Install from WinGet (Windows)
winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf h4rm0n1c/qwen3.5-24b-a10b-IQ4_NL-GGUF:IQ4_NL# Run inference directly in the terminal:
llama cli -hf h4rm0n1c/qwen3.5-24b-a10b-IQ4_NL-GGUF:IQ4_NLUse pre-built binary
# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf h4rm0n1c/qwen3.5-24b-a10b-IQ4_NL-GGUF:IQ4_NL# Run inference directly in the terminal:
./llama-cli -hf h4rm0n1c/qwen3.5-24b-a10b-IQ4_NL-GGUF:IQ4_NLBuild from source code
git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf h4rm0n1c/qwen3.5-24b-a10b-IQ4_NL-GGUF:IQ4_NL# Run inference directly in the terminal:
./build/bin/llama-cli -hf h4rm0n1c/qwen3.5-24b-a10b-IQ4_NL-GGUF:IQ4_NLUse Docker
docker model run hf.co/h4rm0n1c/qwen3.5-24b-a10b-IQ4_NL-GGUF:IQ4_NLqwen3.5b-24b-a10b IQ4_NL GGUF
Quanter's Note: Sibling of sandeshrajx/Qwen3.5-24B-A3B-REAP-0.32, this one has been passed over completely for reasoning distills, I'd be interested to see what difference 7b more active experts can make.
GGUF quantization of sandeshrajx/qwen3.5b-24b-a10b — a 24B-parameter MoE model with 10B active parameters per token. Architecture is Qwen3.5 MoE.
Source: sandeshrajx/qwen3.5b-24b-a10b
Converted by: QuantZhai benchmark pipeline
Quantization: IQ4_NL (importance-matrix 4-bit non-linear)
Model Details
| Property | Value |
|---|---|
| Architecture | Qwen3.5 MoE (Dense + Mamba-2 SSM interleaved) |
| Parameters | 24B total, 10B active per token |
| Experts | 39, 8 active per token |
| Context length | 262144 (256K) |
| Hidden size | 3072 |
| Attention heads | 32, KV heads = 2 |
| Head dim | 256 |
| RoPE | MRope (multimodal), theta = 10,000,000 |
| SSM | Mamba-2 inspired conv/state-space per 4th layer |
| Quantization | IQ4_NL (4.50 bpw) |
| File size | 13.9 GB |
| Tokenizer | Qwen2 (GPT-2 based BPE, vocab 248,320) |
Benchmarks
Hardware: dual-GPU (RTX 3080 10GB + V100-SXM2 32GB, 42 GB total)
Engine: llama.cpp with TurboQuant KV (q8_0 K / turbo3 V)
Perplexity: macvox68 code corpus, ctx=4096, stride=512
| Metric | Cold | Warm |
|---|---|---|
| PPL | 8.0217 | 3.9654 |
| TPS | 26.1 tok/s | 31.4 tok/s |
| TTFT | 1919 ms | 1594 ms |
QuantZhai Ranking
Rank #30 of 46 — combined score 50.5 (equal-weight: TPS, PPL, convergence).
Usage
llama-cli -m qwen3.5b-24b-a10b-IQ4_NL.gguf \
-p "Write a mergesort in Python" \
-n 1024 -t 12 --temp 0.6 --top-p 0.95
llama-server -m qwen3.5b-24b-a10b-IQ4_NL.gguf \
--host 0.0.0.0 --port 8080 -ngl 99 -t 12 \
--cache-type-k q8_0 --cache-type-v turbo3
Recommended: temp 0.6, top-p 0.95, context up to 256K.
License
MIT (this quantization).
Source model by sandeshrajx — review its license separately.
- Downloads last month
- 94
4-bit
Model tree for h4rm0n1c/qwen3.5-24b-a10b-IQ4_NL-GGUF
Base model
sandeshrajx/qwen3.5b-24b-a10b
Install (macOS, Linux)
# Start a local OpenAI-compatible server with a web UI: llama serve -hf h4rm0n1c/qwen3.5-24b-a10b-IQ4_NL-GGUF:IQ4_NL# Run inference directly in the terminal: llama cli -hf h4rm0n1c/qwen3.5-24b-a10b-IQ4_NL-GGUF:IQ4_NL