With 32–64 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
Coding a SaaS product on an M1 Max, one well-scoped feature at a time
Running Qwen 3.8 27B on the same Mac they work on, this developer credits per-feature domain glossaries and tight scoping — not the model — for making local coding reliable.
Using a local model to stress-test the agent a coding agent is building
An unusual layered setup: the local model isn't the coding agent — it's the thing an outer coding agent uses to build and stress-test a separate 'inner' agent.
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.
Qwen 27B coding and infrastructure on two 3090s
A two-3090 local setup runs Qwen 27B for coding and infrastructure work, including bug fixing and quality-of-life automation.
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.
Coding and homelab maintenance on a dual-R9700 server
A garage server with two Radeon Pro AI R9700 cards supports coding, homelab maintenance and an evolving Cline/Open WebUI workflow.
Two RTX 5090s running helpdesk and ops for a small office
An office replaced roughly $2,000 a month of API consumption with two RTX 5090 machines running Qwen 3.8 27B under vLLM, one for helpdesk and one for ops.
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.
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).
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.
Benchmarking Strata n-gram table placement on a modded RTX 4090 48GB with 128GB DDR5
A measured benchmark on a modded 48GB RTX 4090 shows Strata's n-gram table belongs on SSD while conversation parking delivers the large speedup.
A Tesla V100 32GB and an RTX 3090 on Strata: Qwen 3.8 Flash Next at 262K context for local business management
Two single-GPU Ubuntu servers, one with a blower-modded Tesla V100 32GB and one with an RTX 3090, run Strata with Qwen 3.8 Flash Next at 262K context for local business management automation.
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.