Xiaomi-Lens-0

01 / Capabilities

What it does.

Each generated <|mask|> token is decoded into a mask, so the model can point at what it talks about: in a final answer, or in the middle of its reasoning.

[ 01 ]

Referring Expression Comprehension

Finds the object a phrase describes and returns its bounding box, from appearance, action, size, or position.

[ 02 ]

Referring Expression Segmentation

Returns a pixel mask for the referred object, decoded from the visual latent of a generated <|mask|> token.

[ 03 ]

Multi-Instance Grounding

Lists every instance of a category in crowded scenes, each with its own box and mask.

[ 04 ]

Reasoning Segmentation

Resolves functional and implicit requests, such as what could package a gift, after a short reasoning trace.

[ 05 ]

Embodied Referring

Picks the object a person asks to be handed in cluttered scenes, combining category, appearance, and position.

[ 06 ]

Fine-Grained Perception

Thinks with visual latents: grounds each region it refers to while counting or answering high-resolution questions.

02 / Sample playback

Explore model outputs.

Six domains, four recorded cases each. Pick one to replay its output; boxes and masks appear as the corresponding tokens are generated.

Dark brown bear

“dark brown bear at left”

Grid of recorded outputs across all six task domains. — All six domains at a glance.

03 / Method

One backbone for words and masks.

Xiaomi-Lens-0 extends Qwen3.5-4B with a lightweight mask decoder. Joint language and mask training supervises the shared representation used for both spatial predictions and textual answers.

Architecture: text and visual tokens enter the language model backbone; mask-token states go to a mask decoder head that combines them with pre-merge visual features.

The language model consumes spatially merged visual tokens. Each generated <|mask|> token queries the encoder's pre-merge features through the mask decoder. Open full size ↗

01
Mask token → visual latent

The contextual state of each <|mask|> token represents a target region given the image, instruction, and preceding response.
02
One pre-merge feature grid

The decoder reads a single feature grid from the vision encoder before spatial merging. It needs no separate segmentation encoder or multi-scale fusion.
03
Lightweight decoder

Patch-wise compatibility scores between the latent and the grid are reassembled and upsampled into a mask.
04
Thinking with visual latents

Mask tokens placed inside the reasoning trace decode to masks for the regions it discusses. Grounded reasoning traces supervise these intermediate regions.

Training recipe

Two stages.

Both stages mix in-house data with public segmentation, referring, and reasoning datasets.

Stage 1 · Mask decoder pretraining
With the MLLM backbone frozen, the decoder learns to read spatial information from fixed pre-merge features.
Stage 2 · Supervised fine-tuning
The backbone is unfrozen and trained jointly with the decoder, so mask supervision updates the shared visual features alongside next-token supervision.

Donut charts of public and in-house training data for Stage 1 and Stage 2, with per-dataset sample counts.

04 / Data & LensBench

Data for grounding and visual reasoning.

We construct referring-expression data and grounded reasoning traces that link intermediate region references to object masks, and introduce LensBench, an open benchmark for segmentation from explicit and implicit instructions.

Referring expression segmentation data pipeline: visual grounding, instruction synthesis, and consistency filtering.

(a) Candidate objects are localized, their descriptions are checked against image crops, and masks are generated. (b) The masked view and the original image guide referring-instruction synthesis. (c) Re-grounding and contextual verification check that each instruction identifies its target.

LensBench design across four domains and its construction pipeline.

4domains: General, Aerial, Driving, Panoramic

300evaluation units per domain; 1,200 paired units in total

2 × 2Single/Multiple targets × Explicit/Implicit instructions

4,800instruction-level evaluation samples in total

LensBench contains 1,200 human-reviewed paired units and 4,800 instructions across general, aerial, driving, and panoramic imagery (300 units per domain). Each unit combines single/multiple targets with explicit/implicit wording. The two wordings share the same target mask; multiple targets use the union mask and its enclosing box. Targets and masks are fixed before instruction synthesis, followed by automated consistency checks and human review of each complete unit.

Xiaomi-Lens-0 leads mask cIoU in aerial (57.1), driving (68.6), and panoramic (62.3) imagery, and leads both metrics on explicit instructions. Implicit instructions remain challenging: cIoU drops from 71.4 to 54.7, while box Prec@0.5 drops from 85.0 to 70.0. The gap is largest for multiple targets in panoramic scenes, where cIoU falls from 71.8 to 48.9 and box Prec@0.5 from 81.3 to 48.7.

