Llm Jp 4.1 32B A3b Thinking — Hardware Requirements & GPU Compatibility
ChatFunctionsllm-jp-4.1-32b-a3b-thinking is a Mixture-of-Experts reasoning model from Japan's National Institute of Informatics Research and Development Center for Large Language Models, featuring 32 billion total parameters with 3.8 billion active ones. It employs 128 routed experts where only 8 stay active, was pre-trained and mid-trained on 11.7 trillion tokens, then post-trained using supervised fine-tuning and direct preference optimization without reinforcement learning. The model targets Japanese and English while the card notes code data across several programming languages; with only 3.8 billion parameters active per token, decoding runs fast even though full weights still demand a large GPU or a 32 GB-class unified-memory machine, even when quantized. The architecture supports a 65,536 token context window. Released under the Apache 2.0 license for unrestricted commercial and research use, it came out in September 2026. As part of the LLM-jp-4.1 thinking series, this MoE model joins 8B and 33B dense variants and serves as an update to the earlier llm-jp-4 version.
Runs great on a 24 GB GPU
Runs from 16 GB (Q2_K · 32K)Best on 24 GB (Q4_K_M · ~188 tok/s · 64K)Macs from 18 GB (Q2_K · ~45 tok/s)
Apache 2.0
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run Llm Jp 4.1 32B A3b 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
vllm serve llm-jp/llm-jp-4.1-32b-a3b-thinking --max-model-len 16384vLLM repo with its own quantisation
Download: Weights on Hugging Face
Ways to runOther ways
each opens step-by-stepRun in cloud
Runs on RTX A6000 (48 GB) (27 GB headroom) · Q4_K_M
- Generation speed
- ~155 tok/s
- generation speed
- Cost per 1M output tokens
- $0.59
- per 1M output tokens
How Much VRAM Does Llm Jp 4.1 32B A3b Thinking Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | On RTX 4090 (reference) | |
|---|---|---|---|---|
| 14.0 GB | 15 GB | 2-bit · K-quant — 2-bit quantization with K-quant improvements | Fits: 64K context · ~232 tok/s | |
| 16.0 GB | 17 GB | 3-bit medium — 3-bit medium quantization | Fits: 64K context · ~214 tok/s | |
| 19.6 GB | 20.7 GB | 4-bit medium — 4-bit medium quantization — most popular sweet spot | Fits: 64K context · ~188 tok/s | |
| 23.3 GB | 24.3 GB | 5-bit medium — 5-bit medium quantization — good quality/size tradeoff | Fits: 16K context · ~167 tok/s | |
| 26.9 GB | 27.9 GB | 6-bit, very good — 6-bit quantization, very good quality | Does not fit: needs 27.4 GB | |
| 32.5 GB | 33.5 GB | 8-bit, near-lossless — 8-bit quantization, near-lossless | Does not fit: needs 33 GB | |
| 64.7 GB | 65.7 GB | 16-bit · bfloat — Brain floating point 16 — preferred for training | Does not fit: needs 65.1 GB |
Selected Q4_K_M
4-bit medium quantization20.1GBat 16K context
- Weights
- 19.3 GB
- KV cache
- 0.5 GB
- Overhead
- 0.3 GB
Fits RTX 4090 (reference) · 3.9 GB free · ~188 tok/s
≥50 tok/s needs ≥197 GB/s · RTX 5090 ~265, RTX 3090 Ti ~188
Which Hardware Can Run Llm Jp 4.1 32B A3b Thinking?
Q4_K_M · 19.6 GBLlm Jp 4.1 32B A3b Thinking (Q4_K_M) requires 19.6 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 26+ GB is recommended. Using the full 66K context window can add up to 2.1 GB, bringing total usage to 21.7 GB. 8 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti. 137 devices with unified memory can run Llm Jp 4.1 32B A3b Thinking, including Mac Studio M5 Ultra (256 GB), Mac Studio M5 Ultra (512 GB), NVIDIA RTX Spark N1X Laptop (32 GB).
