Run AI models on your own laptop, GPU or server: setup guides, hardware, benchmarks and when local beats the cloud.
Most read by people in Engineering, IT and Legal & Compliance.
A practical guide to distilling a large LLM into a small one: the four-tries headroom test, tokenizer matching, vLLM memory limits, and when to stop training.
Sage A1.5 is a 35B MoE agentic model from Intern. See the reported 38 GB VRAM with llama.cpp, how it was trained, and how it handled sub-agent tasks.
Hands-on test of DiffuSpace's Clover-35B-A3B-Preview diffusion LLM: bug fix, coding, multilingual and logic puzzle results versus its benchmark claims.
Run DiffuSpace's Clover-35B-A3B diffusion model locally: about 69 GB of BF16 weights, Transformers 5.19.0 setup, block diffusion defaults, controls and SGLang.
LFM3-1.7B is pitched as matching 9B models on agent benchmarks. Here is what the card claims, what hands-on tests showed, and where it fell short.
Learn how to fine-tune Liquid AI's LFM3-1.7B with TRL using SFT, DPO, GRPO and continued pre-training, and which recipe fits JSON, RAG, tools or new languages.
How to run Baidu's Unlimited OCR 2 on your own GPU: about 8.3 GB VRAM at full precision, Transformers and SGLang setup, and test results on real documents.
Run Liquid AI's LFM3-1.7B on a phone or laptop: GGUF, MLX, ONNX, FP8 and NVFP4 checkpoints, sampling settings, tool calling, and where it falls short.
Unlimited OCR 2 scores about 77 on a 100-page document test while most rivals fall under 30. How SR-SWA attention works, plus where the model still fails.
Run 2-bit GSQ-RCO GGUF builds of DeepSeek V4.1 Flash on a 12GB GPU. Covers file sizes, Strata vs llama.cpp support, MoE design and benchmark recovery.
GSQ learns how to round each weight and RCO decides how many bits each layer gets. See how the two shrink a huge MoE model to about 2 bits in a standard GGUF.
Nanbeige4.3 33B used just under 88 GB of VRAM on vLLM at long context, about 62 GB at full precision. Here is how to size hardware and cut memory.
Run the 5B Moss Image model in ComfyUI, SGLang or vLLM-Omni. Covers SFT vs base checkpoints, 50-step defaults, 8-step LoRAs and quantized builds.
Nanbeige4.3 33B tops Qwen 3.8 27B and Gemma 4 on most charts, but looped endlessly in a bug-fix and an HTML test. Here is how the claims hold up.
Moss Image is a 5B open-weight text-to-image model under Apache 2.0. See its text rendering, layout control, benchmark claims, and where it falls short.
Kimi K3.1 Pro ranks fourth on Artificial Analysis and sits under a point behind Opus 5.5 on a coding benchmark. Here is what that means for builders.
Run NVIDIA Nemotron-3.5 Bolt E4B locally with llama.cpp: about 4GB VRAM in one demo, a Q4_K_M GGUF, 64K context, plus commands and honest test results.
Hands-on results for NVIDIA's Nemotron-3.5 Bolt E4B: a bug-hunt agent task that failed, a broken HTML scene, and passes on a riddle, physics math and rostering.
Nanbeige4.3-33B leads 17 of 20 benchmarks against Qwen3.8-27B, Gemma4-31B and Muse-Glimmer-30B. See the full table and where it loses.
A 13-task test of Kimi K3.1 Pro against Claude Opus 5.5 in the same Claude Code harness: where Kimi matches Opus, where it cuts corners, and what to check.
Run FLUX 3 Klein locally in ComfyUI: 5.6B transformer, 16-step distilled model, about 26.4GB VRAM at full precision, and the under-$5M commercial license.
Run Nanbeige4.3-33B with Transformers, vLLM or SGLang. Covers the 512K context, thinking flags, tool-call format, and recommended sampling settings.
A free GPU unlocker patch speeds up prefill on RTX 3090, 4090 and other Ampere/Ada cards in vLLM. Steps for Proxmox/Linux plus before and after numbers.
Kimi K3.1 Pro tested against Opus 5.5 and GPT 6.1 Soul: pricing, open-weight caveats, and hands-on demos from a Rubik's cube to a Fall Guys clone.