peculiar-ragdoll/Qwen-Sharp-Chat-Templates

Qwen Sharp Chat Templates

This is a drop-in fix for any Qwen3.5, 3.6, or 3.8 model, optimizing the models for knowledge work and coding.

With the Sharp template, Qwen3.8-27b (medium effort) gets smarter and uses fewer thinking tokens before it answers, and the effect is comparable for other compatible models.

Sharp makes Qwen's models more intelligent per token, and makes them communicate more information per token by cutting filler without sacrificing correctness or substance, saving time and effort for both the model and the user in multi-turn conversations.

On real repository work the same effect shows up as speed: Sharp fixes about as many issues as the stock template, while reaching each one 2.5× faster on the median problem.

Straight to the point

This is froggeric's Qwen-Fixed-Chat-Templates v22.3 with a force-appended system prompt spliced in, plus this repo's v22.3.1 bugfix on top. The base fixes issues, the addition makes it better, and v22.3.1 makes the fast (thinking-off) path coherent — see the changelog below.

v22.3.1 (current — this repo's bugfix, rebased onto froggeric v22.3). Three changes, all scoped to thinking-off (fast) mode; the thinking-on path is byte-identical to upstream v22.3 plus the terseness block, so anything already running with thinking on gets only upstream's fixes.

1. Fast-mode <think> contradiction fixed. With tools and thinking off, upstream still tells the model to put its reasoning inside <think></think> — while the generation prompt has already closed thinking. This gates every <think>-related tool-call instruction on thinking being on, so fast mode no longer asks for a block it can't open. (Tool-call format rules are untouched.)
2. Terseness lead split by mode. The single lead — "Answer directly, after thinking" — is incoherent when thinking is off, so thinking-off gets a terse-neutral "Answer directly and concisely" lead instead. The terseness core (the never/always rules) is identical in both.
3. The last thinking reference on the fast path removed. One tool-call rule still read "output the <tool_call> block IMMEDIATELY after thinking" in fast mode — the same contradiction as (1), surviving in a line the first fix didn't cover. Only the two-word fragment is gated, so thinking-off reads "…IMMEDIATELY, with NO conversational text before it" and thinking-on is unchanged to the byte. With this, a fast-mode prompt contains no reference to thinking at all.

(1) and (2) carry over unchanged from v22.1.1; (3) is new in v22.3.1. None is fixed upstream as of v22.3.

Version-string break. v22.3.1 does not contain v22.1 as a substring, so any checker matching on the old id stops matching. This project's re-embed scripts (publish/retemplate_dirk.py, publish/retemplate_dirk_v3.py) now read the expected version out of the template file at run time instead of hardcoding it, so the next rebase won't break them again. Already-published GGUF/MLX builds still carry v22.1.1 and are unaffected until deliberately re-templated.

What upstream added in v22.2 / v22.3 (and what it closed for us). Rebasing picked up, verified by upstream's own suite:

• Long tool errors escalate again. The old content|length < 500 gate meant a multi-line traceback — the exact case where escalation matters — silently never escalated. v22.2+ replaces it with two tiers: structural signals ("error":, "status": "error", a nonzero exit code, a real Traceback (most recent call last):) escalate at any payload size, while weak signals stay size-gated. This is one of the two issues this repo previously listed as open and deferred — it is now fixed, upstream, and better than the patch that regressed here.
• False retry loops on code search killed. Grep hits containing throw new Error(...), console.error, logger.error no longer count as tool failures.
• Multiple leading system/developer messages merge into one system turn joined by blank lines, instead of only the first being treated as a system prompt.
• In-content reasoning extraction widened to content that starts with <think>/<thinking>, and de-duplicated when reasoning_content/thinking is supplied alongside inline tags.
• Preserved assistant turns always render the <think> wrapper, even when the thought was empty, so rendered history matches what was actually generated and the prefix cache stays valid.
• Tool arguments serialize correctly. Booleans, nulls and numbers now go through tojson instead of | string (which emitted Python True/None); raw string args honour max_tool_arg_chars; JSON tool format no longer truncates tool responses.
• More effort aliases: off, max, ultracode, extreme, with matching <|think_…|> tags.

