Running a coding focused AI locally is absolutely practical now. With a single workstation GPU you can get a fast, private assistant that writes functions, fixes bugs, and drafts tests, without shipping your code to the cloud. Below is a clear, opinionated guide that distills the earlier discussion into one cohesive plan.
Why go local?
-
Privacy and compliance: your source never leaves your machine.
-
Latency and availability: no rate limits, no waiting for a busy service.
-
Cost: once you have paid for hardware, heavy daily usage is cheaper than subscriptions.
The trade off: the very hardest “agentic” coding tasks (whole repo refactors, tool driven multi step plans) are still where frontier hosted models tend to be more reliable.
The runtime: how to host models locally
You need a “model runtime” that downloads weights and exposes an API your tools can call.
-
Ollama - The simplest path. One command pulls a model and starts a local endpoint. Good CLI and REST ergonomics and wide community adoption.
-
LM Studio - Desktop app with a friendly UI and a built in OpenAI compatible server; very easy to point IDE tools at.
-
vLLM - A high performance server (great throughput and latency), ideal if you want to serve multiple IDEs or multiple users over your LAN.
-
Open WebUI (optional) - A polished chat UI that can sit atop Ollama or LM Studio for “ask the repo” style workflows or teammate access.
Recommendation: start with Ollama or LM Studio. Move to vLLM if you need more performance or multi user reliability.
IDE integration and whole repo search
For a coding assistant to be useful, it must understand your repository, not just the file in focus.
-
Continue (VS Code / JetBrains): builds an embedding index over your project and does similarity search to bring the right files and sections into context automatically. It can browse and search the codebase via built in tools.
-
Aider (CLI, editor agnostic): creates a “repo map” (via parsers like tree sitter) summarizing classes, functions, and call sites, then uses that map to propose precise edits across files. It is excellent for “change A → find or modify B or C”.
-
Sourcegraph Cody: first class code search plus AI. It searches the workspace and feeds matched snippets to the model; can also integrate with a self hosted Sourcegraph instance if you want enterprise grade search.
Bottom line, yes, modern plugins can search the entire codebase and relate one file to another. For best results, allow an initial indexing step and refresh the index after large refactors.
Which free models are strongest for coding?
For local code generation and repair, these open models stand out:
-
Qwen2.5-Coder-32B-Instruct
-
Excellent code quality across common benchmarks and practical editing tasks.
-
Long context (useful for multi file prompts).
-
Permissive licensing (Apache 2.0 for most sizes).
-
DeepSeek-Coder-V2 family
-
V2 (Mixture of Experts, “236B total / about 21B active”) is a top performer but heavy to host;
-
V2-Lite-16B is the sweet spot for local use while retaining strong coding ability.
-
Take note of MoE: “total parameters” are large, but active parameters per token are much smaller.
-
Codestral 22B (Mistral)
-
Very capable, long context, strong repository level behavior.
-
License caveat: research or test friendly, but not for production for commercial use without further permissions.
Solid baselines if you are compute constrained: StarCoder2 7B or 15B; generalist but capable: Llama 3.x Instruct (8B or 70B).
Does bigger always mean better? Mostly yes within a family (7B to 14B to 32B tends to improve), but data quality, instruction tuning, context length, and decoding settings also matter. MoE models complicate the “parameter equals power” mental model; look at active parameters and real results.
Hardware: do you need a 4080? What if you have a 4090?
VRAM dictates your ceiling because you must hold weights and the KV cache (which grows with context length and batch size).
-
24GB VRAM (RTX 4090 / 3090): ideal for Qwen2.5-Coder-32B at 4 bit quantization. You will have enough headroom for longer contexts and stable editing sessions.
-
16GB VRAM (RTX 4080 or 4080S or 4070Ti 16GB): best with Qwen2.5-Coder-14B or DeepSeek-Coder-V2-Lite-16B (4 bit). You can attempt 32B with aggressive offloading and shorter context, but it is a compromise.
-
Multi GPU or 32GB plus VRAM: enables bigger unquantized loads or faster 22B class models, and gives breathing room for very long contexts.
Practical tip: for live coding speed, use an efficient 7 to 16B model for autocomplete and quick Q and A, and switch to a 32B model when you need higher stakes edits or complex refactors.
3090 vs 4090/4080: the essentials
click to open
-
Architecture and process: 3090 = Ampere on Samsung 8N; 4090 = Ada Lovelace on TSMC 4N, bringing a clear efficiency gain.
