PNPM244/linne-fern-1-0

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 8, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

PNPM244/linne-fern-1-0 is a 27B parameter decision model built on a standard Qwen3.8 architecture, designed for classification, routing, fact-checking, and agent guardrails. It processes a given state and question with options, returning a single option and probabilities in one forward pass. This model excels at making rapid, agent-safe decisions within product and agent loops, demonstrating strong performance on the Decision Index and agent safety benchmarks across English, Vietnamese, and Arabic.

Loading preview...

Overview

PNPM244/linne-fern-1-0 is a 27B parameter decision model, built on a standard Qwen3.8 architecture, specifically engineered to make rapid, probabilistic decisions within product and agent loops. Unlike general-purpose LLMs, it takes a state (text, document, JSON, conversation, or agent trajectory) and a question with options, then returns a single chosen option along with probabilities for all options in a single forward pass. This design makes it highly efficient for its intended use cases.

Key Capabilities

  • Decision-making: Optimized for classification, routing, fact-checking, and selecting tool responses.
  • Agent Safety: Provides strong step-level guardrails for agents, catching 97.7% of unsafe steps with a low false-block rate (0.3% on held-out agent steps, 1.8% on benign prompts).
  • Multilingual: Supports knowledge, language understanding, retrieval, and classification in English, Vietnamese, and Arabic.
  • Performance: Achieves a Decision Index public score of 58.91 and is fast, with a p50 latency of approximately 46 ms on a single H200.
  • Compatibility: Uses a standard Qwen3.8 architecture and chat template, ensuring drop-in compatibility with transformers, vLLM, and SGLang.

Good for

  • Integrating decisions into products: Ideal for scenarios requiring quick, automated choices based on complex inputs.
  • Enhancing agent reliability: Provides robust guardrails to prevent agents from executing undesirable actions.
  • Classification and routing tasks: Excels at categorizing inputs and directing them appropriately.
  • Multilingual applications: Suitable for decision-making processes in English, Vietnamese, and Arabic contexts.

Finetunes
1 models