42.20	42.40	46.90	12.60*	9.52*	35.37*	34.24*	37.84*
Perceptron-Grounding-ICL	68.50	69.50	72.50	23.50*	17.77*	54.68*	52.93*	58.50*
REFERRING EXPRESSION GROUNDING
REFL4_val	76.24	82.80	87.80	86.20*	72.41*	73.58*	71.23*	78.73*
RefCOCOg	74.06	82.00	85.80	72.50*	54.81*	64.71*	62.64*	81.20
RefCOCO	77.15	86.40	89.20	73.10*	55.26*	67.28*	65.12*	87.60
RefCOCO+	70.73	77.20	82.60	76.10*	57.53*	62.30*	60.30*	79.00
UI & MULTI-IMAGE UNDERSTANDING
ScreenSpot-Pro	8.10	29.60	35.80	11.40	9.58*	30.00*	29.04*	48.50
MuirBench	37.00	37.60	44.30	39.87*	31.20*	37.12*	35.93*	47.40

* Reported from third-party sources or Perceptron's own evaluations

Try Isaac Now

We believe in building in the open and empowering the community. You can start creating with Isaac today.

Try demo

Download weights

SDK

Try Isaac via API

The Road Ahead

Isaac 0.2 moves us further toward action-capable multimodal systems — models that don’t just see, but reason and operate effectively in complex environments. This is core to our long-term vision for VLMs that underpin autonomous control.

We can’t wait to see how Isaac 0.2 is used in real-world pipelines where perception drives decisions.

Next up: our video-native VLM, enabling temporal reasoning, event detection, and cross-frame insight — key components for our emerging multimodal control stack (VLAs).

Get in touch

Technical inquiries: support@perceptron.inc

Commercial inquiries: sales@perceptron.inc

Join us: join-us@perceptron.inc