LFM2.5-24B-A2B Q4_K_M (llama.cpp, expert offload)

Metrics

Checklist score

– / 100

Design (rubric, 0–5)

–

Stop reason

killed

Model time

20s

Tool time (wall minus model)

7s

Wall time

27s

Tool calls

0

Tool errors

0

Errors by type

–

Loop-guard events

0

Tokens in / out

10,173.0 / 339.0

Thinking tokens

–

Decode / prefill tok/s

47.9 / 784.6

Prefix reuse

0.000

Peak VRAM (GPU / engine)

10,888 / 10,880 MiB

Peak engine RAM

14,991 MiB

Engine

llamacpp

Model

LFM2.5-24B-A2B Q4_K_M (llama.cpp, expert offload) (lfm25-24b-a2b)

Settings

{"thinking": "always on (LFM2.5 template opens <think>), server --reasoning-budget 3072", "reasoning_budget_tokens": 3072, "max_tokens": 7168, "sampling": "LFM2.5 card: temp 0.1, top_k 50, repeat_penalty 1.05 (no top_p/min_p)", "context": "65536 (q8_0 KV, --fit on expert offload), 32768 fallback", "tool_format": "LFM2 pythonic, parsed by llama.cpp (--jinja)", "research": "research/lfm2.5-24b-a2b.md", "max_tokens_boost_ceiling": 16384, "max_tokens_route": "agent cap (hermes_learn --max-tokens), not extra_body (2026-10-09)"}

Node

omarchy (RTX 3060 12 GB)

Prompt

creative-v4

Started

2026-10-14T10:15:37Z

Feature scores