MoziAI-35B-V3.8 is a locally deployable open-source multimodal AI large model developed by the team of Chen Yumo, a leading Chinese finance influencer. Built on the open-source base Ornith-1.5-35B-A3B (Qwen3.5-35B-A3B / Qwen3.6-35B-A3B architecture, MoE 35B, MIT license), it integrates the team's self-developed financial data + financial domain capabilities + dynamic seven-dimensional thinking framework + agent LOOP reflection and iteration mechanism + Uncensored characteristic + MoziSmartBit hybrid quantization algorithm.
💡 Size Advantage: only 15.9 GB — the 35B-parameter MoE model is compressed to just 15.9 GB via the self-developed MoziSmartBit quantization (about 30% smaller than standard Q4_K_M ~22GB). It fits in a single installer, runs on ordinary consumer GPUs (20GB VRAM+), reduces cloud token costs to zero, enables 7×24 hour token freedom, and ensures local data privacy and security. Licensed for free commercial use — zero barrier for individuals and enterprises.
2. Key Features
🧠 Dynamic Seven-Dimensional Thinking Framework
MoziAI's self-developed core reasoning framework. For any task, the model first outputs a moziAI-Think marker, then dynamically unfolds structured thinking based on task complexity:
Level
Scenario
Typical Tasks
Dimensions Expanded
Level 0
Simple Q&A
Term explanation, fact lookup, translation, summarization
Complex tasks automatically enter moziAI-Loop iteration mode: Round 1 execute+assess → Round 2 adjust+verify, ensuring output undergoes self-validation before the final answer. The model works like a senior engineer — "decompose problem → evaluate plan → execute → reflect → optimize" — significantly improving accuracy and executability of complex tasks. Simple Q&A automatically skips the Loop.
📦 MoziSmartBit Smart Quantization
Self-developed layered smart quantization compresses the 35B-parameter MoE model to about 15.9 GB — about 6.5 GB (~30%) smaller than standard Q4_K_M (~22 GB) while maintaining ~99% of FP16 accuracy. Traditional quantization applies uniform precision to all layers; MoziSmartBit adopts a smart differentiated strategy for MoE structure, delivering accuracy better than Q4_K_M. Compression ratio: 4.5x.
💰 Financial Vertical Focus
Deeply optimized for financial Q&A, quantitative programming, and tool calling. Finance has extremely low tolerance for hallucination — MoziAI significantly outperforms general models of the same size in this domain.
🛡️ Uncensored Feature
No content moderation restrictions, free output, complete information, local privacy. Suitable for academic research, deep analysis, free discussion and more. (See Section 15)
🌐 Other Features
Multilingual support: 201 languages and dialects, with specially optimized Chinese
General programming: Full-stack development, debugging, architecture design, covering Python/JS/TS/Go/Rust
Multi-framework support: llama.cpp / Ollama / LM Studio / Jan
Multi-Agent support: OpenClaw / Hermes / Cursor / Claude Code / Codex etc., native tool calling and multi-turn task orchestration
3. Version Upgrade Notes
V3.8 is retrained on the same-generation self-developed training dataset system as 27B-V3.8 (identity / dynamic seven-dimensional thinking / LOOP iteration / financial vertical domain), with key enhancements to the dynamic seven-dimensional thinking + LOOP reasoning mode — smarter complexity recognition, higher complex-task completion rates, stronger "think before act" ability. The Uncensored characteristic and financial vertical optimization are carried forward.
MoziAI maintains an active upgrade cadence, staying at the forefront of AI development while continuously making local AI models lighter and more capable through self-developed technology.
Pluggable into live market data, databases, research retrieval and other financial data sources
5. Technical Specifications
Item
Specification
Base Model
Ornith-1.5-35B-A3B (Qwen3.5-35B-A3B / Qwen3.6-35B-A3B architecture, MIT license)
Parameter Count
35B MoE architecture, 256 routing experts + 1 shared expert, 8 experts active per token
Quantization
Self-developed MoziSmartBit smart quantization + GGUF standard format
Context Length
256K (262,144 tokens)
Model Size
~15.9 GB
Minimum VRAM
20GB+ deployable (CPU offload); 24GB+ smooth long context; 32GB+ full 256K + vision
Inference Frameworks
llama.cpp / Ollama / LM Studio / Jan
Inference Speed
With speculative decoding: 140+ tok/s on AMD R9700 GPU / 70+ tok/s on AMD MAX+395 iGPU — token freedom locally
Developer
Chen Yumo Team
6. Quick Start 3 Files 100 Best Inference
⚠️ Key tip: Best inference requires downloading 3 files at once — main model, vision projector, chat template. Missing any one loses the corresponding capability.
6.1 Download Model Files
Download these 3 files from HuggingFace / ModelScope to one local folder (main model at repo root, vision projector under mmproj/35B/, chat template under V3.8/):
💡 LM Studio users: search moziAI in LM Studio to download with one click.
💡 Download tip: go to the HuggingFace repo above, open the "Files and versions" tab, download the main model at the repo root, then the vision projector from mmproj/35B/ and the chat template from V3.8/, keeping all three in the same folder.
