- 🛰️ Building Vergilius, a local OSINT assistant: open sources, a small model on your own GPU, nothing leaves the machine
- 📘 Wrote the NVIDIA GTX 1080 (sm_61) AI Full Guide: everything I made run fast on a 2016 GPU with 8 GB
- ⚡ Running Qwen3.8 Flash-Next with 4 users × 227k tokens on one RTX PRO 6000: blackwell6000-qwen3.8-flash-next
- 🤝 Open to collaboration on inference kernels, speculative decoding and local agents
- 🎲 Fun fact: my GTX 1080 from 2016 runs a ternary 9B model at 84 tok/s
chatroom for Claude Code and Codex: a multi-agent harness built on this year's research on autonomous research agents.
- Research mode: fixed roles (workers, proposer, auditor, registrar, searchers), phases with gates, a pre-registered contract, and a kernel that checks every proposal: sealed evaluator, a worktree per worker, repeated measures, every attempt tagged, a final run on held-out data.
- Harness enforced by code: a live group chat without turns, instant alerts, a task board with leases, handoffs and loans, closing only with evidence, detectors for stalled or unproven work.
- Friends: share your office over a tunnel with a PIN, send your agents into a friend's common room, visit each other's office and watch together.
- Spawn: summon free models from OpenRouter, NVIDIA NIM, Groq, Google AI Studio, Mistral, Cohere, Cloudflare and Hugging Face as roles. A role keeps its mission and notes; when its context fills up it writes a handover, and when a host runs out of quota it moves to another one. Long tool output is compressed for every agent.
- A 3D voxel office where you watch each agent type, read, wait for a permission or take a break.
Influenced by (24 papers)
- The Agentic Scientist: Orchestrating Coding Agents for Automated Scientific (Co-)Discovery
- Harness-Search: Guiding Long-Horizon Search through Multi-Agent Coordination
- PatchBoard: Schema-Grounded State Mutation for Reliable and Auditable LLM Multi-Agent Collaboration
- MemTX: Transactional Belief Commit for Stateful Agent Memory
- SwarmResearch: Orchestrating Coding Agents for Open-Ended Discovery
- CORAL: Towards Autonomous Multi-Agent Evolution for Open-Ended Discovery
- ARIS: Autonomous Research via Adversarial Multi-Agent Collaboration
- Accelerating scientific discovery with Co-Scientist
- IdeaScientist: Orchestrating Agents for Grounded Scientific Ideation
- ScienceFlow: A long-horizon agent for ML research, scientific discovery and beyond
- OneDayAgent: Towards a Long-Horizon Harness for Autonomous Agents
- Rethinking Multi-Agent Collaboration: When More Is Less
- AI Research Agents Narrow Scientific Exploration
- Autonomous Research Agents: A Survey of AI Scientists and the Verification Gap
- ARAC: Benchmarking Auto-Research's Alignment and Completeness on End-to-End Researchs
- A Vocabulary for Multi-Agent Automated Research Systems
- Position: AI Agents in Scientific Teams Should Be Studied as Human-Agent Systems
- Benchmarking Open-Ended Multi-Agent Coordination in Language Agents
- Harnessing the Collective Intelligence of AI Agents in the Wild for New Discoveries
- Mixture of Complementary Agents for Robust LLM Ensemble
- Beyond Consensus: Trace-Level Synthesis in Mixture of Agents
- When Does Combining Language Models Help? A Co-Failure Ceiling on Routing, Voting, and Mixture-of-Agents Across 67 Frontier Models
- Are Diversity Metrics Measuring Diversity? A Capability-Controlled Audit of Majority-Vote Gain in LLM Ensembles
- What Do Evolutionary Coding Agents Evolve?
| Project | Pull request | Status |
|---|---|---|
| stable-diffusion.cpp | #2202 LoRA baked_scale fix |
 |
| PrismML llama.cpp | #368 Pascal PQ2_0 mat-vec, +17% | |
| vLLM | #58863 exact GDN replay commit, adopted in the PR |  |
| llama.cpp | #27721 CUDA graphs on Pascal, +40% MoE |  |
| Model | Downloads (last 30 days) |
|---|---|
| Swift-1.5-Qwen3.8-Flash-Next-MTP |  |
| MiniCPM5-2B-DSpark-FRSpec32k-GGUF |  |
| MiniCPM5-2B-pjev-LoRA |  |
| TernaryPilot-v2-Computer-Browser-Use-LoRA |  |
| TernaryDecider-v1-Decision-LoRA |  |