Still open (so you don't over-trust it): a literal <|think_off|> arriving in tool output still disables reasoning if your harness packs tool results into a user message — upstream's tag scanner reads system/developer/user roles, so a proper tool-role message is safe, and that is unchanged in v22.3. And a separately-reported mid-answer <think> tag remains unreproduced in our stack (2,863 corpus messages + 14 fresh llama.cpp generations, zero repros), with MTP speculative decoding the leading suspect; v22.3.1 does not target it.

v22.3 (upstream base). Covers Qwen 3.8 alongside 3.5/3.6 and adds prompt-directed reasoning-effort steering (none/minimal/low/medium/high/xhigh, plus the aliases above) and inline <|think_…|> control tags. The default effort is medium — a neutral baseline that injects no steering line when the caller asks for nothing. (Earlier v22 forced xhigh by default; froggeric fixed that upstream, so this Sharp build no longer suppresses anything — out of the box you get the tuned terseness behavior and nothing else, exactly as v1.) An explicit effort still renders; pass it via chat_template_kwargs (a bare top-level reasoning_effort field is dropped by OpenAI-style servers before the template sees it):

{"messages": [...], "chat_template_kwargs": {"reasoning_effort": "low"}}

Dagger-Qwen3.6-27B and Nail-Qwen3.6-35B-A3B shipped with the v1 template baked into those builds — that is the exact template embedded in those GGUF and MLX builds (template_version = "qwen3.6-froggeric-v21.3", terseness, no reasoning-effort steering), and it lives here in archive/v1-qwen3.6-froggeric-v21.3/. The chat_template.jinja at the root of this repo is the newest v22.3.1 described above; drop it in to move a model onto it. The template is published separately because it is the portable part — the thing worth reusing is not tied to either model.

Superseded versions are kept verbatim under archive/: the v1 froggeric-v21.3 build (Dagger/Nail), the v22.1 build in archive/v22.1-sharp/, and the immediately-prior v22.1.1 build in archive/v22.1.1-sharp/ — the last one before the v22.3 rebase, and the version embedded in the published Dirk builds.

What it changes

A terseness block, force-appended after your own system prompt. The lead now varies by thinking mode (the v22.3.1 fix); the core never/always rules are identical on both paths, and the thinking-on path is byte-identical to upstream v22.3 + the original terseness block.

{%- if ns_state.thinking %}
    {%- set _terse_lead = 'Answer directly, after thinking. Lead with the answer, then only what it needs to be correct and usable.' %}
{%- else %}
    {%- set _terse_lead = 'Answer directly and concisely. Give the answer with only what it needs to be correct and usable.' %}
{%- endif %}
{%- set _terse_core %}
Never: open with preamble or pleasantries; restate the question; add filler transitions; hedge with niceties; or repeat a point you've already made.
Always: keep essential steps, caveats, uncertainties, and specifics — never drop correctness or a needed warning for brevity. Keep the final answer lean. Use the least structure that conveys it (plain prose when short; lists or code only when they earn their place). If genuinely uncertain, say so and explain why — never omit uncertainty for the sake of brevity.
If a user request is genuinely ambiguous, ask a sharp question, don't guess.
{%- endset %}
{%- set _terse = _terse_lead ~ '\n' ~ (_terse_core | trim) %}
{%- if not _sc %}
    {%- set _sc = _terse | trim %}
{%- else %}
    {%- set _sc = (_sc | trim) ~ '\n\n' ~ (_terse | trim) %}
{%- endif %}

Two things happen here: the if/else on _sc keeps your own system prompt — the terseness block is appended after it, nothing you pass in is replaced; and the lead line matches the reasoning mode so a fast-mode model isn't told to "answer after thinking" when it isn't thinking. Separately, the tool-calling instructions gate every <think> reference — and the phrase "IMMEDIATELY after thinking" — on thinking being on (the other half of the v22.3.1 fix). No effort-suppression is needed: v22.3 already defaults to medium, which injects no reasoning-effort line unless you ask for one (earlier v22 forced xhigh; see the v22.3 note above). An explicit reasoning_effort still renders.

Impact

The terseness instruction targets prose padding: preamble, restating the question, filler transitions. Where the deliverable is mostly code or a structured artifact there is less paddingto remove, so expect less from it — and the prompt deliberately protects those ("lists or code only when they earn their place", "never drop correctness for brevity").

In addition to reducing thinking tokens while retaining or increasing accuracy, like shown in the graphs up top, the template avoids amnesia and loops by turning on thinking retention: with this template, the model remembers what it thought last turn by default, instead of discarding it. This also increases time to first token on subsequent turns by guaranteeing a cache hit, instead of invalidating the cache by ripping out previous thinking blocks.

Use

MLX / transformers — drop chat_template.jinja into the model directory.

hf download peculiar-ragdoll/Qwen-Sharp-Chat-Templates chat_template.jinja \
  --local-dir /path/to/your-model

Two places can hold a template, and old runtimes disagree about which wins. A model directory can carry it as chat_template.jinja and as a chat_template key inside tokenizer_config.json. Anything on transformers ≥ 4.51 — which includes current oMLX and LM Studio — prefers the .jinja file, so the drop-in just works. Older runtimes read only the embedded key and ignore the file, and then the drop-in silently does nothing.

