Two open-weights Qwen3.8-Flash-Next fine-tunes released: Victoria for coding/agents and Maple for Canadian context

Two new open-weights fine-tunes of Qwen3.8-Flash-Next have been released, targeting specific use cases: Victoria for coding and agents, and Maple for Canadian regulatory context.

Victoria: Coding and Agents

Victoria utilizes the REAP technique to prune 44% of experts (reducing from 512 to 288 per layer) and is retrained at 4-bit (NVFP4) precision. Key performance metrics include:

• Terminal-Bench 2.1: 70.0% (averaged over 3 runs), compared to 62.5% for the previous NVFP4 build.
• HumanEval: 159/164.
• Inference Speed: 280 tok/s single stream on a Dell B300 with the draft head, versus 135 tok/s without it.
• Size: 48.0 GiB of weights, including the draft head. A separate 95.4 GiB n-gram table is not counted in that number. A GGUF Q4_K_M version is also available at 49.17 GiB, scoring 75.3% on Terminal-Bench (single run) and 93.2% on HumanEval (averaged over 5 runs).

Maple: Canadian Context

Maple is fine-tuned to default to Canadian answers for taxes, benefits, and regulations, addressing the tendency of base models to default to US contexts. On 600 held-out questions with search enabled:

• Citing official Canadian sources: 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%.
• Coding performance: Remains strong at 157/164 on HumanEval.

Technical Notes

• llama.cpp Compatibility: The GGUF includes a draft head not yet supported by mainline llama.cpp, causing errors. Users must currently build from the rmonsurate/llama.cpp fork (branch qwen4exp-mtp).
• Evaluation Method: Maple's grading was performed by an AI judge panel; human review has not yet occurred.