LensBench detailed results

Marginal results pool single and multiple targets. Each cell reports mask cIoU / box Prec@0.5 (%, higher is better). Bold marks best; underline marks second best.

Method	Explicit		Implicit		General		Aerial		Driving		Panoramic
Method	cIoU	Prec@0.5	cIoU	Prec@0.5	cIoU	Prec@0.5	cIoU	Prec@0.5	cIoU	Prec@0.5	cIoU	Prec@0.5
GPT-5.6-Sol	63.8	73.2	61.2	70.0	68.4	67.9	42.2	58.5	68.4	81.8	57.1	78.3
Gemini-3.6-Flash	61.4	74.5	61.0	70.3	63.6	65.8	50.1	62.5	66.8	81.0	60.8	80.3
MiMo-V2.5	51.4	52.2	47.4	46.8	54.2	45.2	32.6	38.6	50.5	52.8	45.4	61.4
Seed-2.1-Pro	62.1	70.2	55.9	64.5	60.6	56.7	50.1	58.5	66.8	77.4	59.9	76.7
Qwen3.8-Max	67.6	69.2	64.5	63.8	74.2	79.7	33.5	62.5	58.0	57.7	60.8	66.2
Kimi-K3	29.4	39.1	28.3	33.0	38.5	26.1	13.8	44.9	47.2	31.1	13.4	42.1
GLM-5.3-Flash	44.3	37.5	43.5	33.3	48.9	26.3	28.0	33.4	51.6	47.5	30.6	34.4
DeepSeek-V4-Flash-Vision-Exp	41.8	30.7	41.1	27.3	46.1	28.1	25.9	28.9	43.9	27.9	32.8	31.0
SAM 3	17.3	38.1	15.4	19.5	23.8	24.8	4.3	27.3	27.7	35.1	11.1	27.9
Xiaomi-Lens-0_4B	71.4	85.0	54.7	70.0	63.9	79.1	57.1	79.6	68.6	79.4	62.3	72.1

LensBench qualitative comparisons across four visual domains; red overlays show predicted and ground-truth regions.

Single and multiple targets under explicit and implicit instructions across four domains. Click the image for the full-resolution comparison.

LensBench on Hugging Face ↗

05 / Benchmarks

Evaluation results.

The same six domains, in the same order. Select one to compare the reported scores. All scores are percentages; higher is better.

01 · Referring Expression Comprehension

RefCOCO / RefCOCO+ / RefCOCOg

92.9Xiaomi-Lens-0 · mean Prec@0.5 · 8 splits

Ranks first on seven of eight RefCOCO, RefCOCO+, and RefCOCOg splits, 5.6 points above its Qwen3.5-4B backbone on average.

01Score comparison

mean Prec@0.5 · 8 splits · top models

Xiaomi-Lens-0_4B92.9
Youtu-VL_4B91.8
Florence-291.1
Qwen3-VL-Seg_4B89.7
InternVL-3.5_4B89.4
MLLMSeg_{InternVL2.5-4B}88.2

02Xiaomi-Lens-0 by split

Orange marks a best score in the table

RefCOCO Val94.6

RefCOCO TestA96.2

RefCOCO TestB92.3

RefCOCO+ Val92.1

RefCOCO+ TestA94.9

RefCOCO+ TestB88.5

RefCOCOg Val92.5

RefCOCOg Test92.4

View detailed scores

Bold marks the best score and underline the second best.

