Llm Jp 4.1 33B Thinking — Hardware Requirements & GPU Compatibility
ChatFunctionsThe llm-jp-4.1-33b-thinking model comes from the Research and Development Center for Large Language Models at Japan's National Institute of Informatics. This dense reasoning architecture uses 33 billion parameters, spans 64 layers, and underwent multi-stage pre-training before alignment via supervised fine-tuning and direct preference optimization—without reinforcement learning. It handles both Japanese and English, with documentation noting code data in several programming languages. As a dense model of this scale, it requires at least a 24 GB GPU even when quantized, or alternatively a unified-memory machine. The system supports a context window of up to 65,536 tokens. Released under the Apache 2.0 license, it allows unrestricted commercial and research use, with publication in September 2026. Within the LLM-jp-4.1 thinking series, this is the largest dense model alongside its 8B and 32B-A3B siblings, serving as an upgrade to the earlier llm-jp-4-33b-thinking variant.
Decent on a 24 GB GPU
Runs on 24 GB (Q4_K_M · ~32 tok/s · 8K)Macs from 24 GB (IQ3_M · ~14 tok/s)
Apache 2.0
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run Llm Jp 4.1 33B Thinking
One recommended way. Every other method, OS and quant is in the step-by-step guide.Each way opens a step-by-step guide for your OS and quant.
Recommended
ollama run hf.co/modulon-labs/llm-jp-4.1-33b-thinking-GGUF:Q4_K_MQ4_K_M · needs ~24.5 GB · 16K context
Download: llm-jp-4.1-33b-thinking-Q4_K_M.gguf20.16 GBmodulon-labs
Ways to runOther ways
each opens step-by-stepRun in cloud
Runs on RTX 5090 (32 GB) (7 GB headroom) · Q4_K_M
- Generation speed
- ~56 tok/s
- generation speed
- Cost per 1M output tokens
- $2.66
- per 1M output tokens
How Much VRAM Does Llm Jp 4.1 33B Thinking Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | On RTX 4090 (reference) | |
|---|---|---|---|---|
| 15.0 GB | 23 GB | 2-bit · K-quant — 2-bit quantization with K-quant improvements | Fits: 32K context · ~44 tok/s | |
| 15.8 GB | 23.8 GB | 3-bit · i-quant medium — Importance-weighted 3-bit, medium | Fits: 32K context · ~42 tok/s | |
| 17.0 GB | 25.1 GB | 3-bit medium — 3-bit medium quantization | Fits: 16K context · ~39 tok/s | |
| 18.7 GB | 26.7 GB | 4-bit · i-quant compact — Importance-weighted 4-bit, compact | Fits: 16K context · ~35 tok/s | |
| 20.8 GB | 28.8 GB | 4-bit medium — 4-bit medium quantization — most popular sweet spot | Fits: 8K context · ~32 tok/s | |
| 24.5 GB | 32.6 GB | 5-bit medium — 5-bit medium quantization — good quality/size tradeoff | Does not fit: needs 28.3 GB | |
| 28.2 GB | 36.3 GB | 6-bit, very good — 6-bit quantization, very good quality | Does not fit: needs 32 GB | |
| 34.1 GB | 42.1 GB | 8-bit, near-lossless — 8-bit quantization, near-lossless | Does not fit: needs 37.8 GB | |
| 67.3 GB | 75.3 GB | 16-bit · bfloat — Brain floating point 16 — preferred for training | Does not fit: needs 71 GB |
Selected Q4_K_M
4-bit medium quantization24.5GBat 16K context
- Weights
- 19.9 GB
- KV cache
- 4.3 GB
- Overhead
- 0.3 GB
Fits RTX 4090 (reference) · fits up to 8K context · ~32 tok/s
≥50 tok/s needs ≥1.60 TB/s · RTX 5090 ~56
Which Hardware Can Run Llm Jp 4.1 33B Thinking?
Q4_K_M · 20.8 GBLlm Jp 4.1 33B Thinking (Q4_K_M) requires 20.8 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 28+ GB is recommended. Using the full 66K context window can add up to 16.6 GB, bringing total usage to 37.4 GB. 7 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti. 112 devices with unified memory can run Llm Jp 4.1 33B Thinking, including Mac Studio M5 Ultra (256 GB), Mac Studio M5 Ultra (512 GB), Mac Studio M1 Max (32 GB).
