Opt.Gear

July 2026

Opt.Gear: 0.5T 토큰으로 학습한 온디바이스 언어 모델

SRAM·대역폭·전력이 제한된 환경을 출발점으로 설계한 온디바이스 텍스트 생성 모델 Opt.Gear(270M/1B)를 오픈소스로 공개합니다.

글 읽기 →

August 2026

Opt.Gear-QAT: QAT로 지키는 온디바이스 성능

Hessian 기반 scale 초기화와 LSQ를 결합한 Opt.Gear-1B의 Quantization-Aware Training 레시피를 소개합니다.

글 읽기 →

August 2026

Opt.Gear-1M: MCU에서 실행되는 생성 언어 모델

약 1M 파라미터의 언어 모델을 STM32 Cortex-M7에 직접 배포해 20 tokens/s의 자기회귀 생성을 달성했습니다.

글 읽기 →

August 2026

Gear-VLM-1B

Vision-language understanding for on-device multimodal applications.

Coming soon

August 2026

Gear-1B-MTP

Multi-token prediction for faster decoding on edge accelerators.

Coming soon

August 2026

FunctionGear

Structured function calling at the edge — natural language in, typed device intents out.

Coming soon

July 2026

Generative inference on a $60 board: Gear-1M on the STM32H747

How a roughly 1M-parameter model reaches 20 tokens/s on a $60 Cortex-M7 board.

글 읽기 →

The Gear family

One architecture, every device class

The same hybrid design — grouped-query attention for global routing, the ConvKV-Gated Mixer everywhere else — scaled from a server GPU down to a microcontroller.

1B · Flagship

Gear-1B

The strongest quality-per-parameter point in the family. Korean–English bilingual, long-document comprehension, reasoning-tuned SFT — with a decode path built to stay fast.

Context length65,536

Layers / hidden26 / 1,152

TokenizerKORMo · 125K

VariantsBase · Instruct · QAT

Model card →

270M · Mobile

Gear-270M

Low-latency deployment on constrained mobile and edge devices — same hybrid design, same 64K window as the 1B.

Context length65,536

Layers / hidden18 / 640

TokenizerKORMo · 125K

VariantsBase · Instruct

Model card →

1M · Microcontroller

Gear-1M

Natural language to structured board-control intents, inside 1 MB of weights on an ARM Cortex-M7.

Throughput20.0 tok/s @ 400 MHz

Layers / hidden5 / 128

TokenizerBPE · 2,048

Target boardSTM32H747I-DISCO

Model card →

Why Gear is fast

A decode state that doesn’t grow with context

On-device decoding is bound by memory traffic — it is what sets your tokens per second — and the KV-cache is what generates it: every new token adds another position that every later token has to read back. Gear replaces the local attention layers with a fixed-length convolutional state. That is where the speed comes from.

Per-layer decode statelog–log

Global attention · 2·C·d Sliding window · W = 512 ConvKV mixer · L = 3

21,845× lessstate per local layer

At a full 64K context a local layer holds 3 positions of state where global attention holds 65,536. The buffer is sized by the convolution kernel, so it is a property of the layer, not of the prompt.

1

Fixed-size local state

Local mixing keeps a three-position buffer per layer, whether the context is 1K or 64K.

2

Operators an NPU already has

Dynamic QKV matmuls and softmax give way to static linear, depthwise convolution, and element-wise ops — no dynamic shapes in the local path.

3

Global routing kept intact

A small set of GQA layers still carries full-window retrieval, so long-range recall survives the substitution.

Benchmarks

Fast, without giving up quality

Speed is the headline; this is the proof it didn't come at the model's expense. Gear-1B leads every open model at or below 1.2B parameters on all four Korean benchmarks — KMMLU, KoBEST, CLIcK and HAERAE — from 0.5T training tokens and no distillation. The only model ahead on Korean is Qwen3-1.7B: 70% larger, trained on 72× the data.

Korean average vs. pretraining tokens — 1B class↖ better

Korean average is the unweighted mean of KMMLU, KoBEST, CLIcK and HAERAE from the table below. Every peer needed at least 4× the training data; the only higher score needed 72×.

Benchmark	Gear-1B QAT	+ RTN INT4	Gemma-3-1B QAT	+ RTN INT4
English
MMLU	41.3	39.8	35.8	32.5
HellaSwag	54.0	53.9	51.4	48.5
WinoGrande	58.3	56.9	55.0	53.8
Korean
KMMLU	35.0	33.8	26.9	21.2
KoBEST	60.4	59.8	56.2	55.1
HAERAE	42.7	39.9	31.2	23.6

Both models are trained quantization-aware, then the identical RTN INT4 post-training quantization is applied on top (technical report, Table 5). Gear-1B gives up 1.5 points of MMLU and 0.1 of HellaSwag; Gemma-3-1B drops 3.3 and 2.9 on the same pipeline — and 7.6 on HAERAE. A robust QAT foundation is what keeps PTQ from falling off a cliff.

Korean average — 1B class

Bars scale to the highest score in view. Hover a row for detail; the striped bar is outside the size class.

