iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
You can run the WUIC Assistant with generation supplied by a local Ollama model, retrieval kept in the WUIC backend, and VS Code providing the agent loop. The documented 30B setup targets an RTX 4090-class GPU with 24 GB of VRAM, uses about 19 GB of disk for the Q4_K_M model, and connects through a stdio MCP server called wuic-rag. After the one-time hardware and configuration work, model calls have no per-token charge and your prompts and code can remain on your LAN.
This is not a self-contained autonomous programmer. WUIC supplies codebase knowledge and model responses; VS Code decides when to call tools, edit files, and run commands. For difficult, long-running, multi-file tasks, a hosted frontier model can still be more reliable. The practical setup is local Ollama for frequent questions and scoped edits, with a hosted model reserved for work that needs stronger long-horizon reasoning.
| # | Preview | Product | Price | |
|---|---|---|---|---|
| 1 |
|
VIPERA NVIDIA GeForce RTX 4090 Founders Edition Graphic Card | $5,199.99 | Buy on Amazon |
What the local WUIC setup actually runs
There are three cooperating pieces rather than one large application:
- WUIC backend: indexes and retrieves your code locally. Its retrieval pipeline combines BM25, bge-m3 embeddings, and a fine-tuned reranker implemented in native .NET/ONNX.
- Ollama: hosts the language model and exposes an OpenAI-compatible endpoint. Generation moves from a cloud API to this service.
- VS Code and WUIC Assistant: provide the chat interface and agent loop. The editor decides how to plan, edit files, and execute commands, while the MCP server supplies WUIC tools and retrieved context.
The backend and Ollama can share one computer or run on separate machines on the same LAN. In the separate-host topology, the backend and VS Code call Ollama on a GPU machine at port 11434. VS Code reaches the backend through the stdio MCP server named wuic-rag. MCP (Model Context Protocol) is an open standard for giving an LLM client callable tools; it does not itself provide planning or autonomy.
Popular on Amazon
- 16,384 NVIDIA CUDA Cores
- Supports 4K 120Hz HDR, 8K 60Hz HDR and variable refresh rate as indicated in HDMI 2.1A
- New streaming multiprocessors: up to 2x power and power efficiency
- Fourth generation tensor cores: up to 2x AI power
- Third-generation RT cores: up to 2x ray tracing performance
A useful mental model is: retrieval finds the relevant files, Ollama writes the response, and the editor remains responsible for acting on that response. This division is why a local model can answer questions about a private repository without sending the repository to a hosted API, while still allowing the normal VS Code workflow of reviewing diffs and approving commands.
Requirements and hardware choices
Recommended 30B configuration
| Item | Documented value | Why it matters |
|---|---|---|
| GPU | 24 GB VRAM, RTX 4090 class | Target for the qwen3-coder-wuic 30B Q4 setup |
| Model storage | Approximately 19 GB | Size of the GGUF at Q4_K_M quantization |
| Model architecture | MoE 30B, about 3B active parameters | Most weights exist in the file, but each token activates a smaller expert subset |
| Measured throughput | About 197 tokens/s | WUIC Framework figure on a 24 GB GPU; actual speed varies with prompt, context, and system |
| Node.js for MCP | 14.18 or newer | Required by the stateless wuic-rag server |
The official 24 GB guidance calls for flash attention, a q8 key-value cache, a 49,152-token context length, and indefinite keep-alive. These settings trade memory for responsiveness: keeping the model resident avoids reload delays, while the long context leaves room for retrieved code and conversation history. Watch ollama ps after loading. If it reports CPU/GPU splitting, lower the context length, for example to 32,768, until the model fits predominantly in VRAM.
When a smaller GPU is more sensible
The documented alternatives are qwen2.5-coder:14b and gpt-oss:20b. They reduce memory pressure, but the WUIC comparison notes weaker performance on complex tool schemas. Choose one when buying a 24 GB card is not justified, and expect to validate tool calls and edits more carefully. A smaller model can still be useful for repository questions, explanations, and tightly scoped changes.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
One machine or two
- Single machine: simplest networking. Point the backend at
http://localhost:11434/v1and keep Ollama bound to the local interface. - Dedicated GPU host: put Ollama on the machine with the 24 GB card and leave the backend and VS Code on another workstation. Configure Ollama to listen on
0.0.0.0:11434, open only the required LAN firewall path, and use the GPU host’s LAN address in the backend and extension settings.
Binding to all interfaces is a network decision, not a performance setting. Do not expose the Ollama port directly to the public internet; restrict it to trusted addresses with the host firewall or a private network.
Install the WUIC model in Ollama
Use Ollama’s Hugging Face import and then create the short name expected by the WUIC configuration.