20 GB
Q4_K_M · best fit IQ3_M (19.6 GB)
1 GPU
RX 7900 XT~30Simulate output on RX 7900 XT32–40 GB
fits · 7.5 GB free · up to 32K ctx
1 GPU · 3 Mac classes · 9 device families
RTX 5090~56Simulate output on RTX 5090M5 Max*36 GB~16Simulate output on Mac Studio M5 Max (36 GB)M632 GB~6Simulate output on iMac 24" M6 (32 GB)RTX Spark N1X~9Simulate output on NVIDIA RTX Spark N1X Laptop (32 GB)Orin~6Simulate output on NVIDIA Jetson AGX Orin 32GB
By VRAM tier · typical GPUs and Macs at Q4_K_M
Similar models
6| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| MiniCPM V 4.7 35B A3BOpenBMB | — | 22.1 GB | ~169 | same size |
| NVIDIA Nemotron 3.5 Lightning 30B A3B BF16NVIDIA | — | 19.3 GB | ~197 | same size |
| Agents A1InternScience | — | 21.5 GB | ~171 | same size |
| Ornith 1.0 35BDeep Reinforce | — | 21.5 GB | ~171 | same size |
| EXAONE 4.5 33BLGAI-EXAONE | — | 23.9 GB | ~33 | same size |
| Gemma 4 31BGoogle | — | 21.7 GB | ~32 | same size |
Specifications
- Publisher
- llm-jp
- Parameters
- 33.2B
- Architecture
- LlamaForCausalLM
- Context Length
- 65,536 tokens
- Vocabulary Size
- 196,608
- Release Date
- 2026-09-16
- License
- Apache 2.0
Frequently Asked Questions
How much VRAM does Llm Jp 4.1 33B Thinking need?
Llm Jp 4.1 33B Thinking requires 20.8 GB of VRAM at Q4_K_M, or 67.3 GB at BF16. Full 66K context adds up to 16.6 GB (37.4 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 33.2B × 4.8 bits ÷ 8 = 19.9 GB
KV Cache + Overhead ≈ 0.9 GB (at 2K context + ~0.3 GB framework)
Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.
KV Cache + Overhead ≈ 17.5 GB (at full 66K context)
VRAM usage by quantization
Can NVIDIA GeForce RTX 4090 run Llm Jp 4.1 33B Thinking?
Yes, at Q4_K_M (20.8 GB) or lower. Higher quantizations like Q5_K_M (24.5 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.
What's the best quantization for Llm Jp 4.1 33B Thinking?
For Llm Jp 4.1 33B Thinking, Q4_K_M (20.8 GB) offers the best balance of quality and VRAM usage. Q5_K_M (24.5 GB) provides better quality if you have the VRAM. The smallest option is Q2_K at 15.0 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run Llm Jp 4.1 33B Thinking on a Mac?
Yes — Mac Mini M5 Pro (24 GB) and 115 other Macs can run Llm Jp 4.1 33B Thinking, the smallest at IQ3_M. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 20.8 GB of usable unified memory (RAM minus macOS overhead), from Mac Studio M1 Max (32 GB).
Can I run Llm Jp 4.1 33B Thinking locally?
Yes — Llm Jp 4.1 33B Thinking can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is Llm Jp 4.1 33B Thinking?
At Q4_K_M, Llm Jp 4.1 33B Thinking can reach ~231 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~32 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = (bandwidth GB/s ÷ model GB) × efficiency
Example: NVIDIA B200 → 8000 ÷ 20.8 × 0.65 = ~250 tok/s
Estimated speed at Q4_K_M (20.8 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of Llm Jp 4.1 33B Thinking?
At Q4_K_M, the download is about 19.93 GB. The full-precision BF16 version is 66.44 GB. The smallest option (Q2_K) is 14.12 GB.
Which GPUs can run Llm Jp 4.1 33B Thinking?
7 consumer GPUs can run Llm Jp 4.1 33B Thinking at Q4_K_M. Top options include NVIDIA GeForce RTX 5090, AMD Radeon RX 7900 XTX, NVIDIA GeForce RTX 3090. 1 GPU has plenty of headroom for comfortable inference.
Which devices can run Llm Jp 4.1 33B Thinking?
112 devices with unified memory can run Llm Jp 4.1 33B Thinking at Q4_K_M (20.8 GB), including AMD Ryzen 7 8845HS (Hawk Point) Laptop, AMD Ryzen AI 7 350 (Krackan Point) Laptop, AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10. Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.