Maple

Maple is a Canadian-first assistant model built on Victoria.

When an everyday question about tax, employment, benefits, immigration, health or housing does not name a country, Maple answers in Canadian terms by default. Given a web search tool, it looks things up and cites official Canadian sources. It also knows Canadian culture and answers in a polite Canadian tone, while still respecting a different jurisdiction when the user names one.

Maple is an NVFP4 model for vLLM on NVIDIA Blackwell GPUs, with 5.9B active parameters per token. SHA256SUMS lists the sha256 of every model file.

Results

Everyday questions that do not name a country. An answer passes if it is framed in Canadian terms, answers the question directly, is correct against an official source, and covers the key caveats. A question with no answer counts as a fail. Scores are the average of two AI judges.

Held-out test set (600 questions never seen in training)

Answered and passed the full standard	Victoria	Maple
With web search	6.6%	21.8%
Without search	1.3%	7.5%

With search, the gain is +15.2 points (95% CI +12.8 to +17.6). Without search, it is +6.2 points (95% CI +4.8 to +7.8).

Dev set only: comparison with Qwen3.8-Flash-Next

Dev set, without search	Victoria	Maple	Qwen3.8-Flash-Next
Everyday questions (200): answered and passed	1.7%	10.8%	8.6%
Everyday questions: no answer	19.0%	17.5%	1.0%
Canadian culture (100, multiple choice): answered correctly	59%	77%	98%

Caveats:

• All results are AI-judged; no human grading has been done yet.
• The Qwen3.8-Flash-Next comparison is on the dev set only and is not comparable with the held-out table above.
• Qwen3.8-Flash-Next is much stronger on Canadian culture and almost always gives an answer; Maple leads on answering an unmarked everyday question in Canadian terms to the full standard.

Quick start

hf download rmonsurate/Maple --local-dir maple

NVFP4 on vLLM, one NVIDIA Blackwell GPU (B300 or B200), draft head on:

vllm serve ./maple \
  --served-model-name maple \
  --tensor-parallel-size 1 \
  --max-model-len 20480 \
  --max-num-seqs 16 \
  --max-num-batched-tokens 16384 \
  --gpu-memory-utilization 0.80 \
  --enable-prefix-caching \
  --compilation-config '{"inductor_compile_config":{"triton.autotune_at_compile_time":false}}' \
  --kernel-config '{"enable_flashinfer_autotune":false}' \
  --trust-remote-code \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --reasoning-parser qwen3 \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3}'

• The --compilation-config line is required: compile-time Triton autotuning on this model either hangs or blows out startup time.
• We ran vLLM from the Qwen3.8-Flash-Next vLLM image with two fixes backported for a prefix-caching crash on this architecture (vllm#53798 and vllm#54076). On a vLLM without those fixes, leave out --enable-prefix-caching.
• Recommended sampling: thinking on, temperature 0.6, top_p 0.95, top_k 20.
• Give the model a web search tool. Its Canadian-first habit is strongest when it can look things up.

There is no GGUF build of Maple.

License

Maple is a derivative work of Qwen3.8-Flash-Next by Farpoint Technologies and carries the same license, the Qwen Community License 1.0, whose full text ships in this repo as LICENSE. Two conditions travel with it: the copyright and permission notice must be included in all copies, and a licensee running a Model as a Service or AI Work Assistant business must obtain a separate license from Qwen before any commercial use. Attribution goes to Qwen for Qwen3.8-Flash-Next and to Farpoint for Victoria and Maple.

Acknowledgements

Built on an 8x NVIDIA B300 system provided by Dell. Thanks to the Qwen team for Qwen3.8-Flash-Next and to the maintainers of vLLM.