New Open-Weights Models Released via Qwen Fine-Tuning

Two new open-weights large language models, Victoria and Maple, have been released, both based on the Qwen3.8-Flash-Next architecture. According to reports on Reddit, the models were developed using a Dell B300 system in a laboratory setting.

Victoria: Optimization for Coding and Agents

The Victoria model is designed specifically for coding and agentic tasks. According to the developers, the model utilizes a technique known as REAP to reduce the Qwen3.8-Flash-Next architecture by 44%, decreasing the number of experts from 512 to 288 per layer.

The model was retrained at 4-bit (NVFP4) to optimize it for its shipping format. In unverified performance benchmarks shared on Reddit, Victoria achieved a score of 70.0% on Terminal-Bench 2.1, an improvement over the 62.5% scored by a previous NVFP4 build. The model also reportedly achieved a score of 159/164 on HumanEval.

Technical specifications for the Victoria release include:

  • Weight Size: 48.0 GiB (including the draft head), with a separate 95.4 GiB n-gram table.
  • Inference Speed: Reported at 280 tokens per second (single stream) on a single B300 when using the draft head, compared to 135 tokens per second without it.
  • GGUF Version: A GGUF Q4_K_M version is available, which reportedly scored 75.3% on Terminal-Bench and 93.2% on HumanEval.

The developers noted that current mainline llama.cpp versions may encounter errors when running the GGUF version due to the inclusion of the draft head; a specific fork on GitHub is recommended for compatibility.

Maple: Regional Specialization for Canada

The second release, Maple, is a fine-tune optimized for Canadian-specific queries. The developers aim to address a common issue where models default to United States-centric information regarding taxes, benefits, and regulations.

In a test of 600 held-out questions involving search capabilities, the developers reported the following improvements for Maple compared to the base model:

  • Official Canadian Citations: Increased from 6.0% to 62.9%.
  • Fully Correct Answers: Increased from 6.6% to 21.8%.
  • “No Answer” Responses: Decreased from 47.2% to 23.7%.
  • Geographic Accuracy: The frequency of the model incorrectly pushing Canadian defaults onto users living elsewhere dropped from 2.9% to 1.0%.

The model’s coding capabilities remained consistent with the base architecture, reportedly scoring 157/164 on HumanEval according to an AI judge panel. The developers noted that human review of these results has not yet been conducted.