I've been experimenting with a new quantization approach that selectively elevates the precision of key layers beyond what the default IMatrix configuration provides.
In my testing, standard IMatrix quantization underperforms at lower bit depths, especially with Mixture of Experts (MoE) models. To address this, I'm using the --tensor-type option in llama.cpp to manually "bump" important layers to higher precision. You can see the implementation here:
👉 Layer bumping with llama.cpp
While this does increase model file size, it significantly improves precision for a given quantization level.
I'd love your feedback—have you tried this? How does it perform for you?
SmallThinker is a family of on-device native Mixture-of-Experts (MoE) language models specially designed for local deployment,
co-developed by the IPADS and School of AI at Shanghai Jiao Tong University and Zenergize AI.
Designed from the ground up for resource-constrained environments,
SmallThinker brings powerful, private, and low-latency AI directly to your personal devices,
without relying on the cloud.
While frontier large language models (LLMs) continue to push capability boundaries, their deployment remains confined to GPU-powered cloud infrastructure. We challenge this paradigm with SmallThinker, a family of LLMs natively designed - not adapted - for the unique constraints of local devices: weak computational power, limited memory, and slow storage. Unlike traditional approaches that mainly compress existing models built for clouds, we architect SmallThinker from the ground up to thrive within these limitations. Our innovation lies in a deployment-aware architecture that transforms constraints into design principles. First, We introduce a two-level sparse structure combining fine-grained Mixture-of-Experts (MoE) with sparse feed-forward networks, drastically reducing computational demands without sacrificing model capacity. Second, to conquer the I/O bottleneck of slow storage, we design a pre-attention router that enables our co-designed inference engine to prefetch expert parameters from storage while computing attention, effectively hiding storage latency that would otherwise cripple on-device inference. Third, for memory efficiency, we utilize NoPE-RoPE hybrid sparse attention mechanism to slash KV cache requirements. We release SmallThinker-4B-A0.6B and SmallThinker-21B-A3B, which achieve state-of-the-art performance scores and even outperform larger LLMs. Remarkably, our co-designed system mostly eliminates the need for expensive GPU hardware: with Q4_0 quantization, both models exceed 20 tokens/s on ordinary consumer CPUs, while consuming only 1GB and 8GB of memory respectively. SmallThinker is publicly available at this http URL and this http URL .
Performance
Note: The model is trained mainly on English.
Model
MMLU
GPQA-diamond
MATH-500
IFEVAL
LIVEBENCH
HUMANEVAL
Average
SmallThinker-21BA3B-Instruct
84.43
55.05
82.4
85.77
60.3
89.63
76.26
Gemma3-12b-it
78.52
34.85
82.4
74.68
44.5
82.93
66.31
Qwen3-14B
84.82
50
84.6
85.21
59.5
88.41
75.42
Qwen3-30BA3B
85.1
44.4
84.4
84.29
58.8
90.24
74.54
Qwen3-8B
81.79
38.89
81.6
83.92
49.5
85.9
70.26
Phi-4-14B
84.58
55.45
80.2
63.22
42.4
87.2
68.84
For the MMLU evaluation, we use a 0-shot CoT setting.
All models are evaluated in non-thinking mode.
Speed
Model
Memory(GiB)
i9 14900
1+13 8ge4
rk3588 (16G)
Raspberry PI 5
SmallThinker 21B+sparse
11.47
30.19
23.03
10.84
6.61
SmallThinker 21B+sparse+limited memory
limit 8G
20.30
15.50
8.56
-
Qwen3 30B A3B
16.20
33.52
20.18
9.07
-
Qwen3 30B A3B+limited memory
limit 8G
10.11
0.18
6.32
-
Gemma 3n E2B
1G, theoretically
36.88
27.06
12.50
6.66
Gemma 3n E4B
2G, theoretically
21.93
16.58
7.37
4.01
Note: i9 14900, 1+13 8ge4 use 4 threads, others use the number of threads that can achieve the maximum speed. All models here have been quantized to q4_0.
