tarruda/Rio-3.5-Open-397B-GGUF overview

---

quantized_by: tarruda

pipeline_tag: text-generation

base_model: prefeitura-rio/Rio-3.5-Open-397B

base_model_relation: quantized

license: apache-2.0

license_link: https://huggingface.co/Qwen/Qwen3.5-397B-A17B/blob/main/LICENSE

tags:

• gguf
• imatrix
• conversational
• qwen3_5_moe
• multimodal
• vision
• llama.cpp

---

Intro

Rio 3.5 Open 397B is a merge of

Nex-N2-Pro and

Qwen3.5-397B-A17B. In a public

discussion, Nex-AGI reports that the released Rio weights are approximately an

element-wise 0.6 x Nex-N2-Pro + 0.4 x Qwen3.5-397B-A17B blend. I haven't

verified this myself, but the report can be found here:

https://github.com/nex-agi/Nex-N2/issues/4#issuecomment-4702181710

Nex-N2-Pro is itself built on Qwen3.5-397B-A17B. Because attribution around Rio

has been confusing, this quant explicitly credits both Nex-AGI and Qwen.

License metadata is set to Apache-2.0 to match the named upstream components:

Qwen3.5-397B-A17B and Nex-N2-Pro are both listed as Apache-2.0 on Hugging Face.

The chat template from Rio had misleading information about the model training,

so I replaced it with the original one from Qwen 3.5 397B.

Recipes

There are two quants in this repo:

• A 2.46 BPW IQ2_XS GGUF which uses @ubergarm's smol-IQ2_XS mainline recipe,
• A 2.54 BPW IQ3_XXS GGUF which uses the my same Qwen 397B recipe

The imatrix.gguf file was copied from bartowski's GGUF repo.

Why

Despite all the confusing information presented by the team behind Rio, they

created a good model that has become one of my favorite local coding

assistants.

Some people dismiss this model because of its inferior benchmark numbers

compared to N2, but I found that it improves on N2 Pro in some ways. This is

anecdotal, but it is my experience:

• Mixed reasoning style. It normally has N2's "caveman" reasoning style, but it

seems to think less while producing similar results.

• At the same time, it sometimes uses Qwen's style, which can be good for some

non-coding tasks.

• It doesn't get stuck in reasoning loops.

Running

This is the script I use to run it locally:

#!/bin/sh -e

reasoning_budget_message=". Need answer.
Let's draft final."

llama-server --no-warmup --no-mmap \
  --model IQ2_XS/Rio-3.5-Open-397B-IQ2_XS-00001-of-00004.gguf \
  --mmproj Rio-3.5-Open-397B-mmproj-F16.gguf \
  --chat-template-kwargs '{"enable_thinking": true}' \
  --ctx-size 262144 -np 1 \
  --temp 1.0 --top-p 0.95 --top-k 20 \
  --reasoning-budget-message "$reasoning_budget_message"

Credits

• Qwen team for the excellent Qwen3.5-397B-A17B foundation
• Nex-AGI for building on top of it with Nex-N2-Pro
• Prefeitura do Rio / IplanRIO for publishing Rio-3.5-Open-397B
• Nex-AGI / 00INDEX for the public Rio merge-ratio analysis
• @ubergarm for the smol-IQ2_XS recipe
• @bartowski for the Rio 3.5 GGUF imatrix