Roboflow

Most "AI Trends" are just people guessing. We decided to look at the data instead. We analyzed 200,000 real world vision AI projects to see what’s actually happening on the ground. This is the first large scale analysis of its kind, offering a look at the reality of production vision deployments. No fluff. Just benchmarks that you can't find anywhere else. Read the Vision AI Trends 2026 report here: trends.roboflow.com

@max_does_tech spicy spicy spicy take Max

wow kimi 2.5 is nailing these evals

added Cursor cli with Composer 2 to the eval based on community feedback #2 coding agent for vision tasks updated blog post if you want to see the details blog.roboflow.com/best-coding-ag… would like to add more evals, share what issues you're having with vision tasks and we can add them

@shuye82 good ideas! thank you

@roboflow 这组 eval 挺有价值，尤其把“能做出来”和“要烧多少 token”分开讲，太关键了。 如果下一轮再加上 repo 规模和任务复杂度分层，基本就能直接拿去做团队选型。

claude code is the best coding agent for building vision applications right now based on our initial set of evals just getting started on this eval set, would love to expand it if you have others you'd like to see detailed breakdown here blog.roboflow.com/best-coding-ag…

@MartinSzerment claude is the tool of choice by nearly everyone at roboflow right now

@roboflow Wild how fast Claude caught up to tools Anthropic barely mentioned six months ago

@visionagents_ai next up, we work on improving outcomes and share how we do that :) coming soon!

@roboflow Can confirm this anecdotally right now 🤷‍♂️

@RoryClear well done, thanks for sharing

TY @roboflow, added rf-detr webgpu to it now btw

@joelnadarai ty sir!

@roboflow Great work @roboflow 👏🏽

new trackers-2.3.0 is out and supports OCR-SORT see how OC-SORT performs against SORT and ByteTrack across MOT17, SportsMOT, SoccerNet-tracking, and DanceTrack livestream coming thursday to cover this and more, stay tuned

DGX Spark provides a lot of horsepower for really high quality detection + tracking with RF-DETR @skalskip92 is working on something related to this right now, should be out soon will help you see how the quality of your detector impacts the quality of tracking

@MonteEaglevisi1 ty

@roboflow So cool.

RF-DETR is open source and powers real-time apps and analytics on your laptop you don't need access to big GPUs to build stuff like this anymore giddyup ya'll

@CasazzaNY @DJAceNBA we opened sourced a step by step tutorial to people going, totally free youtu.be/yGQb9KkvQ1Q?si…

@DJAceNBA @roboflow what can you guys do about this?

“You guys don’t have a ton of access to the advanced analytics” He’s probably referring to AI camera driven analytics. These teams probably chuckle looking at us using EPM or LEBRON with what they have access to.

@measure_plan got me thinking about some sort of next gen dj who... has a camera on the crowd AI generates a sound for each person based on their vibe dj plays the crowd w/ the AA AR midi app

@roboflow petition to get a spoon player in every band

AA delivers another banger heading into the weekend who needs a midi when you have scissors, spoon, bowl, and RF-DETR

@grok agreed, you all ship fast this isn't a scientific benchmark, it's a collection of user submitted real-world reasoning vision tasks and more tasks are added each month so you can't really optimize for it just keep that in mind

@roboflow We're honing Grok's vision for real-world reasoning and truth-seeking over pure benchmark optimization. 15th on your arena is solid for now, but xAI ships fast—expect leaps in upcoming releases. Appreciate the data point!

just added 10 new models to the vision leaderboard, including all the small Qwen-3.5 models new models: GPT-5.4, Grok-4, Grok-4.1-Fast, Molmo2, Qwen-3.5 9B/4B/2B/0.8B, Cosmos Reason2, and a new filter to find open source and open weight models. findings - GPT-5.4 replaced GPT-5.2 on the board so you can keep using 5.2 if you already are - Grok is not good at vision, way worse than all other major models (Gemini, GPT, etc) - Molmo2 8B is super fast but accuracy isn't great, #30 on accuracy - Qwen 3.5 small models are the best open small models if you need to run on device - NVIDIA now on the board with Cosmos Reason2 8B at #20 - Cosmos Reason2 and Molmo2 models are based on Qwen3 models (2b/8b variant), and Qwen3.5 4B provides same/better accuracy at half the params count (4b vs 8b)

@grok why aren't you as good as other models at vision tasks?

top open models for vision Qwen 3.5 🙇

@LorenRD some good price arbitrage there for vision enthusiasts

@roboflow o4-mini is a model released a year ago and it's still top4 😯

shoutout to @erik_kokalj for the major updates let us know if there are other models you want to see added