You can deploy SmallThinker with offloading support using PowerInfer
Model Card
Architecture
Mixture-of-Experts (MoE)
Total Parameters
21B
Activated Parameters
3B
Number of Layers
52
Attention Hidden Dimension
2560
MoE Hidden Dimension (per Expert)
768
Number of Attention Heads
28
Number of KV Heads
4
Number of Experts
64
Selected Experts per Token
6
Vocabulary Size
151,936
Context Length
16K
Attention Mechanism
GQA
Activation Function
ReGLU
How to Run
Transformers
transformers==4.53.3 is required, we are actively working to support the latest version.
The following contains a code snippet illustrating how to use the model generate content based on given inputs.
python
1from transformers import AutoModelForCausalLM, AutoTokenizer
2import torch
34path ="PowerInfer/SmallThinker-21BA3B-Instruct"5device ="cuda"67tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True)8model = AutoModelForCausalLM.from_pretrained(path, torch_dtype=torch.bfloat16, device_map=device, trust_remote_code=True)910messages =[11{"role":"user","content":"Give me a short introduction to large language model."},12]13model_inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to(device)1415model_outputs = model.generate(16 model_inputs,17 do_sample=True,18 max_new_tokens=102419)2021output_token_ids =[22 model_outputs[i][len(model_inputs[i]):]for i inrange(len(model_inputs))23]2425responses = tokenizer.batch_decode(output_token_ids, skip_special_tokens=True)[0]26print(responses)27
ModelScope
ModelScope adopts Python API similar to (though not entirely identical to) Transformers. For basic usage, simply modify the first line of the above code as follows:
from modelscope import AutoModelForCausalLM, AutoTokenizer
Statement
Due to the constraints of its model size and the limitations of its training data, its responses may contain factual inaccuracies, biases, or outdated information.
Users bear full responsibility for independently evaluating and verifying the accuracy and appropriateness of all generated content.
SmallThinker does not possess genuine comprehension or consciousness and cannot express personal opinions or value judgments.
🚀 If you find these models useful
Help me test my AI-Powered Quantum Network Monitor Assistant with quantum-ready security checks:
The full Open Source Code for the Quantum Network Monitor Service available at my github repos ( repos with NetworkMonitor in the name) : Source Code Quantum Network Monitor. You will also find the code I use to quantize the models if you want to do it yourself GGUFModelBuilder
💬 How to test:
Choose an AI assistant type:
TurboLLM (GPT-4.1-mini)
HugLLM (Hugginface Open-source models)
TestLLM (Experimental CPU-only)
What I’m Testing
I’m pushing the limits of small open-source models for AI network monitoring, specifically:
Function calling against live network services
How small can a model go while still handling:
Automated Nmap security scans
Quantum-readiness checks
Network Monitoring tasks
🟡 TestLLM – Current experimental model (llama.cpp on 2 CPU threads on huggingface docker space):
✅ Zero-configuration setup
⏳ 30s load time (slow inference but no API costs) . No token limited as the cost is low.
🔧 Help wanted! If you’re into edge-device AI, let’s collaborate!
Other Assistants
🟢 TurboLLM – Uses gpt-4.1-mini :
**It performs very well but unfortunatly OpenAI charges per token. For this reason tokens usage is limited.
Create custom cmd processors to run .net code on Quantum Network Monitor Agents
Real-time network diagnostics and monitoring
Security Audits
Penetration testing (Nmap/Metasploit)
🔵 HugLLM – Latest Open-source models:
🌐 Runs on Hugging Face Inference API. Performs pretty well using the lastest models hosted on Novita.
💡 Example commands you could test:
"Give me info on my websites SSL certificate"
"Check if my server is using quantum safe encyption for communication"
"Run a comprehensive security audit on my server"
'"Create a cmd processor to .. (what ever you want)" Note you need to install a Quantum Network Monitor Agent to run the .net code on. This is a very flexible and powerful feature. Use with caution!
Final Word
I fund the servers used to create these model files, run the Quantum Network Monitor service, and pay for inference from Novita and OpenAI—all out of my own pocket. All the code behind the model creation and the Quantum Network Monitor project is open source. Feel free to use whatever you find helpful.
If you appreciate the work, please consider buying me a coffee ☕. Your support helps cover service costs and allows me to raise token limits for everyone.
I'm also open to job opportunities or sponsorship.
