Headless Thrones

ByteDance Seed · Official-source model record

UI-TARS-72B-DPO

UI-TARS · Version UI-TARS-72B-DPO

AvailableChecked Oct 23, 2026
Editorial (0) ↓Download model brief (.md)

Publisher-sourced inventory. Specifications come from official documentation and model cards. “Not documented” means we have no verified value. This inventory is separate from our editorial reviews.

Coverage & sources ↗

Current catalog answers

What to know about UI-TARS-72B-DPO

AI-compiled from linked model documentation. This is a catalog reference, not an independent performance review. Catalog checked October 23, 2026.

Which UI-TARS-72B-DPO does this page describe?

UI-TARS-72B-DPO is a model record from ByteDance Seed in the UI-TARS family. Source

The recorded model identifier or checkpoint is UI-TARS-72B-DPO. Use that exact identifier when checking documentation or a provider listing. Source

How can you access UI-TARS-72B-DPO?

The catalog documents Open weights access for UI-TARS-72B-DPO. Source

As of the catalog check, this version was listed as available. Source

Open weights means the model weights are distributed for use under their stated terms. It does not establish a hosted API, an open-source license, or a hardware requirement.

What can UI-TARS-72B-DPO receive and generate?

Documented inputs: text and image. Source

Documented outputs: text. Source

Input support and output generation are separate capabilities. Reading an image does not establish image generation; a text output listing does not establish speech output. File sizes, formats and media limits depend on the documented implementation.

What license is recorded for UI-TARS-72B-DPO?

The recorded license label is apache-2.0. Source

Read the linked license and associated use terms for this exact version before redistribution, commercial deployment or fine-tuning. A label alone does not establish every permission, restriction or obligation, and it does not determine the terms of a separate hosted service.

What is documented about the UI-TARS-72B-DPO checkpoint?

This evidence refers to ByteDance-Seed/UI-TARS-72B-DPO at recorded revision 43a70a29275e7d1569bb3f3dda5be4a2ff976d5f, checked October 23, 2026. Repository metadata

Safetensors metadata records 73,405,560,320 stored parameters. This is a stored-weight count, not necessarily the number activated per token, a download size or a GPU-memory requirement. Repository metadata

Recorded architecture class: Qwen2VLForConditionalGeneration. Repository metadata

Repository task tag: image-text-to-text. A task tag identifies a distribution category; it does not measure performance on that task. Repository metadata

Listed library: transformers. Check that library's version and model support before loading the weights. Repository metadata

Repository language labels: en. These labels are not independent evaluations of language quality. Repository metadata

Repository access gate: false. This records the repository's gate setting when checked; licensing and use terms still apply. Repository metadata

What the configuration file says

The pinned config.json contains these values. Pinned configuration

max_position_embeddings
32,768
rope_scaling
{"type":"mrope","mrope_section":[16,24,24]}
sliding_window
32,768
torch_dtype
bfloat16
num_attention_heads
64
num_key_value_heads
8
num_hidden_layers
80
hidden_size
8,192

Configuration values describe the checkpoint and its implementation. In particular, max_position_embeddings, sliding_window and rope_scaling do not independently establish an API context limit, tested usable length or hardware requirement. Runtime, quantization, cache, batch size and input length affect deployment behavior.

Read the retrieved model card for setup and qualifications. Model card

Keep the exact version in view

Other UI-TARS models from ByteDance Seed.

These are separate records in the same publisher family. A shared family name does not establish equal capabilities or interchangeable licenses.

The documented record

Specifications, with their sources.

A missing value remains unknown. Input limits, output limits and the total context window are separate measurements.

Publisher
ByteDance SeedSource ↗
Family
UI-TARSSource ↗
Version
UI-TARS-72B-DPOSource ↗
Availability
AvailableSource ↗
Access
Open weightsSource ↗
Inputs
Text · ImageSource ↗
Outputs
TextSource ↗
Context window
Not documented
Maximum input
Not documented
Maximum output
Not documented
Knowledge cutoff
Not documented
Training-data cutoff
Not documented
Training completed
Not documented
Release date
Not documented
License
apache-2.0Source ↗
Input price / 1M tokens
Not documented
Output price / 1M tokens
Not documented

Read this with the numbers.

Official publisher weight listing; availability refers to this distribution, not a guaranteed hosted API. Base, SFT and research variants are identified by their exact published names. Publisher repository evidence (2026-10-23): Publisher task label: image-text-to-text. Published architecture metadata: Qwen2VLForConditionalGeneration. Repository library label: transformers. Model-card language labels: en; these labels are not evaluated proficiency scores. Published safetensors metadata counts 73,405,560,320 stored parameters; this is not an active-parameter count for mixture-of-experts models. Published configuration: max_position_embeddings=32768; sliding_window=32768; dtype=bfloat16; 80 hidden layers. The configuration includes explicit RoPE scaling ({"type":"mrope","mrope_section":[16,24,24]}); no effective context extension is inferred. Checkpoint configuration is not a verified hosted serving limit; context and output limits remain separate documented fields.

Model identifiers and aliases 1
  • ByteDance-Seed/UI-TARS-72B-DPO

Follow the evidence

Official sources.

