MoziAI-27B-3.8 — A Compact Yet Powerful Multimodal AI Model for Free Local Deployment

Release Date: 2026-08-30 · Version: V3.8

📑 Table of Contents

1. Model Overview
2. Key Features — Dynamic 7-Dimensional Thinking / LOOP / MoziSmartBit / Finance Focus
3. Version Upgrade Notes
4. Core Capabilities
5. Technical Specifications
6. Quick Start
7. Model Downloads
8. Launch Commands
9. Recommended Inference Parameters
10. Quantization Format Comparison
11. MTP Speculative Decoding
12. VRAM Configuration Recommendations
13. Deployment Methods
14. Benchmarks
15. License
16. Contact

1. Model Overview

MoziAI-27B-3.8 is a local open-source multimodal AI large model developed by the team of Chinese finance influencer Chen Yumo. Built on the open-source base Qwen3.8-27B (Dense 27B architecture, Apache-2.0 license), it integrates the team's self-developed financial data + financial domain capabilities + dynamic seven-dimensional thinking framework + agent LOOP reflection and iteration mechanism + MoziSmartBit hybrid quantization algorithm. This model lowers the barrier to local deployment for individuals and enterprises, is licensed for free commercial use, can run on consumer GPUs, saves significant cloud token costs, enables 24/7 token freedom, and ensures local data privacy and security.

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	Use Case	Typical Tasks	Dimensions Expanded
Level 0	Simple Q&A	Term explanation, fact lookup, translation, summarization	①Understand task ⑤Resource needs (2-dimension quick answer)
Level 1	Analysis & Diagnosis	Market research, copywriting, data analysis, report interpretation, strategy evaluation	①②③⑤⑥ 5-dimension assessment
Level 2	Complex Dev/Strategy	Code development, architecture design, quant strategy development, multi-step workflows, system design	①②③④⑤⑥⑦ full 7-dimension deep reasoning

Seven dimensions: ①Understand task ②Complexity assessment ③Dependencies ④Risk assessment ⑤Resource needs ⑥Acceptance criteria ⑦Execution strategy

🔄 Agent LOOP Iteration Mechanism

Complex tasks automatically enter moziAI-Loop iteration mode: Round 1 execute + evaluate → Round 2 adjust + verify, ensuring output is self-validated before the final answer. The model works like a senior engineer — "decompose problem → evaluate approach → execute → reflect → optimize" — significantly improving accuracy and executability of complex tasks. For simple Q&A and tasks, Loop is automatically disabled.

📦 MoziSmartBit Intelligent Quantization

Self-developed layered intelligent quantization compresses the 27-billion-parameter Dense model to about 13.7 GB — about 3.3 GB (~20%) smaller than standard Q4_K_M (~17 GB), while maintaining FP16 ~99% precision. Traditional quantization applies uniform precision across all layers; MoziSmartBit uses an intelligent differentiated strategy tailored to Dense model architecture, achieving better precision than Q4_K_M.

💰 Financial Vertical Domain Focus

Deeply optimized for financial Q&A, quantitative programming, and tool calling. The financial domain has extremely low tolerance for model hallucination, and MoziAI significantly outperforms general models of the same size in this domain.

🌐 Other Features

Multilingual support: 201 languages and dialects, with specially optimized Chinese capability
General programming: Full-stack development, debugging, architecture design, covering Python/JS/TS/Go/Rust
Article writing: Research reports, analytical articles, technical docs, creative content and other multi-genre high-quality writing
Vision understanding: Multimodal vision, supports understanding image content from local screenshots
Multi-framework support: llama.cpp / Ollama / LM Studio / Jan
Multi-Agent support: OpenClaw / Hermes / Cursor / Claude Code / Codex, native tool calling and multi-turn task orchestration

3. Version Upgrade Notes

This upgrade mainly strengthens: moziAI's self-developed dynamic seven-dimensional thinking + LOOP iteration reasoning mode, making it smarter at recognizing task complexity, with higher task completion rates for complex tasks, and improving the "think before act" capability.

moziAI will maintain an active version upgrade iteration cadence to stay at the forefront of AI development, and continuously leverage self-developed technology to make local AI models lighter to deploy while becoming more capable.