Reasoning budget tokens (higher for complex tasks)
reasoning_format
deepseek-legacy
deepseek-legacy
Reasoning in separate field
spec-type
default
default
Speculative decoding (ngram, MoE-optimal, see Section 11)
KV cache
q4_0
q4_0
Quantized KV cache (unified kv-unified)
💡 Thinking mode: enabled via --reasoning on — the model reasons internally before answering. reasoning_budget caps the max thinking tokens.
10. Quantization Format Comparison
Format
Size
Accuracy
Notes
FP16 original
~70 GB
100%
Lossless, needs pro GPU
MoziSmartBit (this model)
~15.9 GB
~99%
Self-developed smart quantization, best accuracy per size
Q4_K_M
~22 GB
~98%
Standard GGUF 4-bit
Q5_K_M
~24.7 GB
~99%
Higher accuracy
Q6_K
~28.5 GB
~99.5%
Near-lossless
Q8_0
~36.9 GB
~100%
Lossless
MoziSmartBit keeps ~99% accuracy while compressing the 35B MoE model to 15.9 GB (4.5x compression), ~30% smaller than Q4_K_M — ideal for consumer GPUs.
11. Speculative Decoding Acceleration Key Feature
This model significantly boosts inference speed via Speculative Decoding — locally measured ~1.5-2x faster than disabled.
MoE-optimal config: llama.cpp recommends ngram speculative decoding (--spec-default) for MoE architectures — fastest and most stable in local testing
About the "MTP" in the name: "MTP" refers to the base model's Multi-Token Prediction weights (fully preserved); llama.cpp's MTP draft support for MoE is limited, so MoziAI uses the ngram scheme for the best measured speed
Enable Parameter
--spec-default
Tuning Suggestions
Config
Scenario
--spec-default (default)
Recommended: balanced speed & VRAM
Disable (remove the flag)
Low-VRAM scenarios; slightly slower
12. VRAM Configuration Recommendations
Measured with the MoziSmartBit build (model + vision ~16.4GB total):
Search moziAI in LM Studio / Jan and download the Q4_K_M quantized version (LM Studio reads root-dir models by default; for legacy versions use "Add from URL" to import files from their version directory, e.g. V3.7/).
💡 Ollama's support for mmproj and chat_template is limited — prefer llama.cpp for the full feature set.
14. Benchmarks
MoziAI-35B-V3.8 is fine-tuned, distilled and further developed from the deepreinforce-ai/Ornith-1.5-35B-A3B base, with financial vertical as the core optimization direction. Multi-model comparison (MoziAI general capabilities match base Ornith-1.5-35B-A3B; data carried from V3.7 measurements — V3.8 shares the same base and training system):
Benchmark
moziAI-35B-V3.8 (This model)
Ornith-1.0-35B-A3B
Qwen3.6-35B-A3B
Gemma-4-31B
Muse-Glimmer-30B
Qwen3.5-397B
Coding
Terminal-Bench 2.1 (Terminus-2)
67.8
64.2
52.5
42.1
51.7
53.5
Terminal-Bench 2.1 (Claude Code)
68.5
62.8
49.2
-
-
48.6
SWE-bench Verified
79
75.6
73.4
52
76
76.4
SWE-bench Pro
59.6
50.4
49.5
35.7
51.2
51.6
SWE-bench Multilingual
71.4
69.3
67.2
51.7
-
69.3
DeepSWE
22
0
0
-
-
1
Frontier-Bench v0.1
5.1
1.4
1.4
-
-
1.4
NL2Repo
46.2
34.6
29.4
15.5
-
36.8
SWE Atlas - QnA
39.8
37.1
15.5
-
-
20.4
Reasoning
HLE (no tools)
25.6
20.8
21.4
19.5
22
28.7
HLE (with tools)
33.4
30.1
28.9
26.5
-
48.3
GPQA Diamond
89.2
86.2
86
84.3
83.5
88.4
Agentic
MCP-Atlas
70.2
64.4
62.8
55
75.5
72.3
Toolathlon-Verified
48.7
42.4
41.7
40.8
-
38.3
WideSearch
67.8
63.4
60.1
54.2
-
74
BrowseComp
67.6
63.5
62
-
-
78.6
ClawEval
72.5
69.8
68.7
48.5
-
70.7
MoziAI's financial vertical — earnings interpretation, quant strategy, risk & compliance, agent tool calling — significantly outperforms general models. Gemma-4 / Qwen3.6 numbers are official public results.
15. Uncensored Optimization
This model inherits the Uncensored characteristic of the Ornith-1.5-35B-A3B base:
Advantage
Description
No restrictions
Refuses no topics, including sensitive or controversial content
Free output
Not constrained by safety policies, generates any type of reply
Complete information
Provides unfiltered information, ideal for research and analysis
Local privacy
Local deployment means fully private data, no cloud review
Use cases: academic research, deep analysis, free discussion, unrestricted AI conversation.
Note: This is a locally deployed model — output is entirely controlled by the user; the model bears no content-moderation responsibility.
16. License
This model uses a custom restrictive license:
✅ Allowed — free commercial use, copying and distribution
❌ Forbidden — further development, resale, sub-licensing
📋 Required — retain original copyright notice, credit: moziAI-35B
The model is provided "as is" without warranties of any kind. Model output is for reference only and does not constitute investment advice. Users assume all risk.
