AMAImedia/Qwen3.8-27B-Agens-Pilot-NOESIS-BF16

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 4, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

AMAImedia/Qwen3.8-27B-Agens-Pilot-NOESIS-BF16 is a 27 billion parameter decoder-only multimodal LLM developed by Blockway, fine-tuned from Qwen3.8-27B. It features a hybrid linear + full attention architecture and supports an extended context length of up to 1,000,000 tokens. This model is specifically optimized for Cantonese as a first-class language, providing Hong Kong usage and register, and offers more balanced, multi-perspective answers to complex historical and geopolitical questions while retaining safety guardrails. It is designed for agentic workflows and excels in coding, long-document processing, and vision-language tasks.

Loading preview...

AMAImedia/Qwen3.8-27B-Agens-Pilot-NOESIS-BF16 Overview

This model is a specialized 27 billion parameter multimodal LLM developed by Blockway (BlockWay Link Limited), built upon the robust Qwen3.8-27B base. It retains Qwen3.8's core capabilities, including strong coding performance (97.0 HumanEval), vision integration, and an impressive 1,000,000 token context length (extended with YaRN).

Key Differentiators

  • Cantonese-First Language Support: Unlike its base, Agens Pilot treats Cantonese (廣東話) as a primary language, focusing on Hong Kong usage and register for natural, contextually appropriate responses.
  • Balanced Perspectives: It provides more factual and multi-perspective answers to complex historical and geopolitical questions, achieving 85% multi-perspective responses compared to 60% for the base model, while maintaining robust safety refusals.
  • Agent-Harness Tuned: Optimized for agentic workflows, specifically for Blockway's Claway (Team-AI workforce) and Codeway (coding agent), ensuring strict instruction-following and reliable tool calls.
  • Multimodal Capabilities: Inherits Qwen3.8's vision capabilities, making it suitable for vision-language tasks.

Model Variants

Agens Pilot is available in three official builds, all sharing the same tokenizer and 1M-context configuration:

  • BF16: Reference full precision build (~51 GB).
  • FP8: Optimized for efficiency (~34 GB), matching BF16 quality.
  • INT4: Most compact (~26 GB), with slightly softer performance on borderline factual topics.

Intended Use Cases

  • Cantonese-first applications: Ideal for assistants, customer service bots, and internal tools targeting Hong Kong and Guangdong users.
  • Research and journalism: For tasks requiring factual, multi-perspective answers on complex subjects.
  • Autonomous AI agents: Suitable for multi-step operations and agentic coding workflows.
  • General Qwen3.8-27B tasks: Excels in coding, long-document analysis, and vision-language tasks.

Finetunes
1 models