If the directory has both and you are unsure of your runtime, patch both — that is what chat_template_oneline.txt is for: paste it as the chat_template value. Or run scripts/check_applied.py (below), which reports every source and flags a mismatch.

oMLX — drop chat_template.jinja into the model directory and rescan. Verified on oMLX (transformers 5.12.1) by loading a model with the Sharp template as chat_template.jinja and a deliberately different template embedded in tokenizer_config.json: the .jinja file won, and the model reported the terseness rules with the caller's own system prompt still in force.

GGUF — rewrite the embedded template without requantizing:

pip install gguf
gguf-new-metadata \
  --chat-template-file chat_template.jinja \
  input.gguf output.gguf

tokenizer_config.json — use chat_template_oneline.txt, the minified single-line form. It renders identically to the full template (verified by scripts/verify_template.py).

llama.cpp at runtime, without touching the file — pass it per-run instead:

llama-server -m model.gguf --chat-template-file chat_template.jinja --reasoning-format deepseek -ngl 99
llama-cli    -m model.gguf --chat-template-file chat_template.jinja -ngl 99

Same effect, and it fully replaces whatever is embedded in the GGUF — verified against a build whose embedded template names a specific model: with the flag, the served template is byte-identical to this file and the model name is gone. Check it yourself with curl localhost:8080/props | jq -r .chat_template, or render a prompt through POST /apply-template.

Three caveats. --jinja is enabled by default in current llama.cpp, so you usually do not need it — on older builds you do, and it must come before --chat-template-file. And the flag is per-invocation: forget it once and you silently get the embedded template back. Rewriting the GGUF with gguf-new-metadata is the durable version; the flag is right for trying it out or for running one template across several models.

Third, --reasoning-format deepseek (shown on the server line; it is an API-response setting, so it does nothing for llama-cli). It puts the model's <think> block in the OpenAI reasoning_content field instead of leaving it inline in content — which is what keeps a coding agent from stalling on raw thinking tokens mid-stream. On current llama.cpp it is already a no-op: --reasoning-format defaults to auto, which the source defines as "same as deepseek" — verified at build 9890 (74976e1ae), where COMMON_REASONING_FORMAT_AUTO appears in no behavioural branch at all and every extraction site gates on != none. Pass it anyway if you may be on an older build. The setting that genuinely breaks agents is --reasoning-format none, which leaves the tags inline — don't use it except to inspect raw output.

Did it actually apply?

Point check_applied.py at a model directory or a .gguf. It finds every template source, renders each, and tells you whether they agree — exits non-zero if the prompt is missing or the two sources disagree.

python3 scripts/check_applied.py /path/to/model-dir
python3 scripts/check_applied.py model.gguf

  [chat_template.jinja]  28162 bytes
     terseness prompt ......... yes
     keeps your system prompt . yes
     retains thinking* ........ yes

  [tokenizer_config.json]  8952 bytes
     terseness prompt ......... NO (found 0x)
     keeps your system prompt . yes
     retains thinking* ........ yes

  *** THE TWO SOURCES DISAGREE ***
  Recent transformers uses chat_template.jinja; oMLX and others read the
  copy embedded in tokenizer_config.json. Right now those RENDER DIFFERENTLY,
  so what you get depends on your runtime. Patch both to the same template.

That case — a fresh .jinja dropped in next to a stale embedded copy — is the most common way this silently does nothing. It also warns if the template names a specific model, which happens when the file was taken from a model repo rather than from here.

It compares what the sources render, not how they are spelled. That matters because the documented way to patch both places is to paste chat_template_oneline.txt into tokenizer_config.json — the minified form of the same template, byte-different by construction. A text comparison flags that recommended state as broken; this one reports:

  Both sources render the SAME prompts — whichever your runtime prefers,
  you get the same behaviour (they differ only as full vs. minified text).

