KALYPSO v1.1L (merged)
KALYPSO v1.1L is a code-generation model from GENOMA Labs, produced by merging a LoRA adapter into
Qwen2.5-Coder-14B-Instruct (Apache-2.0). It is the public member of the KALYPSO line; the internal
RL- and robotics-specialized variants are not released.
- Base:
Qwen/Qwen2.5-Coder-14B-Instruct (Apache-2.0)
- Architecture: Qwen2, 14.8B parameters, fp16, 6-shard safetensors (self-contained — no adapter needed)
- Use: general code generation / completion / reasoning; OpenAI-compatible chat
- Serving: vLLM / TGI / Ollama. A 48 GB GPU (A40 / A6000) serves fp16 comfortably with KV headroom.
Honest performance note
v1.1L benchmarks approximately at the Qwen2.5-Coder-14B base — the merge does not materially lift
HumanEval / MBPP over the base model. The value of this release is the open, reproducible pipeline
and GENOMA's RL-coding dataset methodology, not a leaderboard delta. Evaluate it as a clean, well-served
14B coder, not as a frontier model.
Attribution
Built on Qwen2.5-Coder-14B-Instruct (© Alibaba / Qwen team, Apache-2.0). This derivative is released
under Apache-2.0; the base model's license and notices are retained.
About GENOMA Labs
GENOMA Labs builds machine reasoning and perception for heavy equipment, robotic arms, and field
robotics. KALYPSO is the coding-assistant line that supports that work.