4. Core Capabilities

Capability Domain	Description
Market Analysis	Macro/micro economic interpretation, A-share/HK/US stocks, commodities, crypto market trends and logic
Finance & Reports	Financial report key indicator interpretation, research report summarization, valuation & earnings forecast assistance
Risk & Compliance	Product risk assessment, investment advice compliance reminders, financial regulatory policy interpretation
Quant & Strategy	Quant strategy design, Pyramid (PEL) quant, backtest logic, factor construction & tool calling
Tool Calling	Access to real-time market data, databases, research report retrieval and other financial data sources

5. Technical Specifications

Item	Specification
Base Model	Qwen3.8-27B (Dense architecture, hybrid attention 16 full + 48 linear, Apache-2.0 license)
Parameters	27 billion (27B) Dense architecture
Quantization	Self-developed MoziSmartBit intelligent quantization + GGUF standard format
Context Length	256K (262,144 tokens)
Model Size	~13.7 GB
Min VRAM	16GB+ deployable (CPU offload); 20GB+ smooth long context; 32GB+ full 256K + vision
Inference Frameworks	llama.cpp / Ollama / LM Studio / Jan
Inference Speed	R9700 measured (MTP layer Q5_0): 47.3 tok/s off, 61.2 tok/s on (+29.4%) — steady-state decode at ctx 32K
Development Team	Chen Yumo Team

6. Quick Start 3 Files 100 Activate Best Inference

⚠️ Key note: MoziAI's best inference capability requires downloading 3 files simultaneously — main model, vision projector, chat template. Missing any one will lose 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/27B/, chat template under V3.8/):

moziAI-27B-MTP-V3.8-Q4_K_M-Qwen3.8-27B.gguf  ← Main model (required, 13.7 GB, repo root)
V3.8/chat-template-moziai-27B-V3.8.jinja  ← Chat template (required, includes 7D thinking + Loop instructions)
mmproj/27B/moziAI-27B-mmproj-BF16-V1.0.gguf  ← Vision projector (required, 927 MB)

File	Size	Required	Purpose
Main model `.gguf`	~13.7 GB	Required	Model weights, core inference
Vision projector `mmproj`	~927 MB	Required	Multimodal vision understanding; image capability lost without it
Chat template `.jinja`	Tiny	Required	Injects MoziAI identity + 7-dimensional thinking + LOOP instructions

6.2 Launch and Use

llama-server \
  -m ./moziAI-27B-MTP-V3.8-Q4_K_M-Qwen3.8-27B.gguf \
  --mmproj mmproj/27B/moziAI-27B-mmproj-BF16-V1.0.gguf \
  --chat-template-file V3.8/chat-template-moziai-27B-V3.8.jinja \
  -c 131072 -ngl 99 \
  --host 0.0.0.0 --port 8080

Open http://localhost:8080 in your browser to start chatting. Full recommended parameters in Section 9.

7. Model Downloads

Platform	URL
HuggingFace	chenyumo/moziAI-27B-MTP
ModelScope	chenyumo/moziAI-27B-MTP
GitHub	chenyumo166/moziAI-27B-MTP

💡 LM Studio users: search moziAI in LM Studio for one-click download, no manual file download needed.

8. Launch Commands

Minimal Launch (with Three-Piece Set)

llama-server \
  -m ./moziAI-27B-MTP-V3.8-Q4_K_M-Qwen3.8-27B.gguf \
  --mmproj mmproj/27B/moziAI-27B-mmproj-BF16-V1.0.gguf \
  --chat-template-file V3.8/chat-template-moziai-27B-V3.8.jinja \
  -c 131072 -ngl 99 \
  --host 0.0.0.0 --port 8080

