Vulcan 2.5B Coder
A small coding and reasoning model, forged on CherryPit 1.2 by Integer Space Artificial Selection. We bred it directly in GGUF form and picked it from about 30 candidates in one afternoon, on a laptop with 6 GB of VRAM. No training runs, no cluster.
AIME
1 sample · temp 1.0 · Q8_0
Our benchmark harness, no tool use, thinking on, up to 24,000 tokens, temperature 1.0, top-p 0.95, one sample per problem. The AIME 2025 set is MathArena's aime_2025, all 30 problems. Dashed cells are misses; on AIME 2025, three problems hit the token limit before giving an answer.
AIME 2025 compared
- Qwen3 Max75.0%
- Claude Sonnet 474.3%
- Claude Opus 473.3%
- Vulcan 2.5B Coder70.0%
- gpt-oss-120b66.0%
- Gemini 2.5 Flash-Lite (reasoning)58.7%
- Mistral Medium 3.138.3%
- Mistral Large 338.0%
- GPT-4.134.7%
Other models' scores are published figures, not runs on our harness, so their settings differ. Vulcan's score is one sample, so expect a few points of run-to-run noise.
The selection league
Every candidate answered hard coding questions, plus some maths, agentic and reasoning tasks. Each answer scored 0, 0.5 or 1. Vulcan finished first.
The league questions varied between models, so this is not a strict head-to-head. A like-for-like benchmark is coming.
League table
Every candidate from the selection league. Scores are per question, with 0, 0.5 or 1 for each answer. "Last four" is the score on the final four question sets, which include mathematics. Questions varied between models, so treat close scores as ties.
| # | model | tier | answered | correct | score | last four | mix | licence |
|---|
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 league |
| safetensors | BF16 | n/a | coming soon |
Usage
Leave thinking on and give it room. Short token limits cut its answers off. Same architecture and size as CherryPit 1.2, so it runs anywhere GGUF does.
llama-server -m Vulcan-1-Coder-Q4_K_M-league.gguf -ngl 99 -c 32768 --jinja
What went into the forge
- DRUPEs/CherryPit-1.2-2Bbase · Apache-2.0
- openbmb/MiniCPM5-2BApache-2.0
- NANI-Nithin/MiniCPM5-2B-OpenCodeReasoning2-LoRAApache-2.0 · data CC-BY-4.0
- zzzluvst/MiniCPM5-2B-SWE-smith-GRPOApache-2.0
- Caldalis/MiniCPM5-2B-MathApache-2.0
Thanks to the team at DRUPEs for CherryPit 1.2, to NANI-Nithin for the OpenCodeReasoning-2 adapter, and to NVIDIA for the OpenCodeReasoning-2 dataset.