Method	RefCOCO			RefCOCO+			RefCOCOg
Method	Val	TestA	TestB	Val	TestA	TestB	Val	Test
Grounding-DINO-L	90.6	93.2	88.2	82.8	89.0	75.9	86.1	87.0
Florence-2	93.4	95.3	92.0	88.3	92.9	83.6	91.2	91.7
LISA_LLaVA-7B	85.4	88.8	82.6	74.2	79.5	68.4	79.3	80.4
GSVA_Vicuna-7B	86.3	89.2	83.8	72.8	78.8	68.0	81.6	81.8
PixelLM_LLaVA-7B	89.8	92.2	86.4	83.2	87.0	78.9	84.6	86.0
VisionLLM v2_Vicuna-7B	87.9	91.2	84.3	77.6	83.8	70.2	82.9	84.1
Text4Seg_InternVL2-8B	90.3	93.4	87.5	85.2	89.9	79.5	85.4	85.4
UFO_InternVL2-8B	91.4	93.8	88.2	85.7	90.7	79.7	86.8	87.4
MLLMSeg_{InternVL2.5-4B}	91.6	93.3	87.6	86.0	90.0	79.8	88.3	88.6
InternVL-3.5_4B	92.5	94.3	88.2	87.6	92.3	81.6	89.6	89.3
Qwen2.5-VL_3B	89.1	91.7	84.0	82.4	88.0	74.1	85.2	85.7
Qwen3-VL_4B	90.7	92.2	86.7	82.9	89.4	75.6	87.3	87.7
Qwen3-VL-Seg_4B	92.7	94.6	89.8	87.8	92.1	82.2	89.1	89.0
Qwen3.5_4B	90.4	92.3	87.1	84.8	88.9	79.1	88.0	88.2
Youtu-VL_4B	93.6	95.2	90.8	90.1	93.9	85.4	92.2	92.9
Xiaomi-Lens-0_4B	94.6	96.2	92.3	92.1	94.9	88.5	92.5	92.4

RefCOCO-family box F1 (detail)

F1@mIoU is the mean F1 over IoU thresholds from 0.50 to 0.95 in steps of 0.05.

Method	RefCOCO			RefCOCO+			RefCOCOg
Method	F1@0.50	F1@0.95	F1@mIoU	F1@0.50	F1@0.95	F1@mIoU	F1@0.50	F1@0.95	F1@mIoU
GroundingDINO	51.2	23.3	44.3	49.2	20.9	41.4	58.0	23.2	49.7
Rex-Omni	84.4	32.9	71.4	76.9	29.5	64.1	84.7	34.7	72.3
LocateAnything Fast	92.5	44.0	80.6	84.4	39.5	73.1	88.5	40.7	76.2
LocateAnything Hybrid	92.7	44.3	81.4	85.8	40.6	75.1	89.4	41.5	77.7
LocateAnything Slow NTP	92.4	38.4	80.2	85.2	35.7	73.9	88.7	35.6	76.1
Qwen2.5-VL_7B	82.4	17.7	66.9	73.0	16.4	59.3	75.3	15.6	60.3
Qwen3-VL_2B	88.3	26.7	72.4	81.2	24.3	66.5	85.9	25.7	69.8
Qwen3-VL_4B	92.1	32.8	77.6	86.8	30.9	73.4	88.8	32.3	74.7
Qwen3-VL_8B	91.1	29.9	75.8	86.4	29.0	72.0	88.7	28.8	73.3
Qwen3.5_4B	90.1	29.1	74.5	84.1	27.8	69.5	88.4	27.0	71.6
Qwen3.5_9B	92.0	35.1	78.3	87.5	33.9	74.6	89.5	33.5	75.3
RynnBrain1.1	82.6	20.9	65.8	72.5	19.2	57.5	84.3	21.5	67.6
SenseNova-Vision	90.1	45.1	79.7	84.4	41.7	74.3	89.2	44.7	78.7
MiMo-VL-SFT_7B	86.1	14.5	66.3	76.8	13.1	58.9	84.9	14.3	64.8
MiMo-VL-RL_7B	90.4	14.2	68.5	84.8	13.6	64.4	87.7	13.1	64.8
BAGEL	79.4	25.1	65.2	68.3	20.5	54.5	77.7	22.2	62.3
RynnBrain	74.0	13.6	55.6	65.5	11.7	48.9	76.5	14.3	57.9
Xiaomi-Lens-0_4B	94.6	44.6	83.3	92.1	43.6	81.1	92.5	43.8	81.1

02 · Referring Expression Segmentation

RefCOCO / RefCOCO+ / RefCOCOg

81.4Xiaomi-Lens-0 · mean cIoU · 8 splits

Leads the compared models on all eight RefCOCO, RefCOCO+, and RefCOCOg splits.

