0xSero

DeepSeek-V4.1-Flash on two DGX Sparks: EXL3 routed experts, 262k context, 2M-token KV, vision, tools, MTP, bundled with Pi

Point any coding harness (Claude Code, Codex, OpenCode, Pi, OMP, Crush, Copilot, Grok…) at any OpenAI/Anthropic/Responses-compatible inference endpoint. Core library + CLI + macOS tray + web UI.

Wiki for the GLM-5.3 Flash low-bitrate TR3 quantization campaign: architecture, EXL3/TR3 encoding, K2/K3/K4 tiers, validation gates, and quality metrics.

GLM-5.3 Flash on one DGX Spark: M288 mixed-precision mosaic (mul1) + native MTP (mcg) serving recipe with measured evidence

Pinned Local Inference Lab R35 GLM-5.3-Flash Docker deployment for 4x RTX PRO 6000: vision, video, 400k context, image history pruning and measured evidence

Practical model evaluation and compression tools: Terminal-Bench, DeepSWE, GPQA, KL divergence, REAP, EXL3, and reproducible evidence.

Local AI data: every model, build, hardware record, price, benchmark and community measurement behind the Local AI registry (moved from local-ai-registry, full history)

Agent skill: set up and run 1 to 4 NVIDIA DGX Sparks. Linking, best models per fleet size, recipes, fixes.

Open any coding agent on a Hugging Face Inference Providers model, from one button on the Omarchy bar

Six galleries of public-domain paintings with a short line of scripture set into each plate

Augustine of Hippo: his writing broken into six arcs, with every passage quoted verbatim from public-domain sources

Serve StepFun Step-5-Preview (EXL3) on four NVIDIA DGX Spark with vLLM TP4, plus Pi integration

Observation-driven dynamic pruning: configure what a model loses from external observations. Verification-first system design, schemas, and minimal reference code.

Source-checked dates for the 'A Decade of AI Hardware Innovation' chart