lordx64/Qwable-v1

TEXT GENERATIONConcurrency Cost:3Model Size:35.1BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Jun 14, 2026License:agpl-3.0Architecture:Transformer0.1K Open Weights Cold

Qwable-v1 by lordx64 is a 35.1 billion parameter Mixture-of-Experts (3B active) language model with a 32768 token context length. It is a chained distillation of Qwen3.6-35B-A3B, fine-tuned first on Claude Opus 4.7 reasoning traces and then on Claude Fable-5 agentic tool-use traces. This model excels at agentic coding tasks, generating explicit chains-of-thought and emitting XML blocks for file edits, shell commands, and reads when prompted as an agent. It is designed to provide open-weights agentic capabilities for developers.

Loading preview...

Overview

Qwable-v1 is a 35.1 billion parameter Mixture-of-Experts (MoE) model by lordx64, with 3 billion active parameters and a 32768 token context length. It is built through a unique chained distillation process: starting with Qwen3.6-35B-A3B, it was first fine-tuned on Claude Opus 4.7 reasoning traces, and subsequently on Claude Fable-5 agentic tool-use traces. This dual-stage SFT (Supervised Fine-Tuning) imbues the model with distinct capabilities for both advanced reasoning and agentic coding.

Key Capabilities

  • Agentic Coding: When prompted as an agent, Qwable-v1 reliably emits <tool_use> XML blocks for actions like file reads, writes, edits, and shell commands, mimicking Claude-Code-style agent behavior.
  • Explicit Reasoning: Inherits <think>…</think> chains-of-thought from the Opus 4.7 prior, providing transparent reasoning steps.
  • Conditional Behavior: The agentic XML format is system-prompt-conditional; without an agent-style prompt, it defaults to the Opus 4.7 reasoning and explanation style, producing markdown code blocks.
  • Hardware Efficiency: Runs on a single H200 or 2x A100-80GB at bf16, and on 24+ GB consumer GPUs using IQ4_XS quantization.

Good For

  • Agentic Coding Tasks: Ideal for scenarios requiring automated code editing, testing, and codebase navigation, especially when integrated into an agent harness that supplies a tool-use system prompt and tool registry.
  • Pure Reasoning: Matches the performance of the underlying Opus 4.7 distill for general knowledge, math, and science Q&A when not explicitly prompted for agentic behavior.
  • Developers seeking open-weights alternatives for agentic workflows, leveraging the combined strengths of Qwen, Claude Opus, and Claude Fable-5's behavioral patterns.

Finetunes
3 models