Real-world
Hybrid orchestration
Concrete workflows, including the tradeoffs and limitations reported by their authors.
this collection
Frontier plans the tickets, Qwen writes the code
A short but clean division of labor: a frontier model decomposes work into tickets, and a local Qwen model implements them on production code, daily.
Local as a verified subagent — but not trusted to touch files alone
A frontier orchestrator dispatching work to a local Qwen 3.8 27B subagent makes longer, harder jobs safe to hand off — but this person draws a hard line at giving any LLM unsupervised control of local files.
A local model and Codex reviewing each other's work, with a hard stop on disagreement
The same Qwen 3.8 27B instance that drives this person's Hermes agent also sits as an MCP tool for Codex, reviewing its proposals and its final output — and the task simply stops if the two can't agree.
A local model decides what's safe to send to the cloud
Before anything reaches a frontier model, a local pass screens it first — for example, checking with high confidence whether an email is non-confidential enough to triage externally.
An overnight local subagent running unfinished work through cron
Hermes uses a multi-GPU local server as a coding subagent, with cron inventorying unfinished work overnight and a cloud model reviewing the result.
Long-running process orchestration with local models
A user relies on local models for multi-hour process orchestration, video workflows, voice tasks and file operations while reserving frontier tools for coding.
Four DGX Sparks running agentic work all day
A four-DGX-Spark cluster under vLLM serves several concurrent agentic coding and research sessions, with unquantized weights and honest notes on concurrency limits.
The minimum setup for a genuinely useful local agent
After experimenting, this user argues a real computer-manipulating agent needs at least a Strix Halo plus an oculink R9700, or a DGX Spark plus a second GPU for auxiliary models.
An orchestrator dispatching to five fresh-context subagents
A custom harness layers a written skill and an orchestrator agent that dispatches tasks to five subagents, each starting from a clean context with hard test gates.
A frontier model orchestrating a fast local coder via nInfer
A small but clear split: the frontier model plans and verifies, while Qwen 3.8 runs on nInfer as the fast local coder, saving tokens.
27B reasoning model on 2x RTX 2080 Ti (NVFP4, vLLM) for agentic orchestration
Qwen3.8-27B (NVFP4 W4A16 + FP8 KV) served on 2x RTX 2080 Ti 22GB over NVLink via a vLLM TP2 fork. 196K context cap, MTP K=3, prefix caching on. Feeds a 1-orchestrator + 7-worker harness: ~1227 tok/s prefill and ~29 tok/s decode solo at 65K, ~192 tok/s aggregate at 8-way. Rejected AWQ/FP8 weights (72% EOS failure), ExLlamaV3 (3-6x slower), and PP=2 (infeasible on 2 GPUs).
GPT-6.1 Sol as cloud orchestrator with one RTX 3090 worker: wall clock and cost for three small builds
A single RTX 3090 running Qwen 3.8 27B Q4 via llama.cpp, orchestrated by GPT-6.1 Sol in the cloud, finished three small 3D games in 43.4 minutes and $0.17 total.
A three-layer AMD ROCm agent architecture for 24/7 local automation
A Radeon 6650 XT runs quantized Qwen models through vLLM and Ollama, while a custom three-layer harness handles private, deterministic automation around the clock.
Building a two-server local AI ecosystem across 17 GPUs
A self-hosted AI ecosystem spans two Linux servers, assigning different GPUs to long-context coding, inference, speech, RAG and media workloads.
Using local models as a hybrid token-saving layer
A user lets frontier models dispatch and correct local models, keeping repetitive task churn off paid cloud inference.
Local code review for changes made by Claude Code
A developer uses Qwen on an old gaming rig with OpenCode to review code produced by Claude Code at work.
Chose cheap cloud coding subs over building a local rig
Tested the larger Qwen models by API, judged them good only as narrowly-briefed sub-agents, and decided a decade of cheap cloud inference beats the hardware bill for now.