Full Recommended Launch

llama-server \
  -m ./moziAI-27B-MTP-V3.8-Q4_K_M-Qwen3.8-27B.gguf \
  --mmproj mmproj/27B/moziAI-27B-mmproj-BF16-V1.0.gguf \
  --chat-template-file V3.8/chat-template-moziai-27B-V3.8.jinja \
  -c 262144 -ngl 99 -t 28 \
  --batch-size 1024 --ubatch-size 128 \
  --flash-attn auto \
  --cache-type-k q4_0 --cache-type-v q4_0 --kv-unified \
  --poll 0 \
  --reasoning auto --reasoning-budget 1024 --reasoning-format deepseek-legacy \
  --spec-type draft-mtp --spec-draft-n-max 2 --spec-draft-p-min 0.75 \
  --host 0.0.0.0 --port 8080 \
  --temp 0.6 --top-p 0.95 --top-k 20

💡 Disable MTP: remove --spec-type draft-mtp and related parameters; speed drops ~30-50% but VRAM usage is lower.

9. Recommended Inference Parameters

Based on llama.cpp official recommendations and local testing optimizations (AMD Radeon AI PRO R9700 32GB):

Parameter	General Chat	Coding/Agent	Description
temperature	0.7	1.0	Balance creativity and accuracy
top_p	0.95	0.95	Nucleus sampling threshold
top_k	20	20	Truncation sampling
repeat_penalty	1.05	1.05	Repetition penalty
context_length	131072	262144	Chat 128K / Coding 256K (llama.cpp default 128K)
reasoning	auto	auto	Enable reasoning chain (CoT)
reasoning_budget	400	400	Reasoning budget tokens
reasoning_format	deepseek-legacy	deepseek-legacy	Reasoning output to separate field
spec-type	draft-mtp	draft-mtp	MTP speculative decoding (see Section 11)

💡 Thinking mode: enabled via --reasoning auto; the model performs internal reasoning before output. reasoning_budget controls max thinking tokens (recommended 400, adjustable 100-1000).

10. Quantization Format Comparison

Format	Size	Precision	Description
FP16 original	~54 GB	100%	Lossless, needs professional GPU
MoziSmartBit (this model)	~13.7 GB	~99%	Self-developed intelligent quantization, best precision, smallest size
Q4_K_M	~17 GB	~98%	GGUF standard 4bit
Q5_K_M	~20 GB	~99%	Higher precision
Q6_K	~23 GB	~99.5%	Near lossless
Q8_0	~31 GB	~100%	Lossless

MoziSmartBit maintains ~99% precision while compressing the 27B Dense model to 13.7 GB (compression ratio 3.9x), ~20% smaller than Q4_K_M, making it better suited for consumer GPU local deployment.

11. MTP Speculative Decoding Important Speed Feature

This model has built-in MTP (Multi-Token Prediction) speculative decoding layers. This is a native feature of the Qwen3.8 architecture, and MoziAI retains the complete MTP weights.

Measured speedup (2026-11-01, R9700 / ctx 32K / greedy): off 47.3 tok/s → on 61.2 tok/s (+29.4%, acceptance 57.19%) with the MTP layer at Q5_0. The MTP layer must be Q5 or better — Q4_0/Q3_K drop acceptance to 0.00% and yield no speedup at all.

Principle: A lightweight prediction head (Draft Model) is additionally trained in the model architecture to guess subsequent tokens before main model verification, reducing forward passes and lowering inference latency. Wrong guesses are corrected by the main model with no negative impact on output quality.

Enable Parameters

--spec-type draft-mtp \
--spec-draft-n-max 2 \
--spec-draft-p-min 0.75

Parameter	Recommended Value	Description
--spec-type	draft-mtp	Enable MTP speculative decoding
--spec-draft-n-max	2	Max 2 tokens guessed per step (recommended, ~80% acceptance rate)
--spec-draft-p-min	0.75	Minimum acceptance probability threshold (0.0-1.0, higher = more conservative)

Parameter Tuning Suggestions

n-max	Acceptance Rate	Use Case
1	~90%	Most conservative, smallest speed gain
2	~80%	Recommended: balance speed and accuracy
3	~71%	General use, noticeable speed gain
4-5	~60-65%	Creative writing, code generation
6	~50-55%	Pure-text long output (adjust p-min accordingly)