- Pull the Q4_K_M model:
ollama pull hf.co/castricolorenzo/qwen3-coder-wuic-30b-dpo:Q4_K_M - Create the local alias:
ollama cp hf.co/castricolorenzo/qwen3-coder-wuic-30b-dpo:Q4_K_M qwen3-coder-wuic:30b-dpo - Check that the model advertises tool support:
ollama show qwen3-coder-wuic:30b-dpo
Look fortoolsin the output. A model that lacks tool support may produce plausible text but cannot reliably participate in the MCP tool-calling flow.
Keep the alias exactly as shown. It is the name used in the backend and VS Code settings; changing it means every client configuration must be changed to match.
Configure Ollama for a 24 GB card
Apply the flash-attention, q8-KV-cache, 49,152-token context, and indefinite keep-alive recommendations in the Ollama configuration method appropriate to your operating system. The exact environment-variable or service-file mechanism differs by platform, so confirm the values in your Ollama installation rather than copying a shell syntax intended for another operating system.
Recommended Free Tools
Start the model once, then run:
ollama psIf the process is split between CPU and GPU, reduce the context from 49,152 to 32,768 and check again. CPU offload can make responses dramatically slower and may cause timeouts during tool calls. Context is a capacity setting, not a guarantee that every conversation can use the full number of tokens: retrieved chunks, instructions, tool results, and the model’s answer all consume the budget.
Point the WUIC backend at Ollama
Set the WUIC backend’s language-model provider to Ollama. The documented values are:
rag-llm-provider: ollama
rag-llm-api-key: ollama
rag-llm-model: qwen3-coder-wuic:30b-dpo
rag-llm-base-url: http://localhost:11434/v1The key is a non-empty dummy value required by the provider interface; Ollama itself does not use it for authentication in this local arrangement. For a separate GPU host, replace localhost with that host’s LAN name or address, retaining port 11434 and the /v1 path.
After restarting or reloading the backend, verify its health endpoint:
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Repair Windows errors before they cause bigger problems3Scan for outdated or missing drivers - takes under a minuteGET /api/Rag/HealthA successful response reports status: "ok". If health is not OK, fix the backend-to-Ollama connection before installing the editor extension; otherwise VS Code errors can obscure a lower-level networking or model-name problem.
Install and configure WUIC Assistant in VS Code
- From the directory containing the extension package, install the VSIX:
code --install-extension wuic-assistant.vsix - Open VS Code settings and select Ollama as the extension provider.
- Enter the Ollama URL. Use
http://localhost:11434/v1on the same machine or the GPU host’s LAN URL for a split deployment. - Set the model to
qwen3-coder-wuic:30b-dpo. - Set the extension context length to the value that fits your GPU. Start at 49,152 on the recommended 24 GB card; use 32,768 or less if
ollama psshows CPU/GPU splitting. - Enter the WUIC backend URL so the extension can reach retrieval and the MCP service.
- Click the WUIC Assistant icon in the activity bar to open chat.
Run a small, read-only question first, such as asking where a known class is defined. Confirm that the answer includes file paths or line information before asking the agent to edit anything. This separates retrieval failures from editing or command-execution failures.
How the wuic-rag MCP server behaves
The first MCP version exposed two tools: wuic_codebase_search and wuic_ask. A search returns the top eight chunks with file paths, line information, and snippets. The agent can cite those results in an answer or use them as the evidence for an edit. A July update reports nine metadata-aware tools in the server and a 32-tool catalog in the chatbot, so the exact menu depends on the WUIC build you installed.
The server is stateless and uses Node’s built-in HTTP implementation. It requires no npm install and runs on Node 14.18 or newer. WUIC wires it into a generated .mcp.json, which is how the editor discovers the server and its commands. Statelessness means each call carries the information needed for that operation; it does not mean the model can remember prior calls without the host including their results in the conversation.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11MCP adds knowledge, not autonomy. VS Code owns the plan, decides which tool to call, applies file changes, and runs terminal commands. Review the proposed diff and command just as you would with any other coding agent, especially when a local model has produced a syntactically valid but semantically wrong tool argument.
What performance to expect
WUIC’s measurements provide two different snapshots, so do not treat one number as a universal benchmark. A June 2026 benchmark used 89 prompt variants, each repeated three times, on a 24 GB GPU. qwen2.5-coder:32b passed about 91% of routing cases at roughly eight seconds per call. gpt-oss:20b took about 2.6 seconds per call but was weaker on complex schemas.
