Universal smolagents Hardware Advisor & Hugging Face Explorer

Universal Local LLM Hardware Recommender

Find the optimal open-weight models (Qwen 3.8 27B, Qwen 3 14B/8B, Qwen 3 Coder, Llama 3.3, DeepSeek-R1) your computer can run with peak performance and zero cloud dependence.

⚙️ Your System Specs

QUICK HARDWARE PRESETS:

Operating System & Architecture

Total System RAM 16 GB

Dedicated / Unified GPU VRAM 12.0 GB

Primary Optimization Focus

THE #1 SWEET SPOT RECOMMENDATION

Qwen 3 14B Instruct 14B

Next-generation frontier-level instruction following and coding depth. Runs entirely in your 12 GB memory budget with high token throughput and zero swap latency.

🎯 Model Spotlight: Qwen 3.8 27B (`qwen3.8:27b`)

Running 27B on 16GB Apple Silicon: The 27.8B weights require ~16.9 GiB memory. If Ollama reports memory limits on a 16GB Mac, temporarily raise the macOS wired memory ceiling with:
sudo sysctl iogpu.wired_mem_limit=14336
Or use ollama run qwen3:14b for an instant 60+ tok/s daily companion!

Usable LLM Memory 12.0 GB

Recommended Quant Q4_K_M (~10.3 GB)

Max Safe Context 32,768 tokens

🎯 Top Category Picks

General Assistant

Qwen 3 14B

Full 128k context support, massive leap in nuance and accuracy.

ollama run qwen3:14b

Coding & Engineering

Qwen 3 Coder 30B

Next-gen software architecture, multi-file refactoring and debugging.

ollama run qwen3-coder:30b

Reasoning & Math

DeepSeek-R1-Distill-14B

Visible chain-of-thought for hard algorithms, proofs, and logic.

ollama run deepseek-r1:14b

Instant / Lightweight

Qwen 3 8B

Instantaneous response speed; superior intelligence in a 6.4 GB footprint.

ollama run qwen3:8b

🌐 Live Open-Weight Models on Hugging Face

Real-time models fetched directly from huggingface.co/models, analyzed against your host memory.

Model / Repository	Generation & Parameters	Est. Min RAM (Q4)	Downloads	Your System Compatibility	Run / Open

🤖 Ask the Local LLM Hardware Advisor

Ask questions about running Qwen 3.8, Qwen 3, Llama 3.3, memory limits, macOS wired memory, or quantization tradeoffs.

👋 Hello! I am your Universal Local LLM Advisor. Ask me anything like:
• "Can I run Qwen 3.8 27B on my 16GB M4 Mac?"
• "What's the difference between Qwen 3 14B and 27B?"
• "How do I fix 'model requires 16.9 GiB but only 11.3 GiB are available' in Ollama?"

Quick Questions:

📊 Next-Gen Open-Weight LLM Sizing Matrix

Verified memory footprints including ~1.5 GB KV context cache for 16k-32k tokens.

Model Family	Params	Q4_K_M VRAM	Q8_0 VRAM	Min Hardware Tier	Ollama Command
Qwen 3	4B	~3.8 GB	~5.8 GB	Any PC / Mac / 6GB+	`ollama run qwen3:4b`
Qwen 3	8B	~6.4 GB	~10.2 GB	8-16 GB Mac / RTX 4060	`ollama run qwen3:8b`
Qwen 3	14B	~10.3 GB	~16.5 GB	16-24 GB Mac / RTX 4070	`ollama run qwen3:14b`
Qwen 3.8	27B	~18.8 GB	~31.0 GB	16-36 GB Mac (wired mem) / RTX 3090	`ollama run qwen3.8:27b`
Qwen 3 Coder	30B	~20.7 GB	~35.0 GB	24-36 GB Mac / RTX 4090	`ollama run qwen3-coder:30b`
DeepSeek-R1 Distill	14B	~9.1 GB	~15.6 GB	16-24 GB Mac / RTX 4070	`ollama run deepseek-r1:14b`
DeepSeek-R1 Distill	32B	~20.2 GB	~34.8 GB	24-36 GB Mac / RTX 3090/4090	`ollama run deepseek-r1:32b`
Llama 3.3	70B	~42.5 GB	~76.0 GB	48-64 GB Mac / Dual RTX 4090	`ollama run llama3.3:70b`

💻 Run the Agent Directly on Your Physical Machine

The local agent automatically detects physical sysctl memory on Mac, nvidia-smi on Linux/Windows, queries live Hugging Face models, and checks local Ollama model states without any manual configuration.

1. Clone & Run with UV (Recommended):

git clone https://huggingface.co/spaces/maubilli/best-local-model && cd best-local-model && uv run python src/agents/bestLocalModel.py

2. Run Installed Models via Ollama:

ollama run qwen3.8:27b

💡 Key Features of the Local Agent:

Live Hugging Face model integration with automated disk caching.
Dynamic model resolver: Checks Ollama for installed models (like qwen3.8:27b and qwen2.5-coder:7b) and tests responsiveness.
Zero cloud lock-in: Fully autonomous smolagent reasoning cycle executed locally.