12. VRAM Configuration Recommendations

VRAM	Recommended Config	Description
16 GB	Lower context to 64K, CPU offload needed	Entry level, e.g. RTX 4060 Ti
20 GB	128K full config, q4_0 KV cache	Recommended, e.g. RX 7900 XT / RTX 5070 Ti
24 GB	128K full config, ample VRAM headroom	RTX 4090 / RX 7900 XTX
32 GB+	256K full config, strongest setup	Radeon AI PRO R9700 / RTX 5090
128 GB iGPU	256K full config	AMD Ryzen AI Max+ 395 / NVIDIA RTX Spark

💡 Longer context = more VRAM usage. On OOM, gradually lower the -c parameter. Use --fit on to let llama.cpp auto-adjust layer count to fit VRAM. Supports NVIDIA / AMD / Intel GPUs.

13. Deployment Methods

Ollama Deployment

cat > Modelfile << 'EOF'
FROM ./moziAI-27B-MTP-V3.8-Q4_K_M-Qwen3.8-27B.gguf
PARAMETER temperature 0.6
PARAMETER top_p 0.95
PARAMETER top_k 20
PARAMETER num_ctx 131072
PARAMETER num_gpu 99
EOF

ollama create moziAI-27B -f Modelfile
ollama run moziAI-27B

LM Studio / Jan

Search moziAI in LM Studio / Jan and download the Q4_K_M quantized version.

💡 Ollama's mmproj and chat_template support is limited; it is recommended to use llama.cpp for full functionality.

14. Benchmarks

MoziAI-27B-3.8 is fine-tuned on the Qwen3.8-27B base, with the financial vertical domain as the core optimization direction.

Coding

Benchmark	moziAI-27B	Qwen3.6-27B	Qwen3.7-Plus	Opus4.6 Max
Terminal Bench 2.1	73.0	63.4	64.0	78.2
SWE-bench Pro	61.7	53.5	57.6	53.4
DeepSWE 1.1	42.2	13.3	14.2	--
QwenSWEBench	79.0	49.3	59.2	63.8

Agent Capabilities

Benchmark	moziAI-27B	Qwen3.6-27B	Qwen3.7-Plus	Opus4.6 Max
CoWorkBench	70.7	61.0	65.1	68.2
JobBench	33.4	21.8	27.6	--
Agents' Last Exam	42.9	27.3	33.6	--
WebArena-Verified	64.8	48.8	55.3	--
AndroidWorld	81.9	70.3	81.0	62.0

General Capabilities

Benchmark	moziAI-27B	Qwen3.6-27B	Qwen3.7-Plus	Opus4.6 Max
IFBench	79.5	69.1	79.1	62.5
GPQA Diamond	89.2	87.8	90.3	91.3

Multimodal Capabilities

Benchmark	moziAI-27B	Qwen3.6-27B	Qwen3.7-Plus	Opus4.6 Max
MathVision	94.6	85.1	90.3	65.5
BabyVision	85.6	28.9	70.4	12.6
CharXiv RQ	90.2	78.4	85.8	66.0
Vision2Web	62.9	45.0	42.1	--

Competitor data from official public benchmark results. MoziAI significantly outperforms general models in financial vertical domains (financial report interpretation, quant strategy, risk & compliance, Agent tool calling, etc.).

15. License

License: Apache License 2.0 (LICENSE.md; Chinese note LICENSE.zh-CN.md). Distributed as a whole under Apache-2.0 (upstream Qwen3.8-27B likewise Apache-2.0; MoziAI additions likewise Apache-2.0). Attribution: NOTICE; modifications: MODIFICATIONS.md.

16. Contact

HuggingFace: @chenyumo
GitHub: @chenyumo166
Weibo: @rimochen
E-mail: 263515@qq.com

License

License: Apache License 2.0 — authoritative text: LICENSE.md (Chinese note: LICENSE.zh-CN.md). This model is distributed as a whole under Apache-2.0; upstream components (Qwen3.8-27B) are Apache-2.0, and MoziAI's own additions are likewise Apache-2.0. Attribution: NOTICE; modifications: MODIFICATIONS.md.