01Score comparison

mean cIoU · 8 splits · top models

Xiaomi-Lens-0_4B81.4
Qwen3-VL-Seg_4B78.6
Youtu-VL_4B77.7
MLLMSeg_{InternVL2.5-4B}77.5
UniPixel_{Qwen2.5-VL-3B}76.9
Text4Seg_InternVL2-8B75.4

02Xiaomi-Lens-0 by split

Orange marks a best score in the table

RefCOCO Val83.4

RefCOCO TestA84.6

RefCOCO TestB81.2

RefCOCO+ Val79.9

RefCOCO+ TestA83.3

RefCOCO+ TestB76.3

RefCOCOg Val81.1

RefCOCOg Test81.2

View detailed scores

Bold marks the best score and underline the second best.

Method	RefCOCO			RefCOCO+			RefCOCOg
Method	Val	TestA	TestB	Val	TestA	TestB	Val	Test
LISA_LLaVA-7B	74.9	79.1	72.3	65.1	70.8	58.1	67.9	70.6
GSVA_Vicuna-7B	77.2	78.9	73.5	65.9	69.6	59.8	72.7	73.3
PixelLM_LLaVA-7B	73.0	76.5	68.2	66.3	71.7	58.3	69.3	70.5
VisionLLM v2_Vicuna-7B	76.6	79.3	74.3	64.5	69.8	61.5	70.7	71.2
Text4Seg_InternVL2-8B	79.2	81.7	75.6	72.8	77.9	66.5	74.0	75.3
UFO_InternVL2-8B	78.0	79.7	75.6	72.3	76.8	66.6	73.7	74.3
MLLMSeg_{InternVL2.5-4B}	79.5	81.4	77.7	75.4	78.7	70.8	78.1	78.4
SAM 3	75.5	77.6	71.0	67.3	71.1	63.4	73.4	74.0
UniPixel_{Qwen2.5-VL-3B}	80.5	82.6	76.9	74.3	78.9	68.4	76.3	77.0
Youtu-VL_4B	80.7	82.0	78.4	76.2	79.6	71.4	76.5	76.6
Qwen3-VL-Seg_4B	82.3	83.7	79.1	76.2	80.2	70.8	78.2	78.1
Xiaomi-Lens-0_4B	83.4	84.6	81.2	79.9	83.3	76.3	81.1	81.2

03 · Multi-Instance Grounding

COCO 2017

63.9Xiaomi-Lens-0 · F1 · mean over IoU 0.50–0.95

Highest mIoU F1 among the compared models, 0.9 points above GroundingPI.

01Score comparison

F1 · mean over IoU 0.50–0.95 · top models

Xiaomi-Lens-0_4B63.9
GroundingPI63.0
GroundingDINO60.6
LocateAnything Slow NTP59.4
LocateAnything Hybrid59.1

02Xiaomi-Lens-0 by split

Orange marks a best score in the table

IoU 0.50 Recall79.6

IoU 0.50 Precision81.1

IoU 0.50 F180.4

IoU 0.95 Recall27.2

IoU 0.95 Precision27.7

IoU 0.95 F127.4

mIoU Recall63.4

mIoU Precision64.5

mIoU F163.9

View detailed scores

Bold marks the best score and underline the second best.

The mIoU group gives the mean of each metric over IoU thresholds from 0.50 to 0.95 in steps of 0.05.

