rtx-4070

run a 27b local ai on your gaming gpu, rtx 3060 and up, radeon rx 6600 and up in beta, 8gb or more, linux and windows, one line to install

Bonsai 2 27B: 262k context (trained max) on 12 GB, 54 to 97 tok/s decode, 2x prefill, q8_0 KV recipe, grammar tool calls 9/9. Same ternary weights. Windows bundle + receipts.

ComfyUI nodes for MiniMax H3: INT8 X2 VAE and asynchronous NVENC saving. 5-second 1088x1920 video with audio in about 60 s on an RTX 4070 12GB, about 30% faster than the FP16 X2 VAE.

MiniMax H3 audio-first two-pass for ComfyUI: reduces the metallic 'talking in a pipe' voice by generating audio first and locking it. The fast LBH variant is about 18% faster than the standard 4-step workflow on an RTX 4070 12GB.

Faster MiniMax H3 video for ComfyUI: 3 low-resolution steps, LBH latent upscale, then 1 high-resolution step. 14 s 1088x1920 video with audio: 154.8 s -> 116.6 s (-25%) on an RTX 4070 12GB.

Measured llama.cpp and quantization results on Tesla V100 (sm_70), Pascal GTX 1070, and RTX 4070 - hardware most projects don't test on

Reproducible Qwen3.5 MTP and D-Flash benchmarks on an RTX 4070: throughput, acceptance rate, quantization, and language effects.

Workflows de ComfyUI para vídeo con IA (Wan 2.2) en una gráfica de 12 GB · Capitán Tokn

Prebuilt CUDA wheels (sm_89) + build guide to run Microsoft TRELLIS.2 and TencentARC Pixal3D image-to-3D in ComfyUI on NVIDIA RTX 40-series (Ada) under Debian 13 / Linux. Fixes the 'no kernel image is available' (sm_120 vs sm_89) problem.

CachyOS/Xorg black screen fix for MSI Stealth A16 AI+ AMD Radeon 880M + NVIDIA RTX 4070 hybrid graphics.

Local EDU AI LAB for looped (latent-recurrent) transformers - run, patch and measure loop counts on local GPU

Локальный LLM-провайдер для проекта slovo — Ollama c Laguna XS.2 (33B/3B MoE) в Docker с проброшенной NVIDIA GPU. Оптимизирован под single-user агентские сессии на RTX 4070 Ti SUPER + i9-11900K.

Claude Code on a local 125B MoE (Qwen3.8-Flash-Next) on a 12 GB GPU: setup and measurements, a local agent that drives ComfyUI, and VS Code Remote-SSH fixes

Fork of karpathy/nanochat pushed onto one RTX 4070 12GB: a 910M-param base model from 530 GPU-hours, OOM-safe training automation, and 14 documented negative results.