llmrun
XingChen-AGIQwen2_5_VLForConditionalGeneration

TeleOCR — Hardware Requirements & GPU Compatibility

Vision

TeleOCR is a lightweight vision-language model with about 1.2 billion parameters, released by the XingChen-AGI organization to handle document parsing in both Chinese and English. Its main feature, as noted on its card, is managing digital files alongside camera-captured images within one framework, thanks to geometry-aware modeling and a four-stage training process. The project was renamed from NaviDC-OCR back in September 2026. Given its size, it runs fine on a single modest consumer GPU or even a CPU after quantization. The model supports a context window of 128,000 tokens and comes with the Apache 2.0 license, allowing free commercial and research use. Published in August 2026, its card claims it beats MinerU 2.5 Pro and PaddleOCR-VL 1.6 on the Dr.DocBench challenge evaluation, while a community GGUF conversion is available for llama.cpp.

1.4B params · dense128K contextApache 2.019.4K downloads1.3K likes6.3K quant downloads
S

Runs great on a 24 GB GPU

Runs from8 GB Q4_K_M · 64KBest on24 GB Q4_K_M · ~516 tok/s · 125KMacs from8 GB Q4_K_M · ~56 tok/s

Apache 2.0

Reference GPU · RTX 4090 24 GB

Estimates, ±20% · how we estimate

Run TeleOCR

One recommended way. Every other method, OS and quant is in the step-by-step guide.Each way opens a step-by-step guide for your OS and quant.

Recommended

Ollama
one command · needs Ollama installed
$ollama run hf.co/cstr/TeleOCR-GGUF

Q4_K_M fits fully on the RTX 4090 · 21.9 GB free · 16K context

Download: TeleOCR-Q4_K_M.gguf0.85 GBcstr

install, OS, other quants

Ways to runOther ways

each opens step-by-step

Run in cloud

Runs on RTX 3060 12GB (9 GB headroom) · Q4_K_M

Generation speed
~184 tok/s
generation speed
Cost per 1M output tokens
$0.095
per 1M output tokens

How Much VRAM Does TeleOCR Need?

Select a quantization level; the listed hardware adheres to it.

QuantizationVRAM+32KOn RTX 4090 (reference)
1.0 GB2.8 GB2-bit · K-quant — 2-bit quantization with K-quant improvementsFits: 125K context · ~642 tok/s
1.1 GB2.9 GB3-bit medium — 3-bit medium quantizationFits: 125K context · ~590 tok/s
1.3 GB3 GB4-bit medium — 4-bit medium quantization — most popular sweet spotFits: 125K context · ~516 tok/s
1.4 GB3.2 GB5-bit medium — 5-bit medium quantization — good quality/size tradeoffFits: 125K context · ~458 tok/s
1.6 GB3.3 GB6-bit, very good — 6-bit quantization, very good qualityFits: 125K context · ~415 tok/s
1.8 GB3.6 GB8-bit, near-lossless — 8-bit quantization, near-losslessFits: 125K context · ~358 tok/s
3.3 GB5 GB16-bit · bfloat — Brain floating point 16 — preferred for trainingFits: 125K context · ~202 tok/s
est.calculated, no published file yet

Selected Q4_K_M

4-bit medium quantization

2.1GBat 16K context

Weights
0.9 GB
KV cache
0.9 GB
Overhead
0.3 GB

Fits RTX 4090 (reference) · 21.9 GB free · ~516 tok/s

≥50 tok/s needs ≥98 GB/s · RTX 5090 ~917, M5 Ultra ~677

Which Hardware Can Run TeleOCR?

Q4_K_M · 1.3 GB

TeleOCR (Q4_K_M) requires 1.3 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 2+ GB is recommended. Using the full 128K context window can add up to 7.2 GB, bringing total usage to 8.5 GB. 114 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti. 185 devices with unified memory can run TeleOCR, including Mac Studio M5 Ultra (256 GB), Mac Studio M5 Ultra (512 GB).

