tokenizer.chat_template of Ternary-Bonsai-2-27B-PTQ1_0.gguf (8,952 characters) contains, when thinking is enabled (enable_thinking undefined or true):
{%- set resolved_reasoning_effort = reasoning_effort|default('xhigh') %}
{%- if resolved_reasoning_effort not in ('xhigh', 'medium', 'low') %} ... raise_exception ...
{%- if resolved_reasoning_effort == 'xhigh' %}
{%- set reasoning_instructions = 'Reasoning effort is set to xhigh. Please think carefully through the task, validate key assumptions, consider plausible alternatives, and prioritize correctness, consistency, and clarity in the final answer.' %}
{%- elif resolved_reasoning_effort == 'low' %}
{%- set reasoning_instructions = 'Reasoning effort is set to low. Keep your thinking brief and focused, moving directly to the conclusion without unnecessary elaboration.' %}medium sets no instruction. The server's /apply-template confirms it : the default prompt for a one-line user message starts with <|im_start|>system\nReasoning effort is set to xhigh. Please think carefully ...<|im_end|>, the medium prompt has no system message and still opens <think>, enable_thinking: false opens and closes <think> at once.
--reasoning-effort LEVEL server flag (env LLAMA_ARG_REASONING_EFFORT), default keeps the template default. Present on the prebuilt prism-b10685 llama-server and on the bonsai2 bundle binary (both --help outputs checked).--chat-template-kwargs '{"reasoning_effort":"medium"}' server flag."reasoning_effort": "medium" on /v1/chat/completions (none disables reasoning), or "chat_template_kwargs": {"reasoning_effort": "medium"}; the Responses-style reasoning: {effort} is mapped to the same field.--chat-template-file with an edited template.Related: --reasoning-budget N caps thinking tokens server-side, --reasoning-budget-message injects a line before the forced close. Neither was measured here.
Original file, the bonsai2 bundle build (8971d7b) llama-server, -ngl 99 -fa on -c 65536 -np 1 -ctk q4_0 -ctv q4_0 --jinja, greedy (temperature 0, top_k 1), /v1/chat/completions, two runs per cell, two client caps (max_tokens 4096, the common client default, and 16384). Thinking tokens = server reasoning_content counted with /tokenize; content tokens likewise; generated = usage.completion_tokens. Complete = SVG contains <svg and </svg>; HTML contains <html and </html>; Python parses with ast and has at least 100 non-empty lines. Runs 1 and 2 were identical in every cell (greedy).
Tasks: SVG = the prompt from professorpalmer's bench/reason_ab.py with the subject he lists first, "Draw a five-tier pagoda as a single self-contained SVG. No markdown, no explanation, inline SVG only. Temperature-0 style: clean geometric shapes, white background."; HTML = a single self-contained to-do list page with localStorage; Python = a key-value store CLI of at least 100 lines with a self-test.
| cap | effort | task | finish | generated tokens | thinking tokens | content tokens | complete | wall s (run 1, run 2) |
|---|---|---|---|---|---|---|---|---|
| 4096 | default (xhigh) | svg_pagoda | length | 4096 | 4096 | 0 | no, empty | 107.6, 107.6 |
| 4096 | default (xhigh) | html_todo | length | 4096 | 4096 | 0 | no, empty | 107.7, 107.6 |
| 4096 | default (xhigh) | python_kv | length | 4096 | 4096 | 0 | no, empty | 107.7, 107.6 |
| 4096 | medium | svg_pagoda | stop | 2238 | 1610 | 625 | yes | 57.7, 57.7 |
| 4096 | medium | html_todo | stop | 1758 | 22 | 1733 | yes | 45.2, 45.2 |
| 4096 | medium | python_kv | length | 4096 | 2381 | 1713 | no, cut inside the code | 107.6, 107.6 |
| 16384 | default (xhigh) | svg_pagoda | length | 16384 | 16384 | 0 | no, empty | 477.2, 476.9 |
| 16384 | default (xhigh) | html_todo | stop | 12988 | 7839 | 5145 | yes | 368.0, 367.7 |
| 16384 | default (xhigh) | python_kv | stop | 11936 | 9536 | 2397 | yes | 335.0, 335.2 |
| 16384 | medium | svg_pagoda | stop | 2238 | 1610 | 625 | yes | 57.7, 57.5 |
| 16384 | medium | html_todo | stop | 1758 | 22 | 1733 | yes | 45.2, 45.0 |
| 16384 | medium | python_kv | stop | 4679 | 2381 | 2295 | yes | 123.5, 123.4 |
<think>). medium completes the SVG and the page inside the cap; the 100-line script needs about 4,700 tokens and is cut.medium completes all three in 45 to 124 seconds with 22 to 2,381 thinking tokens; on the two tasks both settings complete, medium uses 7.4x and 2.6x fewer tokens.Decision: medium measurably helps. The serve scripts in serve/ and in the prebuilt bundle get --reasoning-effort medium; the README says what it does and what it costs. The tarball already on Hugging Face still carries the old scripts until it is re-uploaded. --reasoning-budget was not measured and is not set.