Ah — that 465 MB file is the MTP draft model (the little speculative-decoding helper inside the MTP/ folder), not the actual model. That's why it's so small.
The full "12 GB version" you want is gemma4-v2-Q8_0.gguf in the root folder (11.8 GB). Grab that one and you're set. If you want something lighter that still runs great, gemma4-v2-Q4_K_M.gguf (6.9 GB) is the sweet spot.
The MTP file is optional — it only helps if you're running speculative decoding in llama.cpp to speed generation up, and it loads alongside the main model, not instead of it.
and yeah — v2 is the upgrade for agentic / tool-driven coding, that's what it was tuned for. If v1's been working well for you it's still solid; v2 just pushes the agentic + coding side further.
but from "Use this model", ollama run hf.co/yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-GGUF:Q8_0, this pulls the 400mb file, shouldn't it pull the 12gb model?
@ly2025 You're right — that's a real gotcha, it's pulling the wrong file. The repo has two files with "Q8_0" in the name: the actual model (gemma4-v2-Q8_0.gguf, ~11.8 GB) and the tiny MTP draft head (MTP/gemma-4-12B-it-MTP-Q8_0.gguf, ~400 MB). Ollama's :Q8_0 tag is ambiguous, so it grabs the small MTP one. (A couple of other tools auto-pick it too — it's on my list to make the repo less confusing.)
Reliable fix — download the exact main file and import it with a Modelfile so the auto-resolver can't pick wrong:
Then a Modelfile: FROM ./gemma4-v2-Q8_0.gguf PARAMETER num_ctx 8192 PARAMETER temperature 1.0 PARAMETER top_p 0.95 PARAMETER top_k 64 PARAMETER repeat_penalty 1.1 ollama create gemma4-v2 -f Modelfile ollama run gemma4-v2
Two notes: for proper tool-calling/agentic urom the official Gemma 4 (ollama pull gemma4→ ollama show --modelfile gemma4) into your Modelfile — a bare import can report "does not support chat." And keep num_ctx capped like above so Gemma 4's 256K context doesn't blow up your KV cache. For the most reliable tool parsing overall, llama.cpp llama-server --jinja is still the best path.
