DEV Community

AI OpenFree
AI OpenFree

Posted on

Darwin-9B-Opus-Code Is Now on LM Studio Hub: Run a VIDRAFT-Tuned Coding Model Locally on Your Desktop

#darwin9bopuscode #lmstudio #localllm #codingmodel

Darwin-9B-Opus-Code Is Now on LM Studio Hub: Run a VIDRAFT-Tuned Coding Model Locally on Your Desktop

TL;DR: Darwin-9B-Opus-Code, a coding-focused model fine-tuned by VIDRAFT's FINAL-Bench team on top of the Qwen3.5 base, is now publicly available on LM Studio Hub — meaning any developer on Windows, macOS, or Linux can pull and run it locally with no API key, no cloud dependency, and no usage fees. The model ships with official recommended sampling parameters for reproducible, high-quality code generation right out of the box.

What it is

Darwin-9B-Opus-Code is a 9-billion-parameter, instruction-tuned language model optimized for software engineering tasks. It is built on the Qwen3.5 architecture and further fine-tuned by VIDRAFT's FINAL-Bench team, which focuses on rigorous evaluation and optimization of coding-capable models.

Key facts from the source:

  • Model identifier: qwen3.5-darwin-9b-opus-code
  • Publisher on LM Studio Hub: dongxiat (FINAL-Bench / VIDRAFT)
  • Visibility: Public
  • Upstream Hugging Face repository: FINAL-Bench/Darwin-9B-Opus
  • LM Studio Hub listing: lmstudio.ai/dongxiat/qwen3.5-darwin-9b-opus-code
  • The listing includes an official preset of sampling parameters recommended by the FINAL-Bench team, so you don't have to guess at configuration.
  • The model is configured with a software-engineering-specific system prompt out of the box: "You are an expert software engineer. Write clean, efficient, well-commented code. Explain your approach before writing."

The LM Studio Hub deployment marks a significant distribution milestone: it brings the model directly into the desktop-native LLM workflow used by a large global base of developers who prefer on-device inference over cloud-hosted endpoints.

How it works

At a conceptual level, Darwin-9B-Opus-Code follows the standard fine-tuning paradigm on top of a strong pre-trained base model:

  1. Base architecture — Qwen3.5: Qwen3.5 is a publicly available transformer-based LLM series from Alibaba Cloud, known for strong multilingual and code understanding capabilities. The 9B parameter scale sits in a practical sweet spot: capable enough for complex reasoning tasks, yet small enough to run comfortably on consumer-grade hardware with a capable GPU or Apple Silicon.

  2. FINAL-Bench fine-tuning: VIDRAFT's FINAL-Bench team applies domain-specific fine-tuning to steer the model toward software engineering tasks — clean code generation, inline documentation, and structured reasoning about implementation approach before writing code.

  3. Sampling preset for code: The FINAL-Bench team publishes an official set of inference parameters alongside the model. These are surfaced directly as a LM Studio preset, so the sampling behavior is tuned for deterministic-leaning, high-quality code output rather than creative or open-ended generation.

The recommended sampling configuration visible in the Hub listing includes:

  • Temperature: 0.6 (moderately low — favors coherent, predictable output)
  • Top P: 0.95
  • Top K: 20
  • Min P: 0 (disabled)
  • Repeat Penalty: 1 (neutral)
  • Presence Penalty: 0 (neutral)
  • Context overflow handling: Truncate Middle

These values represent the FINAL-Bench team's recommended defaults for coding workloads and are safe to use as a starting point.

Benchmarks & results

The source article (LM Studio Hub listing) does not include quantitative benchmark scores such as HumanEval, MBPP, or LiveCodeBench pass rates. The model is positioned as a coding-optimized fine-tune with sampling parameters curated by a team focused on evaluation ("FINAL-Bench"), but specific numbers are not published in this listing.

Developers who want to evaluate performance rigorously should run their own benchmarks against their target tasks, or watch the FINAL-Bench Hugging Face organization for published evaluation results.

How to try it

Option 1 — LM Studio (recommended for desktop use):

  1. Download LM Studio from lmstudio.ai.
  2. In the app, search for dongxiat/qwen3.5-darwin-9b-opus-code in the model browser, or navigate directly to the Hub listing.
  3. Download the model and load the official preset — sampling parameters are pre-configured.

Option 2 — Hugging Face CLI:

The upstream model is publicly available on Hugging Face:

huggingface-cli download FINAL-Bench/Darwin-9B-Opus
Enter fullscreen mode Exit fullscreen mode

Option 3 — OpenAI-compatible local API via LM Studio:

LM Studio exposes a local OpenAI-compatible REST endpoint once a model is loaded. You can point any OpenAI SDK client at http://localhost:1234/v1 (LM Studio's default) to use the model programmatically.

FAQ

Q: Do I need a GPU to run Darwin-9B-Opus-Code locally?
A: LM Studio supports CPU inference, but a GPU (NVIDIA with CUDA, AMD with ROCm, or Apple Silicon with Metal) is strongly recommended for practical inference speed at 9B parameters. Apple Silicon Macs with unified memory handle this model size well.

Q: Is this model free to use?
A: The model is listed publicly on both LM Studio Hub and Hugging Face with no paywall on the Hub itself. Always review the upstream model license in the FINAL-Bench/Darwin-9B-Opus Hugging Face repository for terms covering commercial and research use, as they derive from the Qwen3.5 base model license.

Q: Can I use the sampling preset outside of LM Studio?
A: Yes. The parameters published in the Hub listing (temperature, top-p, top-k, etc.) are standard inference settings you can apply in any inference framework — llama.cpp, vLLM, Transformers, or others — they are not LM Studio-specific.

Originally reported by LM Studio (미국) (2026-07-01) — source article.

Top comments (0)