lmstudio-model-benchmarks.md

LM Studio Model Benchmarks

Date: 2026-01-25 Platform: Framework Laptop 13 (AMD Ryzen AI Edition)

Hardware Specs

Component	Details
CPU	AMD Ryzen AI 9 HX 370 (12 cores, 24 threads)
GPU	AMD Radeon 890M (integrated, shared memory)
RAM	48 GB DDR5
Kernel	6.19.0-0.0.next.20260123.345.vanilla.fc43.x86_64 (linux-next)
OS	Fedora 43
GPU Override	HSA_OVERRIDE_GFX_VERSION=11.0.0

Models Tested

Model	Architecture	Total Params	Active Params	Size	Tool Use	Context
Qwen3 Coder 30B	MoE (qwen3moe)	30B	3B	18.63 GB	✅	262K
Granite 4 H-Tiny	Hybrid Mamba/Transformer MoE	7B	1B	4.23 GB	✅	1M

Coding Task Benchmarks

Test Descriptions

1. FizzBuzz - Write a Python fizzbuzz function
2. Fix factorial - Debug incorrect base case (return 0 → return 1)
3. Explain lambda - Explain a Y-combinator style factorial lambda

Results

Test	Qwen3 Seq	Granite Seq	Qwen3 Parallel	Granite Parallel
FizzBuzz	3.68s	2.97s	6.01s	6.33s
Fix factorial	5.11s	3.38s	7.97s	7.47s
Explain lambda	4.47s	2.85s	6.30s	6.09s
Total	13.26s	9.20s	20.28s	19.89s
Slowdown vs Seq	-	-	1.53x	2.16x

Tool Calling Benchmarks

Test	Qwen3 30B	Granite Tiny	Speedup
Simple (hello)	1.30s	0.72s	1.8x
Multi-tool selection	5.64s	2.11s	2.7x
Code editing	~3s	~2s	1.5x
Bash execution	6.40s	1.76s	3.6x
Multi-turn context	2.36s	1.57s	1.5x

Key Findings

Performance

• Granite is 1.4-3.6x faster than Qwen3 for most tasks
• Both models handle tool calling correctly
• Parallel execution adds ~50-100% overhead (GPU memory contention)

Behavioral Differences

Aspect	Qwen3 Coder 30B	Granite 4 H-Tiny
Approach	Thorough, parallel tool calls	Direct, sequential
Explanations	Verbose	Concise
Tool selection	Calls multiple tools at once	One tool at a time
Caution level	Checks prerequisites first	Executes directly

Resource Usage

• Both models loaded simultaneously: 22.86 GB
• No OOM errors during parallel inference
• Shared GPU memory works with both models active

Recommendations

Use Case	Recommended Model	Reason
Build/Edit Agent	Qwen3 Coder 30B	Better code quality, parallel tools
Plan Agent	Granite 4 H-Tiny	Fast, 1M context, tool-capable
Code Review	Qwen3 Coder 30B	More thorough analysis

Notes

• Linux-next kernel provides latest AMD GPU support
• HSA_OVERRIDE_GFX_VERSION=11.0.0 required for ROCm compatibility with Radeon 890M
• LM Studio v0.3.x with llama.cpp backend
• All tests used Q4_K_M quantization