---
language:
  - en
license: other
license_name: nimbus-layered-open-source
license_link: https://huggingface.co/Nimbus-Labs/Nimbus-4B/blob/main/LICENSES.md
library_name: transformers
pipeline_tag: image-text-to-text
base_model: Qwen/Qwen3.5-4B-Base
base_model_relation: finetune
tags:
  - nimbus
  - qwen3.5
  - coding
  - reasoning
  - local
  - transformers
---

![Nimbus Labs local coding models](assets/nimbus-labs-social.svg)

# Nimbus-4B

A balanced local coding model for implementation, debugging, and bounded agentic workflows.

Nimbus is a family of local coding models: **2B for speed, 4B for balance, and 9B v2.1 for deeper reasoning**. This Transformers repository contains the merged BF16 checkpoint and full multimodal components. The corresponding GGUF repository is text-only.

## Choose a Nimbus model

| Model | Best fit | Transformers | GGUF | MLX |
| --- | --- | --- | --- | --- |
| Nimbus-2B | Fast drafting and focused edits | [`Nimbus-Labs/Nimbus-2B`](https://huggingface.co/Nimbus-Labs/Nimbus-2B) | [`Nimbus-Labs/Nimbus-2B-GGUF`](https://huggingface.co/Nimbus-Labs/Nimbus-2B-GGUF) | [`Nimbus-Labs/Nimbus-2B-MLX-5bit`](https://huggingface.co/Nimbus-Labs/Nimbus-2B-MLX-5bit) |
| Nimbus-4B | Balanced implementation and debugging | [`Nimbus-Labs/Nimbus-4B`](https://huggingface.co/Nimbus-Labs/Nimbus-4B) | [`Nimbus-Labs/Nimbus-4B-GGUF`](https://huggingface.co/Nimbus-Labs/Nimbus-4B-GGUF) | [`Nimbus-Labs/Nimbus-4B-MLX-5bit`](https://huggingface.co/Nimbus-Labs/Nimbus-4B-MLX-5bit) |
| Nimbus-9B v2.1 | Deeper coding and reasoning | [`Nimbus-Labs/Nimbus-9B-v2.1`](https://huggingface.co/Nimbus-Labs/Nimbus-9B-v2.1) | [`Nimbus-Labs/Nimbus-9B-v2.1-GGUF`](https://huggingface.co/Nimbus-Labs/Nimbus-9B-v2.1-GGUF) | [`Nimbus-Labs/Nimbus-9B-v2.1-MLX-5bit`](https://huggingface.co/Nimbus-Labs/Nimbus-9B-v2.1-MLX-5bit) |

![Nimbus family download footprint](assets/nimbus-family-footprint.png)

The adjacent `assets/nimbus-family-footprint.json` contains the plotted values. Download size is not runtime memory: context cache and runtime buffers require additional capacity.


## Artifacts

| Format | Shards | Weight bytes | Integrity |
| --- | ---: | ---: | --- |
| BF16 safetensors | 5 | 9,078,619,688 (9.08 GB) | `SHA256SUMS` |

Text-only local files: [`Nimbus-Labs/Nimbus-4B-GGUF`](https://huggingface.co/Nimbus-Labs/Nimbus-4B-GGUF)

## Evaluation

The released **Q5_K_M** artifact was evaluated in direct mode on the full HumanEval and MBPP suites with llama.cpp `b10007`, one answer per task, temperature 0.6, top-p 0.95, top-k 20, and seed 42.

| Benchmark | Passed | Total | pass@1 |
| --- | ---: | ---: | ---: |
| HumanEval | 121 | 164 | 73.8% |
| HumanEval+ | 112 | 164 | 68.3% |
| MBPP | 285 | 378 | 75.4% |
| MBPP+ | 232 | 378 | 61.4% |

![Nimbus-4B EvalPlus results](assets/nimbus-4b-evalplus.png)

The adjacent `assets/nimbus-4b-evalplus.json` is the machine-readable source for this chart.



## Quickstart

Use a recent Transformers release compatible with Qwen3.5. Load the repository with `trust_remote_code=False`, preserve the supplied processor/tokenizer files, and enable thinking through the supplied chat template where supported. Validate generated code before execution.

```python
from transformers import AutoProcessor, AutoModelForImageTextToText

model_id = "Nimbus-Labs/Nimbus-4B"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
```

## Training and provenance

- Immediate upstream: [`Qwen/Qwen3.5-4B-Base`](https://huggingface.co/Qwen/Qwen3.5-4B-Base)
- Pinned revision: `1001bb4d826a52d1f399e183466143f4da7b741b`
- Foundation: Qwen3.5
- Relationship: Nimbus post-training / fine-tune and verified merge
- Exact uploaded file hashes: `SHA256SUMS`
- Release metadata: `release-manifest.json`

Nimbus prominently credits Qwen3.5 as the foundation for this model family. Nimbus-9B-v2.1 additionally credits DeepReinforce's Ornith-1.0-9B as its immediate upstream.

## Intended use

Local coding assistance, code explanation, debugging, test generation, and bounded tool-aware workflows. Host applications must enforce permissions, sandboxing, timeouts, and verification.

## Limitations

The model can produce incorrect, insecure, incomplete, or non-compiling code. Benchmark performance does not guarantee project-level correctness. GGUF artifacts are text-only even though this Transformers checkpoint includes multimodal components.

## Licenses and notices

See [`LICENSES.md`](LICENSES.md), [`THIRD_PARTY_NOTICES.md`](THIRD_PARTY_NOTICES.md), and the bundled Apache-2.0 text. The model is a derivative distribution; Nimbus attribution does not replace upstream attribution.

## Project

[Nimbus8](https://nimbus8.app) · [GGUF files](https://huggingface.co/Nimbus-Labs/Nimbus-4B-GGUF)