With 25–32 GB,
here is what really runs.
Experiences reported by people using their machines every day. The figure is the reported setup capacity; multi-GPU and unified-memory layouts can behave differently.
in this range
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.
Daily agent workflows on an RTX 5090
An RTX 5090 runs recurring agents for project maintenance, site contributions and personalized news monitoring.
A dedicated local editing model for novels, running on two 5060 Ti cards
A writer runs Qwen 3.8 27B locally on two RTX 5060 Ti cards with a model card fine-tuned only to edit books and detect plot holes, not to write.
One R9700 with card-specific kernels for local coding
A single Radeon Pro AI R9700 runs coding work using Lucebox and a recommended 4-bit quant, relying on custom optimized kernels for the card.
Legal work: a cloud planner driving local anonymisation, OCR and RAG
A legal user has Claude build task plans while a local Qwen 27B executes them, with anonymisation, OCR for difficult documents and a RAG pipeline using nomic embeddings.
Three concurrent coding sessions at 150 tok/s on one 5090
A developer with 20 years of experience runs three concurrent coding sessions at about 150 tok/s each on a single RTX 5090 and has been cloud-free for three months.
Two Radeon RX 9070 XT cards as a 32GB budget rig
A compact configuration note: two RX 9070 XT cards give 32GB total VRAM and 50-60 tok/s at a 200K context on Qwen 3.8 27B in MXFP4.
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.
Coding locally with Qwen 3.8 27B Q6 on an RTX 5090: 160k context, ~80 tok/s, Hermes harness
The author runs Qwen 3.8 27B at Q6 with a 160k context and vision on an RTX 5090, using the Hermes agent harness and SearxNG in Docker, and reports around 80 tok/s but about 15 minutes for a one-shot website.
Building a full SaaS product on Qwen 3.8 27B with vLLM
Switching from llama.cpp to vLLM on AMD hardware was the unlock for running a complete SaaS build loop through a local model.
Company email automation kept off the cloud for legal reasons
This company believes uploading user data to a cloud model would be illegal for their use case, so email and user-text automation runs on a Gemma 4 build on a single Radeon Pro AI R9700.
A daily-refreshed RAG advisor built from your own infrastructure
A Python collector pulls read-only configs, firewall rules, and network health data every day and feeds it into a local RAG pipeline, so the advisor always knows the current state of the homelab without ever being able to touch it.
Dual Radeon Instinct MI50 benchmarks: dense models at 12 to 18 tok/s, MoE models at 47 to 61 tok/s
Two Radeon Instinct MI50 cards limited to 145W each are benchmarked with llama.cpp Vulkan across dense and MoE models, with MoE models far ahead of dense ones.
Compare with another budget.
Every capacity guide, or start from your own machine and a task.