Blame for README.md

---
license: apache-2.0
language:
  - en
library_name: gguf
pipeline_tag: text-generation
base_model:
  - DRUPEs/CherryPit-1.2-2B
base_model_relation: merge
tags:
  - code
  - coding
  - reasoning
  - math
  - aime
  - thinking
  - gguf
  - llama.cpp
  - lm-studio
  - ollama
  - merge
  - model-merge
  - small-model
  - 2b
  - local-llm
  - on-device
  - quantized
  - q4_k_m
  - cherrypit
  - minicpm
  - integer-space
  - vulcan
---

# Vulcan-1-Coder (GGUF)

![image](https://cdn-uploads.huggingface.co/production/uploads/6aa007cf4ae94de9b1aaa30c/uzEAciENqItRjhT0bwydS.png)
Results used no tools and a benchmarking harness. Also quantised to q8. 

A 2.5B coding and reasoning model built by using Integer Space Artificial Selection on **[CherryPit 1.2](https://huggingface.co/DRUPEs/CherryPit-1.2-2B)**, 
an already excellent model made by the team at DRUPEs. This model was generated and tested as a `.gguf`. Details
of the method will be published later and released for all. I'm working hard to put everything together. 

**Made in an afternoon.** Vulcan was produced in a single afternoon along with about 30 other models (12 in its
class), all generated and tested in integer space, on a laptop with a **6 GB VRAM** GPU, using only my own tools.
No training runs, no cluster.

Every run, lasting just a few hours, resulted in success. The techique can work for all types of models. It's a data-led technique where the training is
squeezed out rather than training it on more data. If the information exists in the model, we select for it. 

Named after Vulcan, the Roman god of fire and the forge, the smith who crafts tools for the gods.

**Safetensors version: coming soon.**

## Files

| file | quant | size | notes |
|---|---|---|---|
| `Vulcan-1-Coder-Q4_K_M-league.gguf` | Q4_K_M | 1.56 GB | the exact weights that won the selection league |

## Results

Notes: This model has not been fully tested or trained further after selection. It may have flaws or issues, for example one time it sent a bunch of symbols.
But generally it scores high. All benchmarks are with no tools AND quantised. 

### AIME 2025

Run 1:
21 / 30 = 70.0% with thinking on, up to 24K tokens, temperature 1.0, 1 sample, Q8_0. Benchmarking harness. No tools.
Per-problem results: Will add soon to a site I'm building.

![image](https://cdn-uploads.huggingface.co/production/uploads/6aa007cf4ae94de9b1aaa30c/IUzjwqtu0LM4C3ExVLocX.png)

### AIME 2026

Notes: Although at this level, it could be noise, I wonder if selecting for the model at Q4_K_M found a model that's best at Q4_K_M. I'll get access to a better GPU soon and test.

Run 1:
**23 / 30 = 76.7%** with thinking on, up to 24K tokens, temperature 1.0, 1 sample, Q4_K_M. Benchmarking harness. No tools.
Per-problem results:  Will add soon to a site I'm building.

![image](https://cdn-uploads.huggingface.co/production/uploads/6aa007cf4ae94de9b1aaa30c/jgLV83-gVlDHg_P9rO07w.png)

Run 2:
21 / 30 = 70.0% with thinking on, up to 24K tokens, temperature 1.0, 1 sample, Q8_0. Benchmarking harness. No tools.
Per-problem results: Will add soon to a site I'm building.

### Selection league

Scored on hard coding questions, plus some maths, agentic and reasoning tasks. No repeat questions and a large pool to avoid over-selecting for benchmarks. 
NO BENCHMARK QUESTIONS, all generated by Claude including some questions that are extremely hard. The league questions varied between models and the questions
were not enough to find true representative results. 

(0 / 0.5 / 1 per answer, Q4_K_M, temperature 1.0). Vulcan finished first. Plain CherryPit 1.2 is shown as the baseline. Note: While this is at the baseline,
there are many generated models that were below it. CherryPit was chosen for a reason, it's well above its class.

![league results, part 1](images/league-1.png)
![league results, part 2](images/league-2.png)

| model | score per question |
|---|---|
| **Vulcan-1-Coder** | **0.763** |
| CherryPit 1.2 (base) | 0.611 |

> **Note:** the league questions varied between models, so these scores are not fully representative of a
> head-to-head comparison. A like-for-like benchmark is coming.

## Usage

Recommended sampling: **temperature 1.0, top-p 0.95**, thinking on. Give it room to think (16K+ tokens for
hard problems): short limits cut off its answers.

