KALYPSO v1.1L
KALYPSO v1.1L is the public coding model of the KALYPSO family by GENOMA Labs. It is Qwen2.5-Coder-14B-Instruct fine-tuned with LoRA (r=64, alpha=128, merged to full fp16 weights) on Kraken-Public, a decontaminated dataset of 18,049 records derived from NVIDIA Nemotron coding data.
It is tuned for agentic coding: multi-step programming tasks, tool-calling scaffolds, and plan-then-implement workflows in Python, C++, and SQL.
Benchmarks (EvalPlus)
| Benchmark | KALYPSO v1.1L | Qwen2.5-Coder-14B-Instruct (ref) |
|---|---|---|
| HumanEval | 87.8 | ~89 |
| HumanEval+ | 82.9 | ~87.2 |
| MBPP | 84.7 | ~84 |
| MBPP+ | 70.4 | ~72.8 |
Honest framing: pure code completion is saturated at the 14B scale, and this model scores at parity with its base. v1.1L is published as the reference fine-tune for the Kraken-Public dataset; its value is agentic breadth, not single-shot completion wins.
Usage
vLLM (recommended for serving):
vllm serve genomalabs/KALYPSO-v1.1L --dtype float16 --max-model-len 32768
Transformers:
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "genomalabs/KALYPSO-v1.1L"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="float16", device_map="auto")
messages = [{"role": "user", "content": "Write a Python function that merges overlapping intervals."}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=512)
print(tok.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))
Uses the standard Qwen2.5 chat template. Context length: 32k.
Family
- KALYPSO v1.1L (this model): public 14B dense coder.
- genomalabs/kalypso-v1.2l: successor, Qwen3-Coder-30B-A3B based, coding plus planning, with GGUF quants.
- Original v1.1L release: GenomaLabs-com/KALYPSO-v1.1L.
Attribution and license
Apache-2.0. Built with Qwen (Qwen2.5-Coder-14B-Instruct, Apache-2.0). Training data derived from NVIDIA Nemotron coding datasets; portions of the data lineage may also be subject to the Qwen and DeepSeek license terms of their originating models.
GENOMA Labs · KALYPSO v1.1L · Built with Qwen