Only with CPU offload

LLM-jp 4 33B Thinking at Q4_K_M needs 21.3 GB but only 9.7 GB is addressable, so about 62% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 2.9 tokens per second — usable for batch work, painful for chat.

38% on GPU 8K context Q4_K_M · 18.7 GB Apache 2.0 Released 14 Aug 2026 New this week Not in the Ollama library

Japan’s national-institute reasoning model, Japanese and English. A plain dense Llama-style 33B: Q4 is a tight 24 GB fit.

What hardware do I need for LLM-jp 4 33B Thinking? →

The VRAM budget

weights 18.7 GB

Weights 18.7 GB KV cache @ 8K 2.00 GB Runtime overhead 0.6 GB Over budget 11.6 GB past 9.7 GB

Quantisation ladder

Quant	Weights	Total @ 8K	Max context	Tok/s	Quality	Fit
Q8_0	32.9 GB	35.5 GB	—	~1.4	−0.1% ppl	25.8 GB over
Q6_K	25.4 GB	28.0 GB	—	~1.9	−0.4% ppl	18.3 GB over
Q5_K_M	21.9 GB	24.5 GB	—	~2.3	−0.8% ppl	14.8 GB over
Q4_K_M	18.7 GB	21.3 GB	—	~2.9	−1.9% ppl	11.6 GB over
Q3_K_M	15.1 GB	17.7 GB	—	~4.1	−5.4% ppl	8.0 GB over
Q2_K	12.9 GB	15.5 GB	—	~5.4	−15% ppl	5.8 GB over

Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it.

How to run it

llama.cpp LM Studio vLLM

terminal

$ llama-server \
    -hf llm-jp/llm-jp-4-33b-thinking:Q4_K_M \
    -c 8192 -ngl 24

The engine underneath most of the others. Every knob is exposed. More on llama.cpp.

01Download is 18.7 GB. Keep it on an SSD — a first load off a spinning disk takes minutes.

02Close anything else holding VRAM. A browser with hardware acceleration can sit on 1–2 GB.

03If it falls back to CPU silently, drop the context first, then step down a quantisation.

See all models for this rig Compare with another model