lordx64/Qwable-v2

TEXT GENERATIONConcurrency Cost:3Model Size:35.1BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Jun 24, 2026License:agpl-3.0Architecture:Transformer0.0K Open Weights Cold

Qwable-v2 by lordx64 is a 35.1 billion parameter Mixture-of-Experts (3B active) agentic coding model with a 32768 token context length. It is built by layering Claude Fable-5 agentic tool-use behavior on top of a Claude Opus 4.7 reasoning distill of Qwen3.6-35B-A3B. This iteration features 4x the LoRA capacity and 2x the SFT data of its predecessor, specifically targeting improved agentic tool-use fidelity and reducing early-stop failures in coding tasks. It excels at agentic coding with real Claude Code tool names and field signatures, while also providing strong pure reasoning capabilities.

Loading preview...

Qwable-v2: Agentic Coding with Enhanced Tool-Use

Qwable-v2 is the second iteration of the Qwable lineage, a 35.1 billion parameter Mixture-of-Experts (MoE) model (3B active) with a 32768 token context length. Developed by lordx64, it's engineered for agentic coding, combining Claude Fable-5's tool-use behavior with the reasoning capabilities of a Claude Opus 4.7 distill of Qwen3.6-35B-A3B. This version significantly improves upon v1 by incorporating 4x the LoRA capacity and 2x the SFT data, specifically addressing early-stop failures during tool use.

Key Capabilities

  • Agentic Tool-Use: Emits <tool_use> XML with real Claude Code tool names (Read, Edit, Bash) and correct field signatures, a major improvement over v1's invented variants. This behavior is reliably triggered with an agent-style system prompt.
  • Chains-of-Thought: Inherits explicit <think>…</think> reasoning chains from the Opus 4.7 prior.
  • Enhanced Training: Trained with a larger SFT dataset (9,842 rows) and increased LoRA rank (64), improving the model's ability to learn complex patterns and complete tool-use blocks.
  • Increased Context Length: Supports an 8,192 token sequence length during training, allowing full agentic conversations to fit without clipping.

Good For

  • Agentic Coding Tasks: Ideal for scenarios requiring file edits, shell commands, and code reading, especially when supplied with an agent system prompt.
  • Pure Reasoning: When used without an agent system prompt, it leverages its Opus 4.7 distill for strong performance in math, science, and general Q&A.
  • Developers seeking Claude-like agentic behavior: Offers a structurally cleaner and more faithful implementation of Claude Code's tool surface compared to its predecessor.

Finetunes
2 models