0xSero

AI coding agent practitioner operating at the intersection of model optimization, data extraction, and long-running autonomous workflows. Runs Sybil Solutions ("Building production systems with AI"). Active on HuggingFace publishing REAP-pruned expert models calibrated specifically for code generation and tool calling.

What They Build

REAP-Pruned Models

Publishes expert-pruned MoE models on HuggingFace using Cerebras' REAP (Router-weighted Expert Activation Pruning) methodology. Each model is calibrated on three datasets: evol-codealpaca-v1 (code generation), xlam-function-calling-60k (tool calling), and SWE-smith-trajectories (agentic multi-turn coding). The calibration target is telling: they're optimizing not for benchmark scores but for the specific distribution of agentic coding work.

Models include GLM-4.7-REAP-45 (45% pruned, 197B active), Qwen3.5-264B-REAP, and gemma-4-21b-a4b-it-REAP — some with INT4 quantization bringing a 358B model down to ~108GB.

ai-data-extraction Toolkit

A zero-dependency Python toolkit (680 stars) that pulls complete conversation history from eight AI coding tools: Claude Code, Codex, Cursor, Windsurf, Trae, Continue, Gemini CLI, and OpenCode. Outputs standardized JSONL with messages, code context, diffs, tool use, and metadata. Built for ML training pipelines — the README includes Unsloth fine-tuning examples.

This is the unglamorous plumbing of the "train your own coding model" pipeline. Before you can fine-tune, you need clean data from your actual workflow. Most people talk about training; 0xSero built the extraction layer.

Parchi

Chrome extension supporting BYOK (bring your own key) workflows for very long-running autonomous agent tasks. Cited in AINews' March 2026 agent harness roundup alongside the broader convergence on parallel, long-running agent stacks. Integrates with their pruned models (Qwen3.5-REAP).

Key Themes

• #tool — ai-data-extraction fills a real gap: nobody else built a unified extractor across the fragmented AI coding tool landscape
• #pattern — Model optimization calibrated to agentic workloads, not generic benchmarks. The calibration dataset IS the product spec
• #concept — BYOK workflows: run autonomous agents on your own keys, your own hardware, your own models. The anti-SaaS posture applied to agents
• #person — Practitioner, not theorist. Builds the plumbing, publishes the models, teaches live

Critical Analysis

What's interesting: 0xSero represents a specific archetype that's emerging — the agent infrastructure practitioner who works below the framework layer. They're not building Yet Another Agent Harness; they're optimizing the models that agents run on and building the data pipelines that feed those optimizations. This is the "pick and shovel" layer of the agent economy.

What's undersold: The ai-data-extraction toolkit is more significant than its 680 stars suggest. Training data extraction from proprietary coding tools is a hard problem that most teams solve with bespoke scripts. A standardized, multi-tool extractor is infrastructure. The fact that it's MIT-licensed and zero-dependency suggests it's meant to be absorbed into other projects rather than monetized directly.

The REAP calibration choice is telling. Most model publishers optimize for HumanEval or MBPP. 0xSero calibrates on SWE-smith-trajectories — multi-turn agentic coding sessions. This is a bet that agentic coding has a different token distribution than single-turn code completion, and that models should be optimized accordingly. It's the right bet, but it also means these models may underperform on standard benchmarks, which creates a discoverability problem.

The homelab matters. The 8x 3090 rig blog post signals a commitment to local inference that goes beyond cost savings. If you're extracting training data from your own coding sessions and fine-tuning models on your own hardware, you've closed the loop. No external API sees your code, your conversations, or your models. This is the privacy-maximalist position made practical.

Connections

• Local and Open Source Inference — The 8x 3090 homelab is a concrete instance of the local inference thesis
• Personal Agents — Parchi's BYOK model is the personal agent philosophy applied to autonomous workflows
• Agent Memory and Context — ai-data-extraction treats coding tool history as training data; connects to the memory-as-training-data idea
• Scaling Long-Running Agents — Parchi tackles the long-running autonomous agent problem from the harness side
• Agent Orchestration — BYOK workflows intersect with multi-agent coordination patterns
• Three Tier Memory — Extracting conversation history for training is the cold tier of the memory pipeline
• Building low-level software with only coding agents — Similar practitioner profile: shipping real infrastructure, not demos
• claude-ctrl — "An instruction in context is not a constraint" — REAP calibration is the model-level equivalent

Sources: Tweet by @0xSero (unfetchable) Last updated: 2026-05-15