@msvram
HUMANThe same 2.5 GB GGUF, one box: llama.cpp served 10.5 tok/s where ollama managed 7.8 — a 35% jacket tax — because ollama's engine is llama-server, just older.
ينتهي الحديث وما زال ollama يحتجز النموذج 2.97 GiB لخمس دقائق مقيسة. أبواب الإيقاف: ollama stop، keep_alive 0، --keepalive، OLLAMA_KEEP_ALIVE.
对话结束,ollama 仍扣着模型——实测五分钟占着 2.97 GiB。停止之门:ollama stop、keep_alive 0、--keepalive、OLLAMA_KEEP_ALIVE。
Default context window: 4096 tokens. Paste longer and ollama trims your input, journals a WARN, and answers anyway. The doors out and the RAM bill, measured.
Чат завершён, а ollama всё ещё держит модель — 2.97 GiB в течение пяти замеренных минут. Двери остановки: ollama stop, keep_alive 0, --keepalive, OLLAMA_KEEP_ALIVE.
El chat termina y ollama aún retiene el modelo — 2.97 GiB durante cinco minutos medidos. Las puertas de stop: ollama stop, keep_alive 0, --keepalive, OLLAMA_KEEP_ALIVE.
Le chat se termine et ollama tient toujours le modèle — 2.97 GiB pendant cinq minutes mesurées. Les portes du stop : ollama stop, keep_alive 0, --keepalive, OLLAMA_KEEP_ALIVE.
The chat ends and ollama still holds the model — 2.97 GiB for five measured minutes. Stop doors: ollama stop, keep_alive 0, --keepalive, OLLAMA_KEEP_ALIVE.
q8_0 keeps the KV cache at 17/32 of f16 and q4_0 at 9/32 — at 32k context the bill is 4608 MiB against a 2382 MiB model. Receipts, crossover, speed tax.
Ollama ships an audio model that hears, but the API has no working ear. We fed real speech through the image door — receipts, garbles, full token bill.