Method	IoU 0.50			IoU 0.95			mIoU
Method	Recall	Precision	F1	Recall	Precision	F1	Recall	Precision	F1
GroundingDINO	79.6	83.2	81.4	24.7	28.1	26.3	61.1	63.7	60.6
Rex-Omni	69.9	81.0	75.0	17.9	21.7	19.7	53.9	59.4	56.3
LocateAnything Fast	66.2	73.3	69.6	24.7	26.9	25.8	50.9	59.3	53.1
LocateAnything Hybrid	74.7	80.8	77.6	26.6	28.7	27.6	52.4	65.2	59.1
LocateAnything Slow NTP	75.0	81.9	78.3	25.0	27.2	26.0	54.7	65.1	59.4
Qwen2.5-VL_7B	57.2	62.9	59.9	12.2	13.2	12.7	44.7	49.1	46.3
Qwen3-VL_2B	63.5	67.1	65.3	21.8	22.6	22.2	42.6	44.9	42.5
Qwen3-VL_4B	66.1	70.6	68.3	23.7	24.9	24.3	44.9	47.8	46.5
Qwen3-VL_8B	65.7	69.5	67.5	23.9	25.1	24.5	44.8	47.3	46.3
RynnBrain1.1	36.5	60.5	45.5	11.6	16.5	13.6	28.7	45.3	35.2
SenseNova-Vision	79.1	80.5	79.8	27.2	30.2	28.6	63.1	55.4	57.5
GroundingPI	81.4	80.9	81.2	22.9	22.8	22.8	63.2	62.8	63.0
MiMo-VL-SFT_7B	65.5	72.2	68.7	10.0	11.1	10.5	45.7	50.5	48.0
MiMo-VL-RL_7B	65.6	74.5	69.8	9.2	10.3	9.7	44.5	50.2	47.2
BAGEL	65.6	75.2	70.1	11.0	13.0	11.9	46.3	47.2	46.0
RynnBrain	21.3	49.9	29.8	4.9	10.9	6.7	15.5	35.4	21.5
Xiaomi-Lens-0_4B	79.6	81.1	80.4	27.2	27.7	27.4	63.4	64.5	63.9

04 · Reasoning Segmentation

ReasonSeg

67.6Xiaomi-Lens-0 · mean of 4 gIoU / cIoU scores

Best on all four ReasonSeg scores, validation and test.

01Score comparison

mean of 4 gIoU / cIoU scores · top models

Xiaomi-Lens-0_4B67.6
InstructSAM_Qwen3-VL-2B62.4
LISA++_LLaVA-7B62.2
Sa2VA_InternVL2-8B60.9
Sa2VA_InternVL2-2B58.0
LISA_LLaVA-7B50.7

02Xiaomi-Lens-0 by split

Orange marks a best score in the table

ReasonSeg Val gIoU71.2

ReasonSeg Val cIoU68.5

ReasonSeg Test gIoU68.0

ReasonSeg Test cIoU62.5

View detailed scores

Bold marks the best score and underline the second best.

Method	ReasonSeg
	Val		Test
	gIoU	cIoU	gIoU	cIoU
LISA_LLaVA-7B	52.9	54.0	47.3	48.4
LISA++_LLaVA-7B	64.2	68.1	57.0	59.5
PixelLM_LLaVA-7B	44.7	37.4	44.0	39.7
Sa2VA_InternVL2-2B	59.2	60.0	56.9	55.8
Sa2VA_InternVL2-8B	65.2	61.1	59.5	57.8
X-SAM_Phi-3-3.8B	56.6	32.9	57.8	41.0
InstructSAM_Qwen3-VL-2B	62.5	65.0	61.1	61.0
Xiaomi-Lens-0_4B	71.2	68.5	68.0	62.5

05 · Embodied Referring

RoboRefIt

84.5Xiaomi-Lens-0 · mean of 4 gIoU / cIoU scores

Best on all four RoboRefIt scores, on both test splits.

01Score comparison

mean of 4 gIoU / cIoU scores · top models

Xiaomi-Lens-0_4B84.5
InstructSAM_Qwen3-VL-2B67.7
LISA++_LLaVA-7B52.9
PixelLM_LLaVA-7B38.3
Sa2VA_InternVL2-8B32.9
Sa2VA_InternVL2-2B32.5

02Xiaomi-Lens-0 by split

Orange marks a best score in the table

RoboRefIt TestA gIoU87.0

RoboRefIt TestA cIoU88.4

RoboRefIt TestB gIoU81.5

RoboRefIt TestB cIoU81.2

View detailed scores

Bold marks the best score and underline the second best.

Method	RoboRefIt
	TestA		TestB
	gIoU	cIoU	gIoU	cIoU
LISA_LLaVA-7B	36.1	29.9	28.7	22.7
LISA++_LLaVA-7B	58.7	51.5	53.5	47.8
PixelLM_LLaVA-7B	45.2	38.4	38.2	31.4
Sa2VA_InternVL2-2B	56.8	29.0	34.0	10.2
Sa2VA_InternVL2-8B	58.7	27.9	36.1	9.0
X-SAM_Phi-3-3.8B	41.2	10.3	25.1	4.0
InstructSAM_Qwen3-VL-2B	82.5	58.4	74.4	55.5
Xiaomi-Lens-0_4B	87.0	88.4	81.5	81.2

06 · Fine-Grained Perception

8 perception benchmarks

76.4Xiaomi-Lens-0 · mean accuracy · 8 benchmarks

Improves over its Qwen3.5-4B backbone on all eight benchmarks, including +6.8 points on CountQA.

