Open-Weight LLM · Private & Custom AI
Rio-3.0-Open-Mini
A 4B distilled reasoning model optimized for private deployment in ops workflows requiring math, code, and structured problem-solving without external API dependency.
Rio 3.0 Open Mini is a 4B-parameter transformer distilled from Qwen3-4B-Thinking, enhanced with SwiReasoning—a training-free inference technique that switches between explicit and latent reasoning to improve both accuracy and token efficiency. Built by Rio's municipal IT company, it's MIT-licensed and designed to run on modest hardware while maintaining competitive performance on STEM and code tasks, making it a practical fit for companies building private reasoning layers into ops automation.
Model facts
Private deployment
Run Rio-3.0-Open-Mini in your own environment
Deployable entirely on-premise via standard transformers, vLLM, or SGLang. The 262K token context and ~4B parameters fit comfortably in 8–16 GB VRAM (FP16/BF16), enabling a single GPU or multi-GPU setup without cloud dependency. Data stays in your environment—a structural benefit for finance, legal, or engineering teams processing sensitive calculations or proprietary code.
Operational AI use cases
Finance & Budget Analysis
Automated parsing of departmental expense reports, variance analysis, and forecasting. The model's strong math performance enables root-cause detection in budget anomalies and generation of monthly financial summaries without routing raw financial data to external APIs.
Engineering & Code Review Triage
Internal pipeline to classify and summarize code pull requests, identify technical debt patterns, and flag compliance issues in proprietary codebases. SwiReasoning's token efficiency keeps inference costs low for high-volume triage.
Operational Reporting & Knowledge Synthesis
Digestion of multi-source operational logs, incident reports, and process documentation to generate executive summaries and identify systemic patterns. The 262K context window accommodates large document sets in a single pass.
Custom AI
As a base for custom AI
Strong foundation for building proprietary reasoning agents. Teams can fine-tune or prompt-engineer Rio 3.0 Open Mini into domain-specific reasoning systems (compliance bots, diagnostic assistants, technical workbench tools) without yielding model weights or inference data to third parties. The MIT license permits both commercial products and internal custom applications.
In the operating system
Where it fits
Operates as the reasoning core in a private AI operating system—handling complex inference logic in knowledge retrieval, agentic decision-making, and workflow automation layers. Its efficiency enables real-time or near-real-time response in multi-step operational tasks where larger models would be cost-prohibitive or latency-incompatible.
Data control & security
Self-hosting eliminates data transmission to external LLM providers. Sensitive input (financial records, code, internal documents) remains in your infrastructure. Note: model security and compliance assurance depend on your deployment architecture, access controls, and audit practices—not guarantees from the model itself. Suited for environments with data residency or confidentiality requirements.
Hardware footprint
Estimate: ~8 GB VRAM (FP16), ~16 GB (FP32). Multi-GPU configurations (vLLM tensor parallelism) scale to larger throughput on 4–8 GPUs. Inference latency and token throughput depend on hardware tier and SwiReasoning configuration (explicit vs. latent mode trade-off).
Integration
Standard transformers API via Python; production serving via vLLM or SGLang with multi-GPU support. Chat template support enables drop-in integration into conversational workflows. Inference frameworks expose temperature, top_p, and max_tokens controls for fine-tuning reasoning behavior. Connect via REST API wrappers or direct Python bindings for internal systems.
When it's not the right fit
- —Retrieval-augmented generation or long-context grounding: while the 262K window is large, performance on factual accuracy or knowledge cutoff tasks depends on training data; use supplementary retrieval for high-stakes factual queries.
- —Real-time, sub-100ms latency requirements: 4B models on consumer/mid-tier GPUs will incur inference overhead; latency-critical applications may need quantization or distillation.
- —Multi-lingual code-switching at production scale: model supports many languages, but primary evaluation is English/Portuguese/Chinese; performance on low-resource language mixes is unknown.
- —Streaming and interruption: SwiReasoning's dynamic switching may introduce latency variance; streaming token-by-token in latent mode is not standardized.
Alternatives to consider
Qwen3-4B-Thinking-2507 (base model)
Same footprint, fewer distillation gains (~6–8% lower math/code benchmarks). Useful if you want to fine-tune from the unmodified base or need maximum interpretability.
DeepSeek-R1-Distill-Qwen-4B
Comparable 4B footprint with chain-of-thought distillation. May offer different latency/accuracy trade-off; check your benchmarks and inference framework compatibility.
Llama 3.1 8B
Larger, wider model family, stronger on general tasks and more mature ecosystem. Trade-off: ~2x VRAM, broader utility but not specialized for reasoning-heavy ops workflows.
Related open models
FAQ
Can I run Rio 3.0 Open Mini entirely on-premise without calling external LLM APIs?
Yes. Download the model weights (via HuggingFace), deploy on your infrastructure using transformers/vLLM/SGLang, and serve via local REST API or Python bindings. All inference stays in your environment.
Is the MIT license sufficient for commercial applications?
Yes. MIT permits commercial use, modification, and redistribution with proper attribution. No royalties or vendor lock-in. Verify any custom training or fine-tuning doesn't introduce conflicting dependencies.
How much does SwiReasoning improve inference efficiency in production?
The model card shows 2–17% accuracy gains with SwiReasoning enabled. Token efficiency improvements depend on problem difficulty and confidence thresholds; not specified quantitatively. Test on your workloads.
What's the difference between Rio 3.0 Open Mini and the larger Rio 3.0 Open?
Rio 3.0 Open Mini (~4B) is distilled from Rio 3.0 Open (larger, unlisted size). Mini is cheaper to run and fits tighter deployments; Open has higher accuracy (85.1% GPQA vs. 71.9%). Choose Mini for cost-sensitive ops, Open for maximum reasoning quality.
Build Proprietary Reasoning into Your Ops Stack
Rio 3.0 Open Mini is ready to run on your infrastructure. Let LLM.co help you wire it into workflows, fine-tune for your domain, and monitor private reasoning at scale. Start your custom AI project today.