Xiaomi released an updated 7B reasoning model and VLM version claiming SOTA for their size

Posted by ResearchCrafty1804@reddit | LocalLLaMA | View on Reddit | 38 comments

Xiaomi released an update to its 7B reasoning model, which performs very well on benchmarks, and claims SOTA for its size. Also, Xiaomi released a reasoning VLM version, which again performs excellent in benchmarks. Compatible w/ Qwen VL arch so works across vLLM, Transformers, SGLang and Llama.cpp Bonus: it can reason and is MIT licensed 🔥 LLM: https://huggingface.co/XiaomiMiMo/MiMo-7B-RL-0530 VLM: https://huggingface.co/XiaomiMiMo/MiMo-VL-7B-RL

Reply to Post

38 Comments

[-]

Antique_Job_3407@reddit

No qwen 3 benchmark.

[-]

PaluMacil@reddit

There isn’t a qwen 3 vision model released yet so that would’ve been pretty difficult

[-]

JorgitoEstrella@reddit

They should have used Qwen 3 for comparison

[-]

PaluMacil@reddit

There isn’t a qwen 3 vision model released yet so that would’ve been pretty difficult

[-]

lucidJG@reddit

Anyone tried for ocr? Having a hard time trusting these benchmarks that show qwen is better than Gemma 27b

[-]

PaceZealousideal6091@reddit

I can also confirm that Gemma 3 4B, 12B and 27B performs worse than Qwen 2.5 VL in image processing. Specifically OCR. Gemma tends to hallucinate a lot.

[-]

Eden63@reddit

What is the benefit using a LLM for OCR instead of tesseract?

[-]

R_Duncan@reddit

Try marker + gemini flash and you'll know. Structured pages extracted 100% fine.

[-]

visarga@reddit

Combined layout and semantics. OCR does layout but not semantics. If you want to convert the document to Markdown you need to know how to read the page in a semantic order.

[-]

Yes_but_I_think@reddit

Need not be this complex, even a simple invoice fields extraction is also a unsolved problem.

[-]

PaceZealousideal6091@reddit

This! I totally agree.

[-]

hainesk@reddit

LLMs are better at inferring words and letters from poorly scanned documents, like humans do. I would try out Qwen2.5VL 7B with some documents and you can see how well it understands what's written even if the quality means some of the letters are difficult to read.

[-]

Eden63@reddit

Whats your prompt in this case? "Extract me the words" or "OCR this image?"

[-]

hainesk@reddit

I think it was trained on "Extract the text"

[-]

Eden63@reddit

Thats amazing.

[-]

hainesk@reddit

Honestly, since Ollama released their version of the model, I've just been using the standard 7b Q4KM version with OpenWebui and it works great at OCR.

[-]

Eden63@reddit

I did not know its able to do it. I only thought its like "I see a girl, a blue sky, a red car" and things like that.

[-]

PaceZealousideal6091@reddit

Well, my use case is for doing complex metadata extraction from scientific/research articles. Due to its inherent complexity in data set-up like multi-column organization , special characters and unpredictable location of specfic metadata like DOI, ISBN etc, i always need to ground my data extraction with llms to increase the confidence of metadata extraction. Also, it helps me deal with tables and Images with description better. My pipeline now uses a combination of pymupdf, regex and qwen 2.5 Vl for PDF processing.

[-]

Willing_Landscape_61@reddit

What about Nougat?

[-]

PaceZealousideal6091@reddit

Never heard of it other than in the context of food or Android. Can you elaborate more?

[-]

Willing_Landscape_61@reddit

https://facebookresearch.github.io/nougat/

[-]

Ok_Cow1976@reddit

wow, I wish I knew this earlier. Just tried gemma 27b yesterday, exactly as you said, hallucinate a lot. will try qwen!

[-]

nullmove@reddit

Really? I use Qwen2.5 32B VL a lot and I am happy.

[-]

lucidJG@reddit

It probably comes down to use case. I’ve found at least for mine(manga ocr) Gemma 12b and 27b are much more accurate

[-]

Particular_Rip1032@reddit

At this point, Xiaomi is just making everything lol.

[-]

ResearchCrafty1804@reddit (OP)

https://preview.redd.it/yp3606i4vw3f1.jpeg?width=616&format=pjpg&auto=webp&s=4cd2a80d228b5d534f4e738f98aab55c88e77564

[-]

DinoAmino@reddit

Sooo... interesting that RL is barely an improvement over SFT here.

[-]

You_Wen_AzzHu@reddit

Is the thinking part broken with vllm? It keeps thinking and suddenly stops. Also I have issues that it just outputs in mixed language randomly.

[-]

Iory1998@reddit

It worked once for me. The other times, it just keeps generating the letter G.

[-]

Local_Sell_6662@reddit

GGUF when?

[-]

emsiem22@reddit

[https://huggingface.co/ngxson/MiMo-VL-7B-RL-GGUF](https://huggingface.co/ngxson/MiMo-VL-7B-RL-GGUF)

[-]

GreatBigJerk@reddit

It would be interesting to see it stacked up against Qwen 3 and the new DeepSeek 8B distill.

[-]

robiinn@reddit

Going based on each own given benchmarks, they are (according to themself) these: AIME 24: - R1 Qwen3 8b: 86 - MiMo 7b: 67.5 AIME 25: - R1 Qwen3 8b: 76.3 - MiMo 7b: 52.5 GPQA Diamond: - R1 Qwen3 8b: 61.1 - MiMo 7b: 58.3

[-]

hainesk@reddit

It's weird, the last slide shows different results for MiMo-7B-RL-0530. AIME 24: * R1 Qwen3 8b: 86 * MiMo 7b: 67.5 * MiMo-7B-RL-0530: **80.1** AIME 25: * R1 Qwen3 8b: 76.3 * MiMo 7b: 52.5 * MiMo-7B-RL-0530: **70.2** GPQA Diamond: * R1 Qwen3 8b: 61.1 * MiMo 7b: 58.3 * MiMo-7B-RL-0530: **60.6** So where did that come from?

[-]