File size: 10,951 Bytes
4b905f2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fab3e8d
4b905f2
 
 
 
 
 
 
 
 
 
 
1affbd2
4b905f2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
---
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

Every file is built directly from the model's original full-precision weights. The Q4_K_M is the exact
file that won the selection league; the others are the same model at other sizes, from the same source.

`i1-` = built with an importance matrix (imatrix). `O` at the end = optimised per-tensor mix (S/M/L before it = size within that bit width). The imatrix used is in [`imatrix/`](imatrix/).

| file | quant | size | mean KL vs full precision | top-1 agreement | notes |
|---|---|---|---|---|---|
| `Vulcan-1-Coder-Q8_0.gguf` | Q8_0 | 2.68 GB | 0.0005 | 99.3% | near-lossless |
| `Vulcan-1-Coder.i1-Q6_KO.gguf` | Q6_K optimised mix, imatrix | 2.07 GB | 0.0038 | 98.1% | better than Q6_K at the same size |
| `Vulcan-1-Coder-Q6_K.gguf` | Q6_K | 2.07 GB | 0.0058 | 97.6% | |
| `Vulcan-1-Coder-Q5_K_M.gguf` | Q5_K_M | 1.81 GB | 0.0187 | 95.4% | |
| `Vulcan-1-Coder.i1-Q4_K_MO.gguf` | Q4_K optimised mix, imatrix | 1.56 GB | 0.0355 | 93.4% | closest to full precision at 4-bit size |
| `Vulcan-1-Coder-Q4_K_M-league.gguf` | Q4_K_M | 1.56 GB | 0.0615 | 91.6% | **the exact weights that won the selection league** |
| `Vulcan-1-Coder.i1-IQ4_XS.gguf` | IQ4_XS, imatrix | 1.42 GB | 0.0720 | 90.7% | |
| `Vulcan-1-Coder.i1-Q3_K_LO.gguf` | Q3_K optimised mix, imatrix | 1.40 GB | 0.0841 | 89.9% | best 3-bit |
| `Vulcan-1-Coder.i1-Q3_K_MO.gguf` | Q3_K optimised mix, imatrix | 1.35 GB | 0.1166 | 88.1% | |
| `Vulcan-1-Coder.i1-Q3_K_SO.gguf` | Q3_K optimised mix, imatrix | 1.24 GB | 0.1867 | 85.5% | below our 86% top-1 line |
| `Vulcan-1-Coder.i1-IQ3_M.gguf` | IQ3_M, imatrix | 1.23 GB | 0.1733 | 85.5% | below our 86% top-1 line |
| `Vulcan-1-Coder.i1-IQ3_S.gguf` | IQ3_S, imatrix | 1.19 GB | 0.2044 | 84.1% | smallest; below our 86% top-1 line |

Scores: 40 x 512 held-out tokens, against the original full-precision model (lower KL = closer).

Wrote a script to find mathematically optimal quants for all models. Next step: find the best quants - it's not the same thing. 

![image](https://cdn-uploads.huggingface.co/production/uploads/6aa007cf4ae94de9b1aaa30c/_JM18lVaATlekn4io98xl.png)

**About the quantization bug.** Vulcan is very close to CherryPit 1.2: its change is about one BF16
step per weight. Earlier non-league quants were made the usual way, from a BF16 export, and that BF16
rounding is as large as the change itself. So those files were a slightly different model from the one
that won the league (the export-built Q4_K_M shared none of its tensors' bytes with the winner). All
files here are now built straight from the full-precision weights, so every size is the model that was
selected. The league Q4_K_M was never affected.

## 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. These are mathematics benchmark. This is a coding model. Coding benchmarks coming soon.

### 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.  
*Measured on the earlier BF16-export Q8_0, not the file above; rerun pending.*
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.  
*Measured on the earlier BF16-export Q8_0, not the file above; rerun pending.*
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 | βœ“ | 2,020 |
| A02 | 62 | 62 | βœ“ | 4,937 |
| A03 | 79 | 79 | βœ“ | 3,468 |
| A04 | 68 | 70 | βœ— | 3,198 |
| A05 | 65 | 65 | βœ“ | 2,848 |
| A06 | 441 | 441 | βœ“ | 2,155 |
| A07 | 396 | 396 | βœ“ | 4,241 |
| A08 | 244 | 244 | βœ“ | 3,317 |
| A09 | 29 | 29 | βœ“ | 17,086 |
| A10 | 156 | 156 | βœ“ | 9,186 |
| A11 | 896 | 896 | βœ“ | 11,527 |
| A12 | 161 | 161 | βœ“ | 18,540 |
| A13 | 39 | 39 | βœ“ | 23,518 |
| A14 | 681 | 681 | βœ“ | 7,540 |
| A15 | 2 | 83 | βœ— | 7,237 |
| A16 | 178 | 178 | βœ“ | 1,914 |
| A17 | - | 243 | βœ— (hit token limit) | 24,000 |
| A18 | 505 | 503 | βœ— | 8,668 |
| A19 | 279 | 279 | βœ“ | 5,581 |
| A20 | 190 | 190 | βœ“ | 7,950 |
| A21 | 50 | 50 | βœ“ | 4,905 |
| A22 | 754 | 754 | βœ“ | 5,987 |
| A23 | 245 | 245 | βœ“ | 10,713 |
| A24 | 669 | 669 | βœ“ | 5,061 |
| A25 | 850 | 850 | βœ“ | 6,960 |
| A26 | 132 | 132 | βœ“ | 8,593 |
| A27 | 223 | 223 | βœ“ | 10,525 |
| A28 | 12 | 107 | βœ— | 17,496 |
| A29 | 3 | 157 | βœ— | 18,535 |
| A30 | 243 | 393 | βœ— | 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