README.md

metadata

library_name: mlx
pipeline_tag: text-generation
license: other
tags:
  - mlx
  - code
  - sql
  - text-generation
  - llm
  - tennda

Tennda-Nano

A lightweight code LLM independently developed by the Tennda team Focused on efficient, accurate code generation and instruction following — direct answers, no redundant reasoning.

---

Model Overview

Item	Details
Model Name	Tennda-Nano
Developer	Tennda Team (in-house)
Parameters	Lightweight (<5B)
Architecture	Self-developed Transformer decoder (multimodal input)
Weight Format	bf16, MLX native
Framework	MLX (Apple Silicon Metal GPU acceleration)
Training Data	In-house code instruction dataset (20,022 samples: SQL / Python / algorithms, etc.)
Release Date	2026-08-21

---

Highlights

• Direct answers: instruction → immediate code/answer, extremely fast, short total inference time
• High SQL accuracy: 100% correct on unseen SQL tasks
• Clean output: concise, no verbose reasoning, production-ready answers
• Deployment friendly: lightweight single-machine inference, low resource usage, suitable for edge/personal devices

---

Quality Metrics

Training Convergence (Loss)

Metric	Start	Final	Reduction
Train loss	1.929	0.963	-50%
Val loss	4.500	0.844	-81%

• Stable convergence throughout training, no signs of overfitting (val/train gap remains steady)

Evaluation Scores (3 unseen samples, temp=0.3)

#	Task Type	Result
1	SQL conditional query (age>45)	✅ Directly outputs SELECT * FROM employees WHERE age > 45;
2	Design a Car class (Python)	⚠️ Occasional loop repetition (mitigable via sampling params)
3	SQL aggregate query (MAX salary)	✅ Directly outputs SELECT MAX(SALARY) FROM EMPLOYEE;

SQL accuracy: 100% (2/2) Format alignment (direct answer, no redundant reasoning): 67% (2/3)

Note: current evaluation is a sampled manual assessment; scores are for reference only. A full evaluation pipeline is under continuous development.

---

Inference Performance (measured on Apple M4)

Metric	Result
Model load	~4s
Single-task generation	~3.4s
Output style	Direct answer (no verbose thinking)

Total latency reduced by ~83% on the same task compared with the baseline (output size reduced by 75%).

---

Training Method

• Architecture: self-developed Transformer decoder (35 layers, multimodal encoding)
• Method: in-house lightweight training pipeline (parameter-efficient training, fully local)
• Cost: 26 minutes on a single machine, peak memory 14.4GB, no large-scale compute cluster required
• Artifact: complete inference weights (MLX format)

Details in the training report and comparison report.

---

Usage (MLX)

from mlx_lm import load, generate
from mlx_lm.sample_utils import make_sampler

model, tokenizer = load("MLA299/Tennda-Nano")
sampler = make_sampler(temp=0.3, top_p=0.9)
prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "Write a SQL query to select employees older than 45."}],
    add_generation_prompt=True,
)
print(generate(model, tokenizer, prompt=prompt, max_tokens=256, sampler=sampler))

Or start an OpenAI-compatible API server:

mlx_lm.server --model MLA299/Tennda-Nano --port 8080

---

Deployment Requirements

Config	Requirement
Memory	≥ 16GB (24GB recommended)
Platform	Apple Silicon (M-series, Metal GPU)
Inference	mlx-lm ≥ 0.30
Quantization	Can be 4-bit quantized for 8GB-memory devices

---

Known Limitations

• Chinese instruction capability is still under optimization (training corpus is primarily English code instructions)
• Occasional loop repetition (can be mitigated with repetition_penalty at inference time)
• The concise answer style may omit necessary explanations

---

Tennda-Nano · In-house model · © 2026 Tennda Team