Can local models write code? What the makers report, what a local coding setup needs and where to be careful.
Short answer. Yes, and well enough to be useful, with review. OpenAI reports 62.4 per cent on SWE-bench Verified for gpt-oss-120b and 60.7 per cent for gpt-oss-20b at high reasoning effort, and the Qwen3.8-27B card reports 61.7 on SWE-bench Pro and 90.3 on LiveCodeBench v6. These are the makers’ figures, run with their own setups, so treat them as upper bounds for your codebase. A local model also needs an agent harness, enough context for the files it reads, and a person who reviews and tests every change.
01
What the makers report
Benchmarks differ and are not comparable with each other: SWE-bench Verified and SWE-bench Pro are different test sets. Each maker ran its own harness, so these show direction, not what you will get on your repository.
Model and benchmark
Result
Conditions
gpt-oss-120b, SWE-bench Verified
62.4 per cent at high effort (52.6 medium, 47.9 low)
OpenAI model card; OpenAI runs SWE-bench on a fixed subset of 477 verified tasks
gpt-oss-20b, SWE-bench Verified
60.7 per cent at high effort (53.2 medium, 37.4 low)
OpenAI model card
gpt-oss-120b and 20b, Codeforces Elo with tools
2622 and 2516 at high effort
OpenAI model card
Qwen3.8-27B, SWE-bench Pro
61.7
Qwen card; Claude Code harness, temperature 1.0, 256K context
Qwen3.8-27B, LiveCodeBench v6
90.3
Qwen card
Qwen3.8-27B, Terminal Bench 2.1 (Terminus)
73.0
Qwen card
02
What a local coding setup needs
A model with tool calling, served through an OpenAI-compatible endpoint that a coding agent can use. Ollama, llama.cpp and vLLM all offer one.
A harness: the program that reads files, runs commands and applies edits. The model alone does not touch your code.
Context for the files it reads. Qwen ran its coding tests with a 256K-token context; at 262,144 tokens the cache for Qwen3.8-27B is about 17.2 GB by our formula, on top of the weights.
Memory for the weights: gpt-oss-120b is a 63.4 GB file, gpt-oss-20b 12.1 GB, and Qwen3.8-27B about 16.5 GB at 4-bit.
03
Limits and good practice
Review every change. Models can invent functions or packages and write code that runs but is wrong or insecure.
Run tests the model did not write, and run the agent in a sandbox or container with only the folders it needs.
Keep secrets out of prompts and repositories. Local use keeps code on your machine, which is the point, but logs and sessions still hold it.
Expect weaker results on unusual frameworks, large legacy systems and long multi-step tasks than the headline numbers suggest.
04
Where ANVE fits
ANVE’s Design Studio generates screens from a description and exports HTML or a React component. Coding agents that speak the OpenAI API format can use the models through the runtime’s own OpenAI-compatible endpoint today; the governed ANVE API, with a key per user and every request in the audit trail, is not on ANVE yet. ANVE is pre-release and has not published speed figures.
05
Questions people ask
Which local model should I use for coding?
No single answer holds. Check current benchmark pages such as SWE-bench and LiveCodeBench, then try two or three on your own repository with your own tests. Models change every few months.
Do I need a big GPU for coding assistance?
Memory decides it. A 4-bit 27B model is about 16.5 GB, gpt-oss-20b about 12 GB; long contexts add cache. Smaller models fit smaller machines but are weaker on harder tasks.
Is code written by a local model private?
The model runs on your machine, so prompts do not go to a cloud provider. The code still lives in logs and sessions on that machine, and you remain responsible for reviewing it.
06
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