-
Cache and memory: 3090 has ~6 MB L2 and ~936 GB/s; 4090 has ~72 MB L2 and ~1008 GB/s, which helps memory-bound work.
-
Media engines: 4090 adds AV1 hardware encode; 3090 supports AV1 decode only.
-
Platform: 3090 retains NVLink; 4090 removes it.
Why 3090 can hit inference bottlenecks
Smaller L2 (~6 MB vs ~72 MB) and lower effective bandwidth (~936 GB/s vs ~1008 GB/s) make the 3090 more memory-bound, so long contexts and small batches stall more often.
Why 3090 and 4090 host bigger models than 4080
Both 3090 and 4090 have 24GB VRAM, leaving room for large weights plus KV cache at low-bit quantization. The 4080’s typical 16GB fills sooner, forcing heavier quantization, shorter context, or CPU offload. Also, 3090/4090 use a 384-bit bus; 4080 uses 256-bit, reducing memory throughput. In short, capacity first and bandwidth second make 3090 and 4090 better for big models, while 4080 is more constrained.
Price (2025)
RTX 4090: launch MSRP $1,599; current new listings often ~$2,300–$3,000; used ~$2,100 depending on seller and model.
RTX 3090: launch MSRP $1,499; typical used ~$650–$800; occasional new old stock around ~$900.
RTX 4080 / 4080 SUPER: launch MSRPs $1,199 and $999; street prices vary—used often ~$760–$800, while new cards can sit ~$1,300+ depending on brand and availability.
Note: prices swing with availability, region, and retailer promos; the figures above reflect recent US market snapshots.
How does a local 32B model on a 4090 compare to GPT and Claude?
Where local shines
-
Everyday coding: writing functions, fixing bugs, adding tests, small scope refactors, top open models are highly competitive here.
-
Privacy and cost: your code stays on premises; heavy daily use is cost effective after hardware.
-
Latency control: no shared limits; you can tune decoding for snappy responses.
Where hosted leaders still win
-
Agentic reliability: planning multi step changes, tool orchestration (run or parse tests, apply diffs with fewer mistakes), and obeying exact constraints are still more consistent on frontier closed models (for example, GPT 4.1 and Claude 3.5 tier).
-
Ultra long context: cloud models offer extremely large windows (hundreds of thousands to million token scale in current offerings), which helps for whole repo reads or very large design docs.
-
Edge cases and robustness: the newest proprietary models often sit atop leaderboards like LiveCodeBench and BigCodeBench, especially on complex, real world tasks.
Pragmatic strategy: Use your local 32B as the default assistant. When you hit whole repo surgery or strict tool driven workflows, temporarily call a hosted model as a safety net.
Quick start recipes
A. Easiest path (Ollama and Continue)
B. GUI API server (LM Studio) and Aider
-
Install LM Studio, start its OpenAI compatible server, and download Qwen2.5-Coder-32B (or 14B).
-
Set OPENAI_API_BASE to the LM Studio URL in your shell.
-
Run Aider in your repo root; add only the files you want it to edit, Aider uses its repo map to find related files automatically.
C. High performance and multi user (vLLM)
Workflow tips for better results
-
Index early, refresh often: Embedding indexes (Continue) or repo maps (Aider) make the assistant codebase aware. Refresh after major refactors.
-
Tune context and temperature: For edits, lower temperature and shorter max tokens equals fewer hallucinations and tighter diffs. For brainstorming, loosen them.
-
Quantization matters: 4 bit gets big models onto single GPUs; 8 bit can improve quality slightly if you have VRAM headroom.
-
License check: Qwen2.5-Coder sizes are generally Apache 2.0; Codestral’s default license is not for production use, fine for evaluation, not for commercial deployment without permissions.
A short decision checklist
-
Need maximum quality privately, have at least 24GB VRAM? → Qwen2.5-Coder-32B-Instruct (4 bit) via Ollama or LM Studio plus Continue.
-
Have 16GB VRAM? → Qwen2.5-Coder-14B or DeepSeek-Coder-V2-Lite-16B (4 bit) as daily driver.
-
Heavy multi user or service use? → Host with vLLM.
-
Big, risky refactor with strict tool control or gigantic context? → briefly use GPT 4.1 or Claude 3.5.
Final takeaway
With a 4090, a well tuned Qwen2.5-Coder-32B or DeepSeek-Coder-V2-Lite for speed plus Continue or Aider gives you a strong, private coding copilot that covers most day to day work. Keep a hosted model in your back pocket for the occasional whole repo overhaul or agentic marathon, and you will have the best blend of capability, privacy, performance, and cost.
