Huihui MoE 23B A4B Abliterated — Hardware Requirements & GPU Compatibility
Chat
Huihui MoE 23B A4B Abliterated is a 23.2B-parameter open language model from huihui-ai. It supports a context window of up to 40,960 tokens. At Q4_K_M it needs about 14.43 GB of VRAM — see which GPUs and Macs can run it below.
Runs great on a 24 GB GPU
Runs from 12 GB (Q2_K · 16K)Best on 24 GB (Q4_K_M · ~162 tok/s · 40K)Macs from 16 GB (Q2_K · ~47 tok/s)
Apache 2.0
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run Huihui MoE 23B A4B Abliterated
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 huihui-ai/Huihui-MoE-23B-A4B-abliterated --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 A5000 (24 GB) (8 GB headroom) · Q4_K_M
- Generation speed
- ~134 tok/s
- generation speed
- Cost per 1M output tokens
- $0.33
- per 1M output tokens
How Much VRAM Does Huihui MoE 23B A4B Abliterated Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | On RTX 4090 (reference) | |
|---|---|---|---|---|
| 10.4 GB | 13.2 GB | 2-bit · K-quant — 2-bit quantization with K-quant improvements | Fits: 40K context · ~201 tok/s | |
| 11.8 GB | 14.7 GB | 3-bit medium — 3-bit medium quantization | Fits: 40K context · ~185 tok/s | |
| 14.4 GB | 17.3 GB | 4-bit medium — 4-bit medium quantization — most popular sweet spot | Fits: 40K context · ~162 tok/s | |
| 17.1 GB | 19.9 GB | 5-bit medium — 5-bit medium quantization — good quality/size tradeoff | Fits: 40K context · ~145 tok/s | |
| 19.7 GB | 22.5 GB | 6-bit, very good — 6-bit quantization, very good quality | Fits: 40K context · ~130 tok/s | |
| 23.7 GB | 26.6 GB | 8-bit, near-lossless — 8-bit quantization, near-lossless | Fits: 4K context · ~113 tok/s | |
| 47.0 GB | 49.8 GB | 16-bit · bfloat — Brain floating point 16 — preferred for training | Does not fit: needs 48.3 GB |
Selected Q4_K_M
4-bit medium quantization15.8GBat 16K context
- Weights
- 13.9 GB
- KV cache
- 1.5 GB
- Overhead
- 0.3 GB
Fits RTX 4090 (reference) · 8.3 GB free · ~162 tok/s
≥50 tok/s needs ≥234 GB/s · RTX 5090 ~231, RTX 3090 Ti ~162
Which Hardware Can Run Huihui MoE 23B A4B Abliterated?
Q4_K_M · 14.4 GBHuihui MoE 23B A4B Abliterated (Q4_K_M) requires 14.4 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 19+ GB is recommended. Using the full 41K context window can add up to 3.6 GB, bringing total usage to 18.0 GB. 26 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti, NVIDIA GeForce RTX 5080. 124 devices with unified memory can run Huihui MoE 23B A4B Abliterated, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Mini M5 Pro (24 GB).
16–18 GB
16–18 GB tier →fits · 0.3 GB free · up to 16K ctx
18 GPUs · 1 Mac class
RTX 5080~157Simulate output on RTX 5080RTX 5070 Ti~150Simulate output on RTX 5070 TiRTX 4080 SUPER~130Simulate output on RTX 4080 SUPERM3 Pro*18 GB~30Simulate output on MacBook Pro 14" M3 Pro (18 GB)20 GB
fits · 4.3 GB free · up to 40K ctx
1 GPU
RX 7900 XT~78Simulate output on RX 7900 XT32–40 GB
fits · 16.3 GB free · up to 40K ctx
1 GPU · 3 Mac classes · 4 device families
RTX 5090~231Simulate output on RTX 5090M5 Max36 GB~64Simulate output on Mac Studio M5 Max (36 GB)M532 GB~31Simulate output on MacBook Air 13" M5 (32 GB)Orin~44Simulate output on NVIDIA Jetson AGX Orin 32GBSnapdragon X Elite (X1E-80-100)~29Simulate output on Snapdragon X Elite Copilot+ PC
By VRAM tier · typical GPUs and Macs at Q4_K_M
Similar models
3| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| Magistral Small 2509Mistral AI | 85.6 | 18.1 GB | ~43 | same size |
| LFM2.5 24B A2BLiquid AI | — | 15 GB | ~221 | same size |
| ERNIE 4.5 21B A3B ThinkingBaidu | — | 14.3 GB | ~213 | same size |
Specifications
- Publisher
- huihui-ai
- Parameters
- 23.2B
- Active parameters
- 4.4B per token
- Architecture
- Qwen3MoeForCausalLM
- Context Length
- 40,960 tokens
- Vocabulary Size
- 151,936
- Release Date
- 2025-06-21
- License
- Apache 2.0
Frequently Asked Questions
How much VRAM does Huihui MoE 23B A4B Abliterated need?
Huihui MoE 23B A4B Abliterated requires 14.4 GB of VRAM at Q4_K_M, or 47.0 GB at BF16. Full 41K context adds up to 3.6 GB (18.0 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 23.2B × 4.8 bits ÷ 8 = 13.9 GB
KV Cache + Overhead ≈ 0.5 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 ≈ 4.1 GB (at full 41K context)
VRAM usage by quantization
Can NVIDIA GeForce RTX 4090 run Huihui MoE 23B A4B Abliterated?
Yes, at Q8_0 (23.7 GB) or lower. Higher quantizations like BF16 (47.0 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.
What's the best quantization for Huihui MoE 23B A4B Abliterated?
For Huihui MoE 23B A4B Abliterated, Q4_K_M (14.4 GB) offers the best balance of quality and VRAM usage. Q5_K_M (17.1 GB) provides better quality if you have the VRAM. The smallest option is Q2_K at 10.4 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run Huihui MoE 23B A4B Abliterated on a Mac?
Yes — Mac Mini M2 Pro (16 GB) and 133 other Macs can run Huihui MoE 23B A4B Abliterated, 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 14.4 GB of usable unified memory (RAM minus macOS overhead), from MacBook Pro 14" M3 Pro (18 GB).
Can I run Huihui MoE 23B A4B Abliterated locally?
Yes — Huihui MoE 23B A4B Abliterated can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is Huihui MoE 23B A4B Abliterated?
At Q4_K_M, Huihui MoE 23B A4B Abliterated can reach ~129 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~162 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 → 2.7 GB active ÷ (8000 × 0.65) = 0.527 ms, plus 36 layers × 0.055 ms = 1.980 ms, so 1000 ÷ 2.507 ms = ~399 tok/s
Estimated speed at Q4_K_M (14.4 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of Huihui MoE 23B A4B Abliterated?
At Q4_K_M, the download is about 13.94 GB. The full-precision BF16 version is 46.48 GB. The smallest option (Q2_K) is 9.88 GB.
Which GPUs can run Huihui MoE 23B A4B Abliterated?
26 consumer GPUs can run Huihui MoE 23B A4B Abliterated at Q4_K_M. Top options include AMD Radeon RX 7900 XTX, NVIDIA GeForce RTX 3090, NVIDIA GeForce RTX 3090 Ti, AMD Radeon RX 6800. 7 GPUs have plenty of headroom for comfortable inference.
Which devices can run Huihui MoE 23B A4B Abliterated?
124 devices with unified memory can run Huihui MoE 23B A4B Abliterated at Q4_K_M (14.4 GB), including AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.