The July update tested the WUIC fine-tune against the base model on 102 routing cases. Base qwen3-coder:30b scored 88/102; the DPO fine-tune scored 96/102. That update reports roughly two to seven seconds per call and about 197 tokens per second on the same 24 GB GPU. Prompt length, retrieved context, quantization, thermal limits, and whether any layers are on the CPU all change the result, so use these figures as orientation rather than a promise for your machine.
| Model or approach | Reported result | Interpretation |
|---|---|---|
| qwen3-coder:30b base | 88/102 routing cases | Baseline for the July comparison |
| qwen3-coder-wuic DPO fine-tune | 96/102 routing cases | Better routing in that test; still not infallible |
| qwen2.5-coder:32b | About 91% routing pass rate; about 8 seconds/call | June benchmark, 89 variants × 3 repetitions |
| gpt-oss:20b | About 2.6 seconds/call | Faster in June, but weaker on complex tool schemas |
Local versus hosted coding agents
| Decision factor | Local Ollama and WUIC | Hosted model |
|---|---|---|
| Privacy and residency | Prompts, retrieved code, and responses can stay on your LAN; you control logs and network boundaries. | Data crosses the provider’s service boundary and follows its retention and regional controls. |
| Usage cost | No per-token charge after hardware is purchased, but electricity, storage, and maintenance remain. | Usually metered by tokens or subscription; no local GPU purchase. |
| Latency | Predictable once the model is loaded, with reload and context-size penalties. | Depends on internet path, provider load, and service limits. |
| Hardware burden | Requires a suitable GPU, disk space, drivers, Ollama, firewall rules, and upgrades. | Provider operates the serving hardware. |
| Tool-call reliability | Depends on the model’s tool training, quantization, and adapter; complex schemas can fail. | Frontier hosted models generally handle complicated schemas more consistently, though no service is perfect. |
| Context length | Bound by VRAM, KV-cache settings, and the model’s supported context. | Bound by the provider’s model and plan; often easier to obtain a large context. |
| Long-horizon, multi-file reasoning | More likely to lose the plan or make a plausible but incomplete change on difficult tasks. | Typically stronger for broad refactors and long chains of dependent edits. |
The local option is therefore a control and cost trade, not a blanket replacement for hosted intelligence. As Lorenzo Castrico puts it, “Local LLMs are not magic and they are not free.” The weights can be capable while the surrounding adapter, prompts, schemas, and editor integration remain the engineering work.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Recommended operating pattern
Use local for repeatable, bounded work
- Repository navigation and “where is this implemented?” questions.
- Explanations of a function or module using retrieved snippets.
- Small, reviewable edits with a clear file and acceptance test.
- Frequent internal questions where sending code to a cloud service is undesirable.
Escalate difficult work
Use a frontier hosted model when the task spans many files, requires a large redesign, depends on subtle library behavior, or repeatedly fails tool schemas locally. A hybrid workflow keeps routine traffic local while preserving a higher-quality option for the expensive cases.
Troubleshooting the common failures
Ollama says the model is missing
Run ollama list and verify the exact alias qwen3-coder-wuic:30b-dpo. If only the long Hugging Face name appears, repeat the ollama cp command and use the alias in both backend and extension settings.
ollama show has no tools entry
The selected model is not advertising tool support. Confirm that you pulled the WUIC DPO model and that the configured name points to that copy, not to an unrelated tag with a similar name.
Responses are very slow or time out
Check ollama ps for CPU/GPU splitting. Lower context to 32,768, then test again. Also verify that flash attention and the q8 KV cache are enabled as recommended for the 24 GB configuration. A model reload after every request usually indicates that keep-alive is not set as intended.
Free tools Windows power users keep installed
One-click scans. No signup required.
The health endpoint is not OK
Test Ollama directly from the backend host, confirm port 11434 is listening, and check that the base URL ends in /v1. For a separate host, replace localhost with its LAN address and permit the connection in the GPU host firewall. A non-empty provider key such as ollama is still required by the WUIC configuration.
VS Code cannot discover the MCP server
Confirm Node.js is version 14.18 or newer, that the generated .mcp.json points to the intended WUIC server, and that the backend URL is reachable from VS Code. Because the server needs no npm install, installing unrelated packages will not fix a wrong path or an old Node runtime.
The agent calls tools with malformed arguments
Reduce the task to one operation, ask for a read-only search, and inspect the returned schema and file references. Complex schemas are a known weakness of the smaller alternatives, and even the 30B fine-tune is not perfect. Use the hosted fallback for repeated failures rather than allowing the agent to guess through a large edit.
Is a local coding agent really free?
There is no per-token bill when WUIC sends generation to your Ollama server, and the model is available even when a hosted provider is rate-limited or offline. That is the sense in which the workflow is free. The total cost includes the 24 GB GPU or an existing machine, roughly 19 GB of model storage, electricity, cooling, driver and Ollama updates, firewall configuration, backups, and your time diagnosing serving-layer quirks.
Local serving also changes the cost curve: a heavy upfront purchase can become economical for high-volume internal use, while occasional users may spend less with a hosted plan. Privacy, predictable availability, and the ability to keep code on the LAN can matter more than the arithmetic.
Bottom-line recommendation
For the documented WUIC configuration, install qwen3-coder-wuic:30b-dpo in Ollama, run it on an RTX 4090-class 24 GB GPU, keep WUIC retrieval in the backend, and let VS Code call the backend through wuic-rag. Verify tool support, confirm /api/Rag/Health returns OK, and start with small read-only requests. Treat the local model as a fast, private assistant for routine work; keep a stronger hosted model available for complex multi-file reasoning and recovery from tool-call failures.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
