Tennda-Reason

A high-efficiency code & reasoning assistant fine-tuned by the Tennda Team Structured <think> reasoning followed by complete, production-ready answers — refined thinking, reliable delivery, built for code generation, math, and logical reasoning on Apple Silicon.

---

Model Overview

Item	Details
Model Name	Tennda-Reason
Developer	Tennda Team
Parameters	8.2B (4-bit quantized, ~4.3GB)
Architecture	Standard Transformer decoder, native <think> reasoning support
Weight Format	4-bit, MLX native
Framework	MLX 0.32.1 + mlx-lm 0.31.3 (Apple Silicon Metal acceleration)
Training	QLoRA (rank=16, scale=32, last 16 layers, 19.4M trainable params / 0.237%)
Training Data	Multi-teacher distillation SFT corpus (2,000 curated samples): math 27% · code 27% · reasoning 20% · instruction 14%
Context Length	1024 (trained), extensible via base capabilities
Release Date	2026-08-24

---

Highlights

• Refined thinking: <think> reasoning chains compressed to 300–550 token key-point style — no rambling, no wasted tokens
• Complete delivery: trained on "short thinking + complete answer" patterns; 0/5 test failures from runaway reasoning (baseline: 2/5)
• Multi-domain: balanced across math, code, logical reasoning, and instruction following
• Apple Silicon native: MLX 4-bit, ~6GB peak inference memory, runs on a single M-series machine

---

Training Details

Convergence (Loss)

Metric	Start	Best	Final
Train loss	1.70	—	0.421
Val loss	1.702	0.590 (iter 700)	0.794

• 2,000 iterations ≈ 2 epochs (batch=2, seq=1024, ~67 tokens/s, ~8h on a single M4)
• GPU peak memory 8.5GB, stable throughout, zero OOM
• Released checkpoint: iter 700 (best validation loss), selected via blind A/B output comparison against the final checkpoint

Checkpoint Selection

	iter 700 (released)	iter 2000
Val loss	0.590	0.794
Factual accuracy (networking task)	✅ correct	❌ detail error
Format compliance	✅ verified	✅

---

Evaluation (5 prompts, temp=0.3, vs pre-training baseline)

Task	Tennda-Reason	Baseline
Python quicksort	✅ complete runnable code + complexity analysis	❌ reasoning runaway, no answer produced
JS closures	✅ full structured tutorial	⚠️ thin output
SQL top salary per dept	✅ window-function solutions	❌ reasoning runaway, no answer produced
TCP 3-way handshake	✅ vivid analogy, correct steps	✅ concise & correct
Python HTTP server	⚠️ multi-approach, minor rough edges	✅ concise & correct

Summary: format compliance 5/5; runaway-reasoning failures reduced from 2/5 (baseline) to 0/5; overall usability substantially improved.

---

Usage (MLX)

from mlx_lm import load, generate

model, tokenizer = load("MLA299/Tennda-Reason")
messages = [{"role": "user", "content": "Write a quicksort in Python"}]
prompt = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
response = generate(model, tokenizer, prompt=prompt, max_tokens=2048, verbose=False)
print(response)

Command line:

mlx_lm.generate --model MLA299/Tennda-Reason \
    --prompt "Write a SQL query: highest salary per department" \
    --max-tokens 2048

Recommended sampling: temp=0.3–0.7, max_tokens ≥ 2048 (thinking chain + full answer)

---

Limitations

• Training data is English-dominant; Chinese works but is not specifically optimized
• Post-SFT the model is more confident; factual-detail hallucinations are slightly higher than baseline — verify critical details in production
• Trained at 1024 context; longer inputs rely on native capabilities
• Contains synthetic distillation content; upstream data terms apply

---

License

Apache-2.0. See the license terms for redistribution conditions.

---

Citation

@misc{tennda-reason-2026,
  title = {Tennda-Reason: A Distillation-Fine-tuned Model for Code and Reasoning on Apple Silicon},
  author = {Tennda Team},
  year = {2026},
  month = {August},
  publisher = {Hugging Face},
  url = {https://huggingface.co/MLA299/Tennda-Reason}
}

---

Tennda-Reason · © 2026 Tennda Team