01Score comparison

mean accuracy · 8 benchmarks · top models

Xiaomi-Lens-0_4B76.4
Qwen3.5_4B72.3
Qwen3-VL_8B71.2
Qwen3-VL_4B69.1
MiMo-VL-RL_7B68.4
Qwen2.5-VL_7B63.5

02Xiaomi-Lens-0 by split

Orange marks a best score in the table

General Perception MME-RW-en69.7

General Perception MME-RW-cn67.5

General Perception CV-B88.1

High-Resolution Perception V*87.4

High-Resolution Perception HR-4K82.5

High-Resolution Perception HR-8K80.6

Counting and Color CountQA50.6

Counting and Color ColorB84.4

View detailed scores

Bold marks the best score and underline the second best.

Method	General Perception			High-Resolution Perception			Counting and Color
Method	MME-RW-en	MME-RW-cn	CV-B	V*	HR-4K	HR-8K	CountQA	ColorB
Qwen2.5-VL_7B	60.8	58.3	75.3	78.5	71.6	67.9	18.9	76.4
MiMo-VL-RL_7B	63.4	59.8	84.3	81.2	74.4	72.9	28.3	82.8
Qwen3-VL_4B	63.5	63.6	85.0	80.1	78.3	72.9	28.1	81.6
Qwen3-VL_8B	66.0	66.7	85.4	86.4	78.9	74.6	29.0	82.8
MiniCPM-V-4.5_9B	58.2	56.2	80.2	70.2	69.9	63.6	23.4	79.8
Qwen3.5_4B	66.5	60.3	85.4	82.2	81.5	75.6	43.8	82.7
Xiaomi-Lens-0_4B	69.7	67.5	88.1	87.4	82.5	80.6	50.6	84.4

Radar charts comparing Xiaomi-Lens-0 with baselines on box grounding, referring segmentation, perception, and LensBench.

At a glance, from left to right: box grounding F1, referring segmentation cIoU, fine-grained perception accuracy, and LensBench cIoU.

06 / Citation

Cite this work.

If you find Xiaomi-Lens-0 useful, please cite the technical report.

@misc{xiaomilens0,
  title         = {Xiaomi-Lens-0 Technical Report},
  author        = {Pei Fu and Yike Liu and Yue Sun and Hanmo Wang and
                   Xiuwen Xi and Ruoceng Zhang and Shaojie Zhang and
                   Chao Jiang and Hang Li and Bingrui Liu and Zhenbo Luo and
                   Heng Qu and Changjian Xu and Ming Xu and Jian Luan},
  year          = {2026},
  eprint        = {2610.23279},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2610.23279}
}

Xiaomi-Lens-0

What it does.

Referring Expression Comprehension

Referring Expression Segmentation

Multi-Instance Grounding

Reasoning Segmentation

Embodied Referring

Fine-Grained Perception

Explore model outputs.

Dark brown bear

One backbone for words and masks.

Mask token → visual latent

One pre-merge feature grid

Lightweight decoder

Thinking with visual latents

Two stages.

Stage 1 · Mask decoder pretraining

Stage 2 · Supervised fine-tuning

Data for grounding and visual reasoning.

Evaluation results.

RefCOCO / RefCOCO+ / RefCOCOg

01Score comparison

02Xiaomi-Lens-0 by split

RefCOCO-family box F1 (detail)

RefCOCO / RefCOCO+ / RefCOCOg

01Score comparison

02Xiaomi-Lens-0 by split

COCO 2017

01Score comparison

02Xiaomi-Lens-0 by split

ReasonSeg

01Score comparison

02Xiaomi-Lens-0 by split

RoboRefIt

01Score comparison

02Xiaomi-Lens-0 by split

8 perception benchmarks

01Score comparison

02Xiaomi-Lens-0 by split

Cite this work.