CMKL/MANGO1.5-Qwen3.5-9B

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 19, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

CMKL/MANGO1.5-Qwen3.5-9B is a 9 billion parameter, text-only, bilingual (Thai/English) instruction-tuned model developed by CMKL University. Fine-tuned from Qwen/Qwen3.5-9B using LoRA on a 5 million sample corpus, it excels in Thai and English chat, reasoning, summarization, and translation. This model demonstrates superior performance on the ThaiLLM Leaderboard compared to its base and other 12B Thai LLMs, particularly in exam and NLU tasks.

Loading preview...

Model Overview

CMKL/MANGO1.5-Qwen3.5-9B is a 9 billion parameter, text-only, bilingual (Thai/English) instruction-tuned model developed by CMKL University. It is fine-tuned from the Qwen/Qwen3.5-9B base model using LoRA on a substantial 5 million sample English/Thai instruction corpus. This release focuses on providing a smaller, dense model with broad bilingual capabilities, superseding previous larger, omni-modal efforts.

Key Capabilities & Features

  • Bilingual Proficiency: Optimized for fluent interaction in both Thai and English, handling mixed-language inputs.
  • Instruction Following: Excels in chat, instruction-following, reasoning, summarization, translation, and Q&A across various domains.
  • Strong Performance: Achieves a normalized average of 0.444 on the ThaiLLM Leaderboard, outperforming its base model (0.305) and the 12B Typhoon2.1-Gemma3-12B (0.349).
  • Reasoning Enhancement: Demonstrates significant gains in exam and NLU tasks when reasoning mode is enabled, improving efficiency over the base model.
  • Comprehensive Training Data: Trained on 5 million samples (65% English, 35% Thai) covering chat, math, code, STEM, knowledge, creative writing, and more.

Intended Use Cases

  • Bilingual Chatbots: Ideal for applications requiring seamless conversation in both Thai and English.
  • Instruction Following: Suitable for tasks like summarization, translation, and complex Q&A in either language.
  • Research & Development: A valuable artifact for further research in bilingual LLMs, particularly for Thai language applications.

Limitations

This is an SFT-only checkpoint, meaning it has not yet undergone dedicated preference-alignment (DPO/RLHF), which may impact response style and refusal calibration. Evaluation is currently limited to its base model and one other Thai LLM, without a broader sweep or LLM-as-judge assessment.

Finetunes
3 models