WELCOME TO THE LOCAL SIDE
Small machines.
Big possibilities.
A little desk space. A lot to discover. We put local AI through its paces on real hardware, so you can find the right model for your own corner of the world.
Serious tests. Room to play.
YOUR DESK
IS THE LAB
IS THE LAB
FRESH FROM THE BENCH
GLM-5.3-Flash · SGLang
2× GB10 (DGX Spark head + ASUS GX10 worker) · 256GB total · NVFP4 (LibertAIDAI)
Explore this setup ↗Latest · SWE-bench Verified60.0%2026-10-22
Benchmarks tested4898 scored trials
Leaderboard
20 results · Scientific reasoningScoring & protocols ↗
| Model / setup | Runs on | Evidence | |
|---|---|---|---|
| GPT-6 AstraOpenAI evaluation | 96.0% | Cloud / API | Published ↗ |
| Gemini 3.8 FlashOpenAI comparison | 95.3% | Cloud / API | Published ↗ |
| GPT-5.6 SolOpenAI evaluation | 94.6% | Cloud / API | Published ↗ |
| Gemini 3.1 Pro PreviewGoogle DeepMind evaluation | 94.3% | Cloud / API | Published ↗ |
| Gemini 3.1 Pro PreviewOpenAI comparison | 94.3% | Cloud / API | Published ↗ |
| Claude Fable 5.1OpenAI comparison | 93.7% | Cloud / API | Published ↗ |
| Claude Opus 5OpenAI comparison | 93.7% | Cloud / API | Published ↗ |
| Claude Opus 4.8Anthropic system card | 93.6% | Cloud / API | Published ↗ |
| GPT-5.6 TerraOpenAI evaluation | 92.9% | Cloud / API | Published ↗ |
| Claude Fable 5OpenAI comparison | 92.6% | Cloud / API | Published ↗ |
| GPT-5.6 LunaOpenAI evaluation | 92.3% | Cloud / API | Published ↗ |
| GLM-5.3-Flash NVFP4NVIDIA evaluation | 92.1% | Unspecified | Published ↗ |
| Claude Opus 4.8OpenAI comparison | 92.0% | Cloud / API | Published ↗ |
| Qwen 3.8 27BBF16 reference | 89.2% | Unspecified | Published ↗ |
| Qwen 3.8 Flash-Next · llama.cppUD-IQ3_XXS | 88.4% | LocalGB10 · 128GB | Measured ↗ |
| GLM-5.3-Flash · TensorFoldEXL3 TR3 4bpw | 87.4% | Local2× GB10 (DGX Spark head + ASUS GX10 worker) · 256GB total | Measured ↗ |
| GLM-5.3-Flash · SGLangNVFP4 (LibertAIDAI) | 86.9% | Local2× GB10 (DGX Spark head + ASUS GX10 worker) · 256GB total | Measured ↗ |
| GLM-5.3-Flash · vLLMEXL3 TR3 4bpw | 85.9% | Local2× GB10 (DGX Spark head + ASUS GX10 worker) · 256GB total | Measured ↗ |
| DeepSeek V4 Flash · llama.cppUD-IQ3_XXS | 75.8% | LocalGB10 · 128GB | Measured ↗ |
| Qwen 3.8 27B · llama.cppUD-Q4_K_XL | 73.2% | LocalGB10 · 128GB | Measured ↗ |
Reported scores; setups and evaluation protocols differ. How to compare ↗