GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
DMindAI
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
DMindAI
Model Tree
Qwen/Qwen3.5-4B
Input Modalities
Output Modalities
Supported Functionality
The DMind lineage was born from a singular conviction that decentralized finance deserves decentralized intelligence. This journey began with DMind-1, which shattered the monopoly of closed-source AI by releasing the world's first Web3-native LLM. It continued with DMind-2, which proved that domain-specific fine-tuning could outmaneuver trillion-parameter giants in vertical benchmarks.
🚀 A Paradigm Shift DMind-3 represents our most significant evolutionary leap yet. We recognized that in the high-stakes environment of DeFi, standard knowledge retrieval is a liability. A model must do more than recite facts. It must possess Reflective Intelligence (System 2 Thinking) to navigate risk.
DMind-3-mini embodies this philosophy. Positioned as the Brain within our local ecosystem, it bridges the gap between the real-time reflexes of the edge-side DMind-3-nano and the macroscopic foresight of the cloud-native DMind-3. It is engineered not as a chatbot, but as a Computational Financial Actuary designed to bring institutional-grade logic to the individual sovereign user within a privacy-first, offline-capable engine.
DMind-3-mini introduces Contrastive Chain-of-Correction Supervised Fine-Tuning (C³-SFT). Unlike standard SFT which models a direct mapping \(P(y|x)\), C³-SFT forces the model to navigate a Correction Trajectory by contrasting against plausible but flawed reasoning.
(Figure 1: The C³-SFT training pipeline, illustrating the Triplet Data Structure and Contextual Loss Masking)
The optimization objective \(\mathcal{L}_{C^3}\) is defined as maximizing the conditional probability of the correction path given the error context:
LC3(θ)=−ED[t=1∑TαDuring inference, DMind-3-mini operates in two distinct topological modes based on the presence of a trigger token \(\tau\). Let \(\mathcal{I}(x)\) denote the inference function:
y^=⎩⎨⎧yargmaxPDMind-3-mini is not just a coder. It is a Financial Risk Assessor.
For maximum security, we recommend the DMind Local Stack:
(Figure 2: The On-Device Inference Ecosystem showing the synergy between Nano and Mini)
The model was fine-tuned on 82,000 high-value private samples:
| Data Source | Proportion | Description |
|---|---|---|
| Institutional Alpha Reports | 40% | Deep dive reports from top-tier firms (e.g., Paradigm, Delphi), structured into logic chains. |
| Financial Post-Mortems | 30% | Historical analysis of collapses (Luna, FTX, Euler Hack), focusing on pre-crash indicators. |
| Smart Contract Audits | 20% | C³-SFT formatted pairs: \(\text{Vulnerable Code} \to \text{Exploit Analysis} \to \text{Fix}\). |
| On-Chain Behavior Logs | 10% | Parsed intent analysis of "Smart Money" wallet operations during high volatility events. |
Evaluated on three key benchmarks: DMind Benchmark (Web3 Native Logic), FinanceQA (Financial Domain Knowledge), and AIME 2025 (Advanced Mathematical Reasoning).
The evaluation compares DMind-3-mini (4B) against top-tier frontier models (GPT-5.1, Claude Sonnet 4.5) and other efficient models. Despite its compact size, the Mini model demonstrates exceptional efficiency, particularly in specialized domain tasks where it outperforms significantly larger generalist models.