LFM2.5-24B-A2B · Harbor · Claude Code (LoRA, GGUF)

The GGUF version of harness-hub/LFM2.5-24B-A2B-harbor-claude-code-lora, a LoRA adapter for LiquidAI/LFM2.5-24B-A2B trained with reinforcement learning inside Claude Code, for llama.cpp.

In short: no measured change from the base model after 100 training steps. The original adapter card describes the training and the evaluation.

LFM2.5-24B-A2B-harbor-claude-code-lora-F16.gguf is the adapter only (about 37 MB). It runs on top of LiquidAI's base GGUF.

Run

1. Download the base model and the adapter

hf download LiquidAI/LFM2.5-24B-A2B-GGUF LFM2.5-24B-A2B-Q8_0.gguf --local-dir .
hf download harness-hub/LFM2.5-24B-A2B-harbor-claude-code-lora-GGUF LFM2.5-24B-A2B-harbor-claude-code-lora-F16.gguf --local-dir .

The adapter was trained on the BF16 weights. The closer the base is to them, the closer you get to the trained behavior. Q8_0 is a good default; Q4_K_M is smaller, but drifts further from what the adapter learned against.

2. Serve

llama-server -m LFM2.5-24B-A2B-Q8_0.gguf --lora LFM2.5-24B-A2B-harbor-claude-code-lora-F16.gguf \
    --jinja -c 65536 --port 8080 --alias lfm-harbor

--jinja is required: it turns on tool calling, which Claude Code depends on.

3. Point Claude Code at it

llama-server also serves the Anthropic Messages API that Claude Code speaks (announcement):

ANTHROPIC_BASE_URL=http://127.0.0.1:8080 \
ANTHROPIC_API_KEY=dummy \
ANTHROPIC_AUTH_TOKEN=dummy \
ANTHROPIC_DEFAULT_OPUS_MODEL=lfm-harbor \
ANTHROPIC_DEFAULT_SONNET_MODEL=lfm-harbor \
ANTHROPIC_DEFAULT_HAIKU_MODEL=lfm-harbor \
claude

The key and token can be any string, but ANTHROPIC_AUTH_TOKEN must be set. Mapping all three model tiers to the adapter keeps Claude Code's background calls on your server too.

Conversion

Converted from the PEFT adapter with llama.cpp's convert_lora_to_gguf.py, at F16:

python llama.cpp/convert_lora_to_gguf.py <adapter dir> --base <LiquidAI/LFM2.5-24B-A2B config dir> --outtype f16

Limitations

• Not re-evaluated in GGUF. The scores on the original adapter card come from vLLM at BF16. This file has not been evaluated through llama.cpp, and a quantized base will move the results further.
• Small, single-domain evaluation. See the original card: 100 data-analysis tasks, about ±5 points of noise on pass@1.