
MiniCPM-V 4.7 vs LLaDA-UI: A Phone-Ready Generalist Against a Diffusion GUI Agent
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 187 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 414 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 87 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 138 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0748 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 90 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 79 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 116 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 111 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 67 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 91 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 42 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 210 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 109 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 168 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
Put MiniCPM-V 4.7 next to LLaDA-UI and the obvious frame — generalist versus specialist — turns out to be the least useful way to read the matchup. MiniCPM-V 4.7, which OpenBMB documented and licensed on October 12, 2026, is a 35.2-billion-parameter sparse mixture-of-experts vision-language model whose entire argument is efficiency: the card claims 13-plus tokens per second on an iPhone 17 Pro Max and a better performance-per-inference-time position than Qwen3.6 35B-A3B. LLaDA-UI, inclusionAI's roughly 16.7-billion-parameter mixture-of-experts GUI agent, is a different species altogether — it does not generate tokens left to right at all, it denoises a block of them at once. The question a builder actually has is narrower than "which is better": it is which of these two has been measured on the job you are hiring for, and on what evidence.
The two evidence bases are not the same kind of thing
Both models are vendor-benchmarked, and neither has been independently reproduced. Past that, the comparison splits.
• MiniCPM-V 4.7 — the evaluation section of its model card consists of five images: overall vision-language performance, language performance, the 1B variant's chart, a performance-versus-inference-time frontier plot, and an on-device speed comparison. The prose makes two claims, both comparative rather than absolute. There is no per-benchmark score table in the public documentation.
• LLaDA-UI — the technical report, arXiv paper and project page give named benchmarks with named baselines: ScreenSpot-Pro 52.9 against Qwen3-VL-8B's 52.7, AndroidWorld 53.5 against 47.9, MobileWorld 25.6 against 9.4, WebVoyager 56.9 against 45.2, with ScreenSpot-V2 described as close and OSWorld-Verified framed by inclusionAI itself as the remaining gap.
• What that means in practice — one model publishes task-level numbers you can argue with, and the other publishes a composite you cannot decompose. "Higher overall score in less inference time" is a claim about a chart, and the weighting behind the composite is not published. A ScreenSpot-Pro score of 52.9 is a claim about a defined test with a defined protocol.
• Neither is independently verified — the LLaDA-UI figures come from its own project page and technical report; the MiniCPM-V 4.7 figures come from charts the vendor drew. Treating either as settled would be a mistake, but treating them as equally informative would be a different mistake.
What each model actually is
The architecture gap is wider than the parameter counts suggest, and it decides most of the practical differences downstream.
• Base — MiniCPM-V 4.7 rides a qwen3_5_moe_text backbone with 256 experts, 8 active per token, 40 layers and a three-to-one linear-to-full attention pattern; LLaDA-UI rides LLaDA2.0-mini-base, a masked diffusion language model.
• Generation — MiniCPM-V 4.7 is autoregressive, one token at a time; LLaDA-UI decodes block-wise, denoising a whole block of output positions in parallel and reconstructing under bidirectional conditions inside that block.
• Parameters — 35.2 billion total for MiniCPM-V 4.7 with 8 of 256 experts active per token; approximately 16.7 billion total for LLaDA-UI as a mixture of experts.
• Context — 262,144 tokens maximum position for MiniCPM-V 4.7; LLaDA-UI's card documents a --max-length default of 8192 for its inference script, with large screenshots needing it raised.
• Vision — MiniCPM-V 4.7 uses a 27-layer, 1,152-wide tower at patch 14 and image size 980 with mixed 4x/16x token compression; LLaDA-UI uses a native-resolution ViT initialised from SigLIP with 2D rotary embeddings and a four-to-one spatial merge into a two-layer MLP, explicitly to preserve the aspect ratio of phone, desktop and web screenshots.
• Output — MiniCPM-V 4.7 produces text, with image and video understanding as the input side; LLaDA-UI produces normalised coordinates in a 0-to-999 range and tagged reasoning-plus-action sequences intended to be executed.
• Licence — Apache-2.0 for MiniCPM-V 4.7, declared on October 12; no licence is stated anywhere on the LLaDA-UI project page or model card.
That last bullet is not a footnote if you are shipping. One of these two you can put in a commercial product today without a conversation; the other you cannot.
Decoding is where the real split lives
Block-wise diffusion is the reason LLaDA-UI exists as a separate line of work rather than as another fine-tune. In an autoregressive agent loop, every action token depends on every token before it, so the click coordinate at the end of a response waits on the reasoning that precedes it. In a diffusion agent, the block is refined as a unit, and the practical consequence inclusionAI is selling is latency: on a paired API test the project page reports 4.8 to 6.4 seconds against Qwen3-VL-8B's 17 to 53 seconds, described as mean speedups of 3.58x on web, 6.83x on desktop and 8.95x on mobile inputs, while emitting more tokens per call. On MobileWorld inputs the reported means are 5.9 seconds against 53.0.