  1. UI-TARS-72B-DPO official model card ↗huggingface.co
  2. UI-TARS-72B-DPO: official publisher repository metadata ↗huggingface.co
  3. UI-TARS-72B-DPO: publisher model card at revision 43a70a29275e ↗huggingface.co
  4. UI-TARS-72B-DPO: published checkpoint configuration at revision 43a70a29275e ↗huggingface.co

From the Headless Thrones newsroom

Editorial about UI-TARS-72B-DPO

No published articles are linked to this exact model yet. New coverage will appear here as it is published.

Your starting point

UI-TARS-72B-DPO

ByteDance Seed · Open weights

Browse all models ↗
Inputs
Text · ImageSource ↗
Outputs
TextSource ↗
Context window
Not documented
Knowledge cutoff
Not documented

A place to begin comparing

281 models to consider.

Matches must share an access route, support every documented input type of your starting model, and return text. When both publish a context limit in the same field, the candidate must fall between half and twice that limit. Unknown or differently defined limits remain as broader matches, after verified near-context matches. Within each group, recent documented releases appear first. These criteria do not establish equal quality. The full current inventory is considered.

MiniMax

MiniMax-M3.1-Flash

Text + Image input; text output · Shared access: Open weights · Capacity match not established

NVIDIA

Nemotron-3.5-Parse

Text + Image input; text output · Shared access: Open weights · Capacity match not established

Mistral AI

Mistral Large 4

Text + Image input; text output · Shared access: Open weights · Capacity match not established

NVIDIA

NVIDIA-Nemotron-Parse-2.0

Text + Image input; text output · Shared access: Open weights · Capacity match not established

Meta

Muse Glimmer 30B

Text + Image input; text output · Shared access: Open weights · Capacity match not established

MiniMax

MiniMax-M3

Text + Image input; text output · Shared access: Open weights · Capacity match not established

IBM

granite-vision-4.1-4b

Text + Image input; text output · Shared access: Open weights · Capacity match not established

IBM

granite-4.0-3b-vision

Text + Image input; text output · Shared access: Open weights · Capacity match not established

Microsoft

Phi-4-reasoning-vision-15B

Text + Image input; text output · Shared access: Open weights · Capacity match not established

A record you can follow

The source belongs beside the fact.

Official sources

Model records cite the developer’s documentation, release notes and model cards. A listing alone is not a tested service or a recommendation.

Dates mean different things

Knowledge cutoff, training-data cutoff, training completion and release date are separate fields. A model’s name or repository upload date does not establish any of them.

Limits need context

Some publishers use rounded K/M token figures. Deployment, reasoning settings and regional access can change usable limits. Read the source before choosing.

Coverage and last check

Independent editorial model library assembled from primary vendor documentation and official publisher model cards. This is a dated, auditable inventory of the sources reviewed, not a claim to every model worldwide. First-party language-model inventory for OpenAI, Anthropic, Google Gemini/Gemma, xAI, Meta and Reflection. Includes vendor-listed legacy/retired identities with explicit statuses; official pretrained and instruction-tuned checkpoints remain distinct. Excludes embeddings, moderation/guard models, image/audio/video-only generation, community fine-tunes, speculative draft helpers, format/runtime/quantization duplicates, and PaliGemma dataset-specific benchmark fine-tunes. Availability means documented vendor API or downloadable official weights, not guaranteed account eligibility. Unknown specifications and dates remain null. Rounded K/M limits preserve vendor granularity; exact integers are used only where documented. No OpenRouter facts are used. Primary-source inventory of language-capable models from Alibaba/Qwen, DeepSeek, Moonshot, MiniMax, Z.ai, ByteDance Seed, Tencent, Baidu, Xiaomi, StepFun, 01.AI, Shanghai AI Laboratory, Baichuan and OpenBMB. Includes full reachable official public weight inventories plus documented hosted API models and explicit retirements. Embeddings, reward/classifier/router/tokenizer/draft-head components, standalone OCR/document converters and image/video/audio-only generators are excluded; multimodal language models retained. Official packaging variants are folded into aliases where a canonical listing was identifiable; remaining research/base/SFT variants retain exact identities. Not a guarantee every historical/private/regional model is available. Unknown specifications remain null; K/M API display quantities use decimal normalization where explicitly noted, rather than implying verified exact serving ceilings. Model-name dates and repository creation timestamps are never used as release/training dates. Open weights does not mean unrestricted or OSI-open-source licensing. Some model-card bodies were unreachable; their public inventory metadata is retained and disclosed. Official Mistral, Cohere, AI21, NVIDIA Nemotron, Microsoft Phi/MAI, Amazon Nova, IBM Granite, AllenAI OLMo/Molmo, TII Falcon, Nous Research, Arcee and Essential AI model inventory checked 2026-10-23. Includes base, instruction, reasoning, coding and text-producing multimodal models, including legacy API versions. Excludes embeddings/rerankers/reward and safety classifiers, image/audio/video-only generation, robotics-only models, duplicate quantization/runtime packages, adapters, intermediate checkpoint experiments, and third-party mirrors. Arcee coverage focuses on its named product families rather than exploratory merge repositories. Repository listing means listed, not independently verified API availability. K/M figures are nominal decimal normalizations when the source supplies shorthand; deployment-specific exact token limits may differ. Unverified values remain null. This is a source-backed inventory, not a complete claim-level independent editorial review. Supplemental official Liquid AI, Reka and Inception generative language catalog. Excludes media-only generators and decision/classification-only models. No assumed training dates.

Catalog edition updated . Source links are provided on each model record.

View the full coverage report and discovery sources ↗