AI-Lab Learning &
Experimentation Journal
Hands-on notes, reproducible benchmarks, and curated AI research digests from our private lab playground — shared with friends and peers.
Latest AI & Tech Intelligence
Daily synthesis gathered and analyzed every morning by agent Hermes, reranked by semantic relevance, and summarized for engineering action.
Loading curated digests from edge cache...
Sovereign Model Roster
Hardware-accelerated open-weight foundation models deployed on unified memory architecture.
Qwen 3.8 Flash Next
UD-IQ4_XS Quantization · 125B MoE Backbone
Unified backend for agentic reasoning, autonomous news ingestion, Honcho cognitive memory formation, and OpenWebUI client queries.
BGE-M3 + BGE-Reranker-v2-m3
Dense, Sparse & Multi-Vector Semantic Search
Drives high-relevance filtering of RSS feeds, HN digests, and Honcho memory recall before synthesis.
Gemma 26B & GLM-Flash
Specialized Reasoning & Mathematical Proving
Routinely evaluated against lm-eval suites to benchmark reasoning fidelity against open-source updates.
Model Quality & Benchmark Matrix
Standardized lm-eval evaluations tracking model accuracy across factual knowledge, multi-step reasoning, math, and instruction following.
| Model Identifier | Mode | Knowledge (MMLU-Pro) | Reasoning (BBH CoT) | Math (GSM8k) | Instruct (IFEval) | Tested |
|---|---|---|---|---|---|---|
| Loading benchmark results from edge cache... | ||||||
Edge & Sovereign Lab Topology
Decoupled multi-agent architecture isolating private execution from public edge distribution.
High-Density Inference Engine
Dedicated inference container utilizing hardware passthrough on AMD Strix Halo unified LPDDR5X memory. Serves OpenAI-compatible endpoints with speculative MTP decoding and persistent Radix prefix caching.
Autonomous Hermes Agent
Automated scheduled workflows performing morning RSS/HN ingestion, semantic reranking via local BGE models, daily news digest synthesis, and persistent Obsidian knowledge graph authoring.
Cognitive Memory Engine (Honcho)
Multi-agent dialectic memory tracking user preferences, conclusions, and agent state over time. Provides cross-session continuity across desktop, telegram, and IDE agents.
Scrubbed Edge Distribution
An automated build and scrub gate strips internal network details, IP addresses, and private topologies, pushing sanitized static content directly to Cloudflare Pages for global low-latency reading.