Works in llama.cpp, LM Studio, Ollama and anything else that runs GGUF. Same architecture and size as CherryPit 1.2.

```
llama-server -m Vulcan-1-Coder-Q4_K_M-league.gguf -ngl 99 -c 32768 --jinja
```

For questions, contact ryan@filioai.com

## Thanks

- **The team at DRUPEs** for [CherryPit 1.2](https://huggingface.co/DRUPEs/CherryPit-1.2-2B), the base of Vulcan.
- **NANI-Nithin** for the [OpenCodeReasoning-2 LoRA adapter](https://huggingface.co/NANI-Nithin/MiniCPM5-2B-OpenCodeReasoning2-LoRA).
- **NVIDIA** for the [OpenCodeReasoning-2](https://huggingface.co/datasets/nvidia/OpenCodeReasoning-2) dataset that adapter was trained on ([CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/)).

## Things to come - progress

#### Paper and files released - 50%

Hand writing the paper and hand-making the diagrams without AI so I'll release it as soon as I can as a pre-print. 
I will remove secrets from the folder and clean it up. Folder currently has thousands of files and I'd say about 25% are relevant. 

#### More benchmarks - ?

Whenever I feel up to it, adding some now. 

Edit: Results from AIME 2026: Vulcan-1-Coder (Q4_K_M) benchmark harness no tools.

Settings: thinking on, max 24,000 tokens, temperature 1.0, top-p 0.95, 1 sample per problem, llama.cpp (CUDA).

| problem | Vulcan's answer | correct answer | result | tokens used |
|---|---|---|---|---|
| A01 | 277 | 277 | ✓ | 2,020 |
| A02 | 62 | 62 | ✓ | 4,937 |
| A03 | 79 | 79 | ✓ | 3,468 |
| A04 | 68 | 70 | ✗ | 3,198 |
| A05 | 65 | 65 | ✓ | 2,848 |
| A06 | 441 | 441 | ✓ | 2,155 |
| A07 | 396 | 396 | ✓ | 4,241 |
| A08 | 244 | 244 | ✓ | 3,317 |
| A09 | 29 | 29 | ✓ | 17,086 |
| A10 | 156 | 156 | ✓ | 9,186 |
| A11 | 896 | 896 | ✓ | 11,527 |
| A12 | 161 | 161 | ✓ | 18,540 |
| A13 | 39 | 39 | ✓ | 23,518 |
| A14 | 681 | 681 | ✓ | 7,540 |
| A15 | 2 | 83 | ✗ | 7,237 |
| A16 | 178 | 178 | ✓ | 1,914 |
| A17 | - | 243 | ✗ (hit token limit) | 24,000 |
| A18 | 505 | 503 | ✗ | 8,668 |
| A19 | 279 | 279 | ✓ | 5,581 |
| A20 | 190 | 190 | ✓ | 7,950 |
| A21 | 50 | 50 | ✓ | 4,905 |
| A22 | 754 | 754 | ✓ | 5,987 |
| A23 | 245 | 245 | ✓ | 10,713 |
| A24 | 669 | 669 | ✓ | 5,061 |
| A25 | 850 | 850 | ✓ | 6,960 |
| A26 | 132 | 132 | ✓ | 8,593 |
| A27 | 223 | 223 | ✓ | 10,525 |
| A28 | 12 | 107 | ✗ | 17,496 |
| A29 | 3 | 157 | ✗ | 18,535 |
| A30 | 243 | 393 | ✗ | 15,340 |

**Score: 23/30 = 76.7%**

#### Quant perplexity-solver

Find the quants with the best perplexity

This unfortunately requires a runtime to go  with it.

#### Further processing -  30%

I've discovered another technique that works to post-train one of these models. I'm still perfecting it but I'm already getting good results. I think the models can be optimised further.

#### Qwen 27b - 20%

Started the analysis process.

#### STEM - 50%

Static Embedding. **Experimental native runtime** for the `.stem` model format and embedding engine. The current runtime remains intentionally narrow: 
**static embedding tables, token IDs, FP32 or INT8, mean pooling and optional L2**. It is not yet a general-purpose encoder/ONNX/GGUF runner.
It creates a .STEM format that is faster, smaller, and more accurate.

## Licence

Released under the [Apache License 2.0](LICENSE). Vulcan-1-Coder is a modified work that includes weights from:

- [DRUPEs/CherryPit-1.2-2B](https://huggingface.co/DRUPEs/CherryPit-1.2-2B): Apache-2.0
- [openbmb/MiniCPM5-2B](https://huggingface.co/openbmb/MiniCPM5-2B): Apache-2.0
- [NANI-Nithin/MiniCPM5-2B-OpenCodeReasoning2-LoRA](https://huggingface.co/NANI-Nithin/MiniCPM5-2B-OpenCodeReasoning2-LoRA): Apache-2.0
  (trained on NVIDIA OpenCodeReasoning-2, CC-BY-4.0)
- [zzzluvst/MiniCPM5-2B-SWE-smith-GRPO](https://huggingface.co/zzzluvst/MiniCPM5-2B-SWE-smith-GRPO): Apache-2.0
- [Caldalis/MiniCPM5-2B-Math](https://huggingface.co/Caldalis/MiniCPM5-2B-Math): Apache-2.0