Benchmark	Gear-1B	Gemma-3-1B	Llama-3.2-1B	EXAONE4-1.2B	LFM2.5-1.2B	Qwen3-1.7B
Training tokens	0.5T	2T	9T	12T	28T	36T
Distilled^†	no	yes	yes	yes	yes	yes
Korean
KMMLU	36.0	30.7	29.9	32.6	29.2	41.6
KoBEST	60.3	59.5	51.8	50.7	59.1	62.8
CLIcK	39.7	37.8	30.6	32.9	38.3	49.4
HAERAE	44.2	35.3	32.6	30.2	33.9	52.0
English
MMLU	43.2	39.8	46.1	37.2	51.1	60.3
GPQA	30.3	23.2	28.8	25.3	27.3	24.8
ARC-Easy	70.8	70.0	72.4	58.3	79.2	80.0
ARC-Challenge	38.6	40.1	41.7	34.0	51.0	53.0
HellaSwag	56.0	60.0	61.1	42.4	62.1	60.3
PIQA	71.2	72.5	74.8	55.4	72.9	72.4
WinoGrande	58.2	58.7	62.1	53.9	60.0	60.9
OBQA	36.8	36.4	36.2	27.8	40.8	39.4

^†Whether the reported model was trained against a larger teacher. Gear was not.

^▲Highest score in the comparison set, at any parameter count.

^*All figures from Language Model Evaluation Harness under one shared configuration, evaluated July 2026 — they may differ from numbers published elsewhere. Per-task few-shot settings are listed in the technical report.

On real hardware

Measured on the silicon it ships to

Not a simulator, not a server extrapolation. Gear-1M runs generative inference on a $60 microcontroller board — and every number below was read off the device.

20.0 tok/sdecode, Cortex-M7 @ 400 MHz

50 msper token, W4A32

1.59 MB.text — the whole binary

449 KBpeak RAM

INT4 without the cliff

Post-training quantization costs small models the most — they have no spare capacity to absorb the rounding. Gear-1B-QAT is trained against INT4 weights from the start, so the deployed artifact is the thing that was optimized: 0.6 GB on disk, 98% of the FP16 Korean average.

Korean benchmark	FP16	PTQ INT4	QAT INT4
KMMLU	36.0	33.2	35.5
KoBEST	60.3	56.8	59.7
CLIcK	39.7	36.4	39.1
HAERAE	44.2	40.9	43.0
Average	45.1	41.8	44.3

Model	Precision	Device	Decode (tok/s)	Per token (ms)	Weights
Gear-1M	W4A32	STM32H747I · Cortex-M7 @ 400 MHz	20.0	50.0	0.9 MB
Gear-270M	INT8	Snapdragon 8 Elite Gen5 · Hexagon NPU	96	10.4	310 MB
Gear-270M	INT8	iPhone 17 Pro (A19 Pro) · Neural Engine	88	11.4	310 MB
Gear-1B	INT4 · QAT	Snapdragon 8 Elite Gen5 · Hexagon NPU	41	24.4	620 MB
Gear-1B	INT4 · QAT	iPhone 17 Pro (A19 Pro) · Neural Engine	37	27.0	620 MB

Single-stream decode, greedy sampling. Preliminary — prefill throughput, long-context decode curves and full methodology ship with the technical report.

Download and deploy Opt.Gear

Getting the weights and putting them on a device are different moments. Both are one step: every model ships with its export artifacts already built.

Gear-1B-Instruct Transformers

# pip install transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "OptAI/Gear-1B-Instruct"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

Download model weights

Open weights in the format your stack already reads.

Hugging Face safetensors GGUF llama.cpp ONNX opset 17

From weights to hardware

Prebuilt exports for the runtimes and accelerators we profile.

Opt.GearGGUF · ONNX · safetensors

RuntimeTransformers · vLLM · ExecuTorch

Qualcomm
Hexagon NPU Apple
Neural Engine ARM
Cortex-M

Run on device

Prebuilt binaries for the accelerators we profile on — the conversion pipeline is our job, not yours.

Transformers vLLM ExecuTorch ONNX Runtime

Get started with Opt.Gear

Download the weights, evaluate the model, and bring fast inference to your device.

View the models ↗

Opt.Gear: 0.5T 토큰으로 학습한 온디바이스 언어 모델

Opt.Gear-QAT: QAT로 지키는 온디바이스 성능

Opt.Gear-1M: MCU에서 실행되는 생성 언어 모델

Gear-VLM-1B

Gear-1B-MTP

FunctionGear

Generative inference on a $60 board: Gear-1M on the STM32H747

Use cases

Chatbot

Agent

One architecture, every device class

Gear-1B

Gear-270M

Gear-1M

A decode state that doesn’t grow with context

Fixed-size local state

Operators an NPU already has

Global routing kept intact

Fast, without giving up quality

Apple iPhone 17 Pro Apple A19 Pro · Neural Engine · CoreML

Samsung Galaxy S26 Qualcomm Snapdragon 8 Elite Gen 5 · Hexagon NPU · Genie

Measured on the silicon it ships to

INT4 without the cliff

Download and deploy Opt.Gear

Download model weights

From weights to hardware

Run on device

Get started with Opt.Gear