Benchmarked M2.7, M3 and M2 against the same four pt-BR prompts (a
Minecraft question, a general-knowledge one, an unanswerable one, and a
jailbreak attempt).
M2.7 wins on the axis that matters for chat. All three held the
no-server-access line under the jailbreak, but:
- M3 is a reasoning model, and its thinking counts against max_tokens.
At the 300 chat needs, its Minecraft answer was cut off mid-word at
39 characters while reporting 590 tokens used. Raising the budget
enough to fix that costs more per question than the answer is worth.
- M2 answered the brewing question wrong, going straight from a water
bottle to magma cream and skipping the nether wart.
- M2.7 answered correctly in 2.7-5.2s.
The benchmark also showed both markdown and emoji in every model's
output. Minecraft chat renders neither: `**negrito**` arrives as
literal asterisks, and emoji are empty boxes on Bedrock. The system
prompt now asks for plain text, and sanitise() strips emphasis, code
fences, headings and emoji regardless, since a prompt is a request and
not a guarantee.
marcos
0b3be45d76