Mixture-of-Experts at frontier scale
397B parameters split across specialized experts, with a router that activates just 17B per token. You get the capacity of a massive dense model at a fraction of the compute cost.
prefeitura-rio / Rio-3.5-Open-397B
Rio-3.5-Open397B
An open, MIT-licensed mixture-of-experts model from Rio AI Labs. 397 billion parameters, 17 billion active per token, and a one-million-token context window built for long-context reasoning, coding, and multilingual work.
MIT LICENSE · MIXTURE-OF-EXPERTS · 1M CONTEXT
moe-router · top-k routing17B / 397B ACTIVE
Each token activates a sparse subset of experts. Only 17B of 397B parameters fire per token, so inference stays light while capacity stays large.
397BTotal parametersFull mixture-of-experts capacity
17BActive per tokenSparse, token-efficient inference
1MContext window1,010,000 tokens
MITLicenseOpen weights, free to build on
WHY RIO-3.5-OPEN
Built around a sparse expert architecture and a million-token window
One-million-token context
A 1,010,000-token window holds entire codebases, document sets, and long conversations in memory, so the model reasons over the whole problem instead of a fragment of it.
SwiReasoning integration
Built on the Qwen3.5-397B-A17B base with SwiReasoning, the model switches reasoning depth on demand to stay accurate on hard problems without wasting tokens on easy ones.
Token-efficient inference
Sparse activation and adaptive reasoning keep latency and cost low. Run long-context workloads in production without paying for parameters every token doesn't need.
MODEL CARD
Specifications at a glance
- Model name
- Rio-3.5-Open-397B
- Developer
- Rio AI Labs (Prefeitura do Rio de Janeiro)
- Parameters
- 397B total · 17B active (Mixture-of-Experts)
- Context window
- 1,010,000 tokens (1M)
- Architecture
- Based on Qwen3.5-397B-A17B with SwiReasoning integration
- Capabilities
- Long-context reasoning, coding, multilingual tasks, token-efficient inference
- License
- MIT
CAPABILITIES
What it's good at
Long-context reasoningHold and reason over 1M tokens
CodingWhole-repo understanding
MultilingualStrong cross-language tasks
Efficient inference17B active per token
TRAINED BY RIO AI LABS
A public research effort from the Prefeitura do Rio de Janeiro
Rio AI Labs is the applied AI group of the Prefeitura do Rio de Janeiro. Rio-3.5-Open-397B is its open release: a frontier-scale mixture-of-experts model trained for efficient, long-context reasoning and shipped under a permissive MIT license so anyone can run, fine-tune, and build on it.
MIT licensedOpen weightsResearch-ledMultilingual
Run Rio-3.5-Open-397B today
Weights, the model card, and usage details live on Hugging Face. MIT licensed and ready to deploy.
Open the repositoryhuggingface.co/prefeitura-rio/Rio-3.5-Open-397B