Setting reasoning effort

By default there is no reasoning-effort instruction — you get the tuned terseness behavior and nothing else (that is exactly what medium renders). To turn steering on for a request, set reasoning_effort to low, high, or xhigh. How you pass it depends on the runtime, and one obvious-looking channel does not work:

How you pass it	oMLX	llama.cpp	transformers	Works?
chat_template_kwargs: {"reasoning_effort": "low"} (in the request body)	✅	✅	—	yes — use this
apply_chat_template(..., reasoning_effort="low") (Python)	—	—	✅	yes
top-level reasoning_effort field (the OpenAI API param)	❌	❌	—	no

{"messages": [...], "chat_template_kwargs": {"reasoning_effort": "low"}}

The last row is the trap. The OpenAI-style top-level reasoning_effort field is consumed by the server (oMLX and llama.cpp both use it internally to pick reasoning-parse behavior for formats like harmony/gpt-oss) and is never handed to the chat template — so a custom Qwen template can't see it, and it silently has no effect here. This isn't something the template can fix: a template only reads the variables the runtime binds at render time. If you need the literal top-level field to work against these servers, put a thin proxy in front that copies reasoning_effort into chat_template_kwargs before forwarding. Otherwise, use the chat_template_kwargs channel above — it works everywhere and needs no code.

Verified on both runtimes: with chat_template_kwargs the steering line renders (oMLX prompt grows +38 tokens for xhigh, +26 for low; llama.cpp/minja POST /apply-template shows the same line); with the bare top-level field it does not.

Tests

froggeric ships a test suite upstream; this repo vendors it under scripts/ and runs it against the Sharp template rather than a stock one, so the fork is held to upstream's own invariants.

Script	Covers	v22.3.1
test_v22.py	100 cases — effort steering and aliases, inline <|think_…|> tags, tool-call wire formats, system merging, error escalation, vision parts	100 / 100
test_v21.py	9 cases — the v21-era retention and rendering baseline	9 / 9
fuzz_template.py	property fuzzer, 9 invariants: render, oneline parity, tag balance, content, XML fidelity, JSON validity, warning precision, prefix stability, empty-think prefill	clean over 2,000 conversations

pip install jinja2
python3 scripts/test_v22.py
python3 scripts/test_v21.py
python3 scripts/fuzz_template.py --cases 2000

test_v22.py carries one local change, marked in the file: a shim that strips Sharp's appended terseness block before each assertion. Sixteen upstream tests pin the exact end of the system turn, which is precisely where Sharp appends — without the shim they fail on a difference this repo makes on purpose, and sixteen permanently-red tests would hide a real regression the next time upstream bumps. The shim removes only the block Sharp adds; every other upstream assertion still runs against our rendering. It keys off the terseness marker, so the same file scores a pristine upstream template 100/100 as well, which is how it was checked for being a genuine no-op. Run against the pre-rebase v22.1.1 template it still reports 70/100 — it hides Sharp's intended divergence, not real breakage.

scripts/verify_template.py is this repo's own check and complements those: it re-fetches upstream live, asserts the thinking-on path is still byte-identical to upstream-plus-terseness, and fails if froggeric has moved past the base recorded in BASE — which is what caught the v22.1 → v22.3 drift.

What it doesn't do

• It is not a fine-tune, despite the base_model_relation: finetune tag — that is the closest vocabulary HuggingFace offers for "derived from," and it exists so this repo is linked from froggeric's. No weights are involved. It changes what the model is asked for, not what it knows.
• It does not fix thinking retention by itself — that comes from froggeric's upstream template, which this builds on. If you splice only the terseness block into a stock Qwen template, you get the brevity and not the retention.
• It is not tuned per model. Every model responds a little differently to a terseness instruction; measure yours. The numbers above are from a 27B; a 4B may need firmer wording.
• The table's figures are Qwen3.6 (the plate above is the 3.8 result). The template covers 3.5, 3.6, and 3.8 alike — upstream unified them into one file — but every figure in the table was measured on a 3.6 model.

Credits

Everything structural here is froggeric's work — the retention fix, the tool-calling handling, the error-escalation tiers, the whole template. This repo adds a system prompt, the two fast-mode fixes, and nothing else.

scripts/test_v22.py, scripts/test_v21.py and scripts/fuzz_template.py are froggeric's test suite, vendored so this fork is measured against upstream's invariants; test_v22.py carries the documented shim described under Tests. scripts/minify_jinja.py is froggeric's with one patch: it now preserves newlines inside {% set %}…{% endset %} blocks, which upstream's template doesn't contain and this one does. scripts/check_applied.py and scripts/verify_template.py are this repo's.

Apache-2.0, matching upstream.

Citation

@misc{Qwen-Sharp-Chat-Templates,
  title  = {Qwen Sharp Chat Templates},
  author = {Saga Ishtardottir},
  year   = {2026},
  url    = {https://huggingface.co/peculiar-ragdoll/Qwen-Sharp-Chat-Templates},
  note   = {froggeric's fixed Qwen3.5/3.6/3.8 chat template with an always-on terseness system prompt (v22.3.1)}
}