Those are vendor-measured, paired, non-cached numbers on the lab's own harness, and they should be read as such. But the direction is worth taking seriously even with the caveat, because agent loops multiply latency. An interface agent that takes ten steps pays the per-step cost ten times, and a per-step difference of ten seconds is the difference between a demo and a product.
MiniCPM-V 4.7 attacks the same problem from the opposite end. It keeps autoregressive decoding and makes each step cheap through expert scheduling — a cross-layer predictor preloads experts the next layer will likely need, prioritising the ones the real router actually selected, with the card explicit that prediction drives prefetch only and the model's own routing still governs computation. Add layer-wise quantization and DSpark speculative sampling and the claim is 13-plus tokens per second for a 35-billion-parameter MoE on a handset. Different mechanism, same objective: make the loop fast enough that a device can run it.

What each one is good at, stated as narrowly as the evidence allows
• Pick LLaDA-UI when the task is an interface — clicking, typing, navigating a mobile app, a desktop window or a web page. It is the one of the two with published grounding and navigation scores, a documented coordinate convention and packaged example requests for mobile, desktop and web.
• Pick MiniCPM-V 4.7 when the task is understanding — describing a photograph, reading a document, parsing a PDF, answering a question about a video, running on hardware you own. It is the one of the two with documented attention to OCR, PDF parsing and visual grounding in its cookbook, with stock support across SGLang, vLLM, llama.cpp and Ollama, and with quantized variants published for download.
• The overlap is smaller than it looks — MiniCPM-V 4.7's card lists visual grounding as a cookbook guide but publishes no grounding score at all. LLaDA-UI's card is about one job. If your workload is GUI automation you are comparing a measured specialist against an unmeasured generalist; if your workload is document and image understanding you are comparing a documented generalist against a model that was never built to do it.
• Deployment friction runs the other way — MiniCPM-V 4.7 offers GGUF, AWQ, GPTQ, int4 and FP8 checkpoints plus an OpenVINO path and claimed day-0 support for Rockchip, Houmo, Intel, AMD and Arm silicon. LLaDA-UI's serving path requires the lab's own SGLang multimodal diffusion patches and its serve_llada_ui.sh script; the model card states plainly that the server cannot be started from a stock SGLang installation.
The cost question nobody publishes
Neither model has a public per-token price, because neither is a hosted endpoint you rent by the million tokens. You host both, which means the cost is your hardware and your latency. That reframes the choice: a 16.7-billion-parameter checkpoint at about 32 GB in BF16 is a materially smaller commitment than a 70 GB 35-billion-parameter one, and if the agent task only needs the smaller model, the larger one is not buying you anything on that workload.
The hybrid case is where a routing layer earns its keep. Run the small, purpose-built model locally for the routine interface work, and hand the genuinely hard frames — a dense screenshot, an unusual document, a question that needs a bigger model's world knowledge — to a hosted endpoint. OrcaRouter puts 200-plus hosted models behind one key with each provider's list price passed through at 0% markup and automatic failover when a provider degrades, and its routing DSL lets a single call fan a hard case out to more than one model when the answer needs to be right rather than cheap. Neither MiniCPM-V 4.7 nor LLaDA-UI is served through it — both are open-weights checkpoints you run yourself, and the router is what sits on the far side of that handoff.

Three questions worth answering before you commit
Is LLaDA-UI's speed advantage real on your hardware? The 3-to-9x figures are paired, non-cached measurements on the lab's own harness against one baseline. Block-wise diffusion has different memory characteristics from autoregressive decoding, and a speedup measured on the vendor's GPU does not transfer automatically to yours. This is the single number most worth reproducing before you build on it.
Can MiniCPM-V 4.7 be pointed at an interface task at all? Plausibly — it is a vision-language model with grounding tooling documented — but the card publishes no GUI-agent benchmark, no coordinate convention and no action vocabulary. Building a GUI agent on it today means defining the output format yourself and hoping the model respects it. LLaDA-UI solved that problem deliberately, with tagged reasoning plus action outputs and per-platform vocabularies.
Which one will still be supported in six months? OpenBMB has shipped a full family around MiniCPM-V 4.7 — two base variants, five quantization formats, two DSpark draft models, a cookbook and an edge-deployment repository — in under a week, with Apache-2.0 terms. LLaDA-UI has a technical report, an arXiv paper, a project page and a code repository, and no stated licence. Breadth of surrounding infrastructure is a reasonable proxy for how much a lab intends to keep something alive.
The verdict
These two are not substitutes, and the honest recommendation follows from that rather than from the parameter counts. If you are building interface automation, LLaDA-UI is the more finished article despite being the smaller model — measured on the task, with a documented output format and a serving path, and carrying the licence risk you need to resolve before shipping. If you are building anything that involves looking at images, reading documents or running on a phone, MiniCPM-V 4.7 is the one with the ecosystem, the permissive licence and the deployment story, and its efficiency claims are exactly as strong as an unreproduced chart can make them. Teams that need both run them together, and the routing layer is the cheapest part of that arrangement.

