LM Studio benchmarks
A desktop app for running local LLMs, with a chat window, a model browser and a local API server.
36 community reports
LM Studio results by GPU
Hardware with reported LM Studio results, with each report's model, quant and speed, newest first. Runs on several cards, with speculative decoding, with batched requests or with part of the model in system RAM say so, since each describes a different setup.
LM Studio on M6 32GB15 reports
- Qwen3.5 9B30.3 t/s
- Gemma 4 31B7.5 t/s
- Gemma 4 12B18.8 t/s
- Gemma 4 26B · 4B active48.2 t/s
- gpt-oss 21B · 3.6B active45.2 t/s
LM Studio on M1 Max 64GB3 reports
- Qwen3.8 27B · Q4_K_M16.9 t/s
- Model not named45.0 t/s
- Gemma 3 12B · QAT 4-bit33.0 t/s
LM Studio on M5 Max 36GB3 reports
- Qwen3.6 35B · 3B active · 4-bit124.5 t/s
- Qwen3.5 9B · 4-bit81.1 t/s
- Qwen3.8 27B · 4-bit26.0 t/s
LM Studio on AMD Strix Halo 128GB2 reports
- Nemotron 3.5 Super 120B · 12B active · Q4_K_M31.0 t/s
- Qwen3.8 27B · Q4_K_M26.5 t/s(speculative)
LM Studio on NVIDIA RTX 4050 6GB2 reports
- Qwen3 30B · 3B active · Q4_K_S14.0 t/s(part in system RAM)
- LFM2.5 1.2B129.0 t/s, 8,500 t/s prompt
LM Studio on NVIDIA RTX 4080 Super2 reports
- Qwen3.8 27B10.0 t/s
- Qwen3.8 27B39.0 t/s
LM Studio on AMD RX 6900 XT 16GB1 report
- Qwen3.8 27B · IQ4_XS19.0 t/s(speculative)
LM Studio on AMD RX 7900 XT1 report
- Qwen3.8 27B7.0 t/s(part in system RAM)
LM Studio on M1 Ultra 128GB1 report
- DeepSeek V4 · IQ3_XXS16.0 t/s
LM Studio on M2 Max 64GB1 report
- Qwen3.6 35B · 3B active · 4bit83.8 t/s
LM Studio on M2 Ultra 192GB1 report
- Qwen2.5 72B · 8 bit7.7 t/s
LM Studio on M5 Max 128GB1 report
- DeepSeek V4 Flash 284B · 13B active · IQ2XXS26.0 t/s, 190 t/s prompt
LM Studio on M5 Ultra 256GB1 report
- Qwen3 14B · 4-bit80.3 t/s
LM Studio on NVIDIA RTX 30701 report
- Phi-3 · q497.0 t/s
LM Studio on NVIDIA RTX 40601 report
- Gemma 4 26B · IQ75.0 t/s, 1,500 t/s prompt(2 cards)
LM Studio on NVIDIA RTX 4060 Ti 8GB1 report
- Gemma 4 26B · 4B active13.0 t/s(part in system RAM)
LM Studio on NVIDIA RTX 4070 Ti Super1 report
- Model not named · UD-Q2_K_XL125.0 t/s(speculative)
LM Studio on NVIDIA RTX 40901 report
- Qwen3.6 27B · Q4_025.0 t/s
LM Studio on NVIDIA RTX 5060 8GB1 report
- Qwen3.6 35B · 3B active · Q4_K_XL30.0 t/s(part in system RAM)
LM Studio on NVIDIA RTX 5060 Ti 16GB1 report
- Gemma 4 8B · Q4_K_M73.9 t/s
LM Studio on NVIDIA RTX 50701 report
- Qwen3.6 35B · 3B active35.0 t/s, 1,000 t/s prompt(part in system RAM)
LM Studio on NVIDIA RTX 50801 report
- Qwen3.8 27B30.0 t/s(3 cards)
LM Studio, GPU not identified2 reports
- Qwen3.8 27B · Q4_K_M31.6 t/s
- Bonsai 2 27B · PQ2_060.0 t/s
LM Studio is a desktop app for Mac, Windows and Linux. It runs models through llama.cpp, and through MLX on Apple Silicon, behind a chat window and a model browser.
It also runs an OpenAI-compatible API server, which makes it a common sandbox before moving to a headless setup.
Compare GPUs running LM Studio
Every card is compared on one model, Qwen3.6 35B · 3B active at 4-bit, the size and bit class that most cards here have plain runs of. A card with no run of that model shows none rather than a figure from a different model.
| GPU | VRAM | Reports | Median speed, Qwen3.6 35B · 3B active 4-bit |
|---|---|---|---|
| M6 32GBapple | 32GB | 15 | Gen 63.8 t/s |
| AMD Strix Halo 128GBamd | 128GB | 4 | no Qwen3.6 35B · 3B active 4-bit run reported |
| M1 Max 64GBapple | 64GB | 3 | no Qwen3.6 35B · 3B active 4-bit run reported |
| M5 Max 128GBapple | 128GB | 3 | no Qwen3.6 35B · 3B active 4-bit run reported |
| M5 Max 36GBapple | 36GB | 3 | no Qwen3.6 35B · 3B active 4-bit run reported |
| NVIDIA RTX 4050 6GBnvidia | 6GB | 3 | no Qwen3.6 35B · 3B active 4-bit run reported |
| NVIDIA RTX 5060 Ti 16GBnvidia | 16GB | 3 | no Qwen3.6 35B · 3B active 4-bit run reported |
| AMD RX 7900 XTamd | 20GB | 2 | no Qwen3.6 35B · 3B active 4-bit run reported |
LM Studio against other engines
Pairs of plain runs on the same card, of the same model size at the same bit class: one device, one request, no speculative decoding, the whole model in memory. Context length and build still differ between the two sides, and each side shows its own.
| Model and GPU | LM Studio | Other engine |
|---|---|---|
| Qwen3.6 35B · 3B active 4-bitM2 Max 64GB | 83.8 t/sone generation report, 256K context | Ollama71.3 t/sone generation report, 32K context |
Top models on LM Studio
Frequently asked
Is LM Studio faster than other engines?
llamaperf has 1 matched pair of LM Studio and other-engine runs on the same card, model size and bit class, but none is comparable enough to say which engine is faster: the sides have too few reports or were run at different context lengths. A pair counts toward a direction only when each side has at least 3 generation reports with a stated context, and the two sides' contexts overlap. The measured figures and their conditions are in the table on this page.
Is LM Studio free?
Yes. It's a closed-source desktop app that's free to download and use.
What's the difference between LM Studio and Ollama?
LM Studio leads with a graphical chat window and model browser. Ollama leads with a command line and a background API. Both run GGUF models, and LM Studio can also run MLX models on a Mac.