20 GB
tight · up to 8K ctx
1 GPU
RX 7900 XT~89Simulate output on RX 7900 XT32–40 GB
fits · 11.9 GB free · up to 64K ctx
1 GPU · 3 Mac classes · 9 device families
RTX 5090~265Simulate output on RTX 5090M5 Max*36 GB~73Simulate output on Mac Studio M5 Max (36 GB)M632 GB~38Simulate output on iMac 24" M6 (32 GB)RTX Spark N1X~73Simulate output on NVIDIA RTX Spark N1X Laptop (32 GB)Orin~52Simulate 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 |
|---|---|---|---|---|
| Hy Memory 30B A3BTencent | — | 19.1 GB | ~184 | same size |
| Granite Switch 4.1 30BIBM | — | 22.7 GB | ~35 | same size |
| Muse Glimmer 30BMeta | — | 19.6 GB | ~36 | same size |
| NVIDIA Nemotron 3.5 Lightning 30B A3B BF16NVIDIA | — | 19.3 GB | ~197 | same size |
| Hy MT2 30B A3BTencent | — | 19.1 GB | ~170 | same size |
| EXAONE 4.5 33BLGAI-EXAONE | — | 23.9 GB | ~33 | same size |
Specifications
- Publisher
- llm-jp
- Parameters
- 32.1B
- Active parameters
- 3.8B per token
- Architecture
- Qwen3MoeForCausalLM
- 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 32B A3b Thinking need?
Llm Jp 4.1 32B A3b Thinking requires 19.6 GB of VRAM at Q4_K_M, or 64.7 GB at BF16. Full 66K context adds up to 2.1 GB (21.7 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 32.1B × 4.8 bits ÷ 8 = 19.3 GB
KV Cache + Overhead ≈ 0.3 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 ≈ 2.4 GB (at full 66K context)
VRAM usage by quantization
Can NVIDIA GeForce RTX 4090 run Llm Jp 4.1 32B A3b Thinking?
Yes, at Q5_K_M (23.3 GB) or lower. Higher quantizations like Q6_K (26.9 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.
What's the best quantization for Llm Jp 4.1 32B A3b Thinking?
For Llm Jp 4.1 32B A3b Thinking, Q4_K_M (19.6 GB) offers the best balance of quality and VRAM usage. Q5_K_M (23.3 GB) provides better quality if you have the VRAM. The smallest option is Q2_K at 14.0 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run Llm Jp 4.1 32B A3b Thinking on a Mac?
Yes — MacBook Pro 14" M3 Pro (18 GB) and 117 other Macs can run Llm Jp 4.1 32B A3b Thinking, the smallest at Q2_K. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 19.6 GB of usable unified memory (RAM minus macOS overhead), from Mac Mini M5 Pro (24 GB).
Can I run Llm Jp 4.1 32B A3b Thinking locally?
Yes — Llm Jp 4.1 32B A3b Thinking can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is Llm Jp 4.1 32B A3b Thinking?
At Q4_K_M, Llm Jp 4.1 32B A3b Thinking can reach ~145 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~188 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = 1000 ÷ (active GB ÷ (bandwidth GB/s × efficiency) × 1000 + layers × routing ms)
Mixture-of-Experts: just the active experts get read for each token, along with a fixed routing cost per layer.
Example: NVIDIA B200 → 2.3 GB active ÷ (8000 × 0.65) = 0.45003 ms, plus 32 layers × 0.055 ms = 1.76000 ms, so 1000 ÷ 2.21003 ms = ~452 tok/s
Estimated speed at Q4_K_M (19.6 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 32B A3b Thinking?
At Q4_K_M, the download is about 19.28 GB. The full-precision BF16 version is 64.28 GB. The smallest option (Q2_K) is 13.66 GB.
Which GPUs can run Llm Jp 4.1 32B A3b Thinking?
8 consumer GPUs can run Llm Jp 4.1 32B A3b Thinking at Q4_K_M. Top options include NVIDIA GeForce RTX 5090, AMD Radeon RX 7900 XT, AMD Radeon RX 7900 XTX. 1 GPU has plenty of headroom for comfortable inference.
Which devices can run Llm Jp 4.1 32B A3b Thinking?
137 devices with unified memory can run Llm Jp 4.1 32B A3b Thinking at Q4_K_M (19.6 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.