By VRAM tier · typical GPUs and Macs at Q4_K_M
8 GB · RTX 4060Q4_K_M · 2.1 GB · ~139 tok/s · 64K8 GB tier →
12 GB · RTX 3060 12GBQ4_K_M · 2.1 GB · ~184 tok/s · 125K12 GB tier →
16 GB · RTX 5060 Ti 16GBQ4_K_M · 2.1 GB · ~229 tok/s · 125K16 GB tier →
24 GB · RTX 4090referenceQ4_K_M · 2.1 GB · ~516 tok/s · 125K24 GB tier →
Mac 32 GB · Mac Mini M4Q4_K_M · 2.1 GB · ~66 tok/s · 125KMacs →
Mac 64 GB · Mac Studio M4 MaxQ4_K_M · 2.1 GB · ~301 tok/s · 125KMacs →

Alternatives to TeleOCR

Similar models

6
ModelScoreVRAMtok/sWhy it is similar
Inkling Small DSpark PreviewRadixArk—1.5 GB~555same size
OLMo 2 0425 1B InstructAllen AI—1.7 GB~449same size
Granite 3.0 1B A400m InstructIBM—1.3 GB~524same size
Llama Guard 3 1Balpindale—1.7 GB~516same size
Deepseek Coder 1.3B InstructDeepSeek—4.3 GB~434same size
Phi 1 5Microsoft—0.9 GB~697same size

Specifications

Publisher
XingChen-AGI
Parameters
1.4B
Architecture
Qwen2_5_VLForConditionalGeneration
Context Length
128,000 tokens
Vocabulary Size
151,936
Release Date
2026-08-14
License
Apache 2.0

Frequently Asked Questions

How much VRAM does TeleOCR need?

TeleOCR requires 1.3 GB of VRAM at Q4_K_M, or 3.3 GB at BF16. Full 128K context adds up to 7.2 GB (8.5 GB total).

VRAM = Weights + KV Cache + Overhead

Weights = 1.4B × 4.8 bits ÷ 8 = 0.8 GB

KV Cache + Overhead ≈ 0.5 GB (at 2K context + ~0.3 GB framework)

Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.

KV Cache + Overhead ≈ 7.7 GB (at full 128K context)

VRAM usage by quantization

1.3 GB
8.5 GB

Learn more about VRAM estimation →

What's the best quantization for TeleOCR?

For TeleOCR, Q4_K_M (1.3 GB) offers the best balance of quality and VRAM usage. Q5_K_M (1.4 GB) provides better quality if you have the VRAM. The smallest option is Q2_K at 1.0 GB.

VRAM requirement by quantization

Q2_K
1.0 GB
Q4_K_M ★
1.3 GB
Q5_K_M
1.4 GB
Q6_K
1.6 GB
Q8_0
1.8 GB
BF16
3.3 GB

★ Recommended — best balance of quality and VRAM usage.

Learn more about quantization →

Can I run TeleOCR on a Mac?

Yes — MacBook Air 13" M3 (8 GB) and 158 other Macs can run TeleOCR. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 1.3 GB of usable unified memory (RAM minus macOS overhead).

Can I run TeleOCR locally?

Yes — TeleOCR can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.

How fast is TeleOCR?

At Q4_K_M, TeleOCR can reach ~3780 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~516 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.

tok/s = (bandwidth GB/s ÷ model GB) × efficiency

Example: NVIDIA B200 → 8000 ÷ 1.2700 × 0.65 = ~4094 tok/s

Estimated speed at Q4_K_M (1.3 GB)

~4095 tok/s
~516 tok/s
~4095 tok/s
~3780 tok/s

Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.

Learn more about tok/s estimation →

What's the download size of TeleOCR?

At Q4_K_M, the download is about 0.85 GB. The full-precision BF16 version is 2.83 GB. The smallest option (Q2_K) is 0.60 GB.

Which GPUs can run TeleOCR?

114 consumer GPUs can run TeleOCR at Q4_K_M. Top options include AMD Radeon RX 5600 XT, AMD Radeon RX 5700 XT, AMD Radeon RX 580 8GB. 114 GPUs have plenty of headroom for comfortable inference.

Which devices can run TeleOCR?

185 devices with unified memory can run TeleOCR at Q4_K_M (1.3 GB), including AMD Ryzen 7 8845HS (Hawk Point) Laptop, AMD Ryzen AI 7 350 (Krackan Point) Laptop, AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10. Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.