Qwen-Sharp arrow — agent chat templates

Jinja chat templates for Qwen3.6-format models, built for agent harnesses (tool use, multi-step work). Plain chat is unchanged from Qwen-Sharp v22.5.0; with tools, arrow turns the model's reasoning into action.

File	For
`chat_template_sharp-v22.5.0-arrow4.0.jinja`	Qwen3.6-format models
`chat_template_sharp-v22.5.0-arrow4.0-nex.jinja`	Nex-N2.5 (Nex reasoning-level opener, see below)
`chat_template_sharp-v22.5.0-arrow4.0-minicpm.jinja`	MiniCPM-V 4.7 (openbmb's multimodal content format: bare `<\|image_pad\|>`, parts joined by a newline)

llama-server -m <model>.gguf --jinja --chat-template-file chat_template_sharp-v22.5.0-arrow4.0.jinja

How it works

Lineage: Qwen official → froggeric/Qwen-Fixed-Chat-Templates (thinking retention, reasoning effort, XML/JSON tool formats, error coaching) → peculiar-ragdoll/Qwen-Sharp v22.5.0 (terse block) → arrow (ours). Both upstreams are Apache-2.0; what arrow changes is stated here.

How the prompt is built — one system block, in this order:

Reasoning-effort line (only at low / xhigh; none at the default medium).
# Tools list and tool-call format rules — upstream text, verbatim (format-critical).
The harness's own system prompt.
.arrow project-memory rules (only when file or shell tools exist).
Agent rules (tools present) or Sharp's terse answer rules (no tools) — last, closest to the conversation.

Plain chat is untouched: with no tools, arrow4.0 renders byte-identical to Sharp v22.5.0.
Agent rules: think once, then act; seen once = known; live evidence beats memory; deliver once evidence is enough; smallest change that tests it; stay in scope (project root — a mentioned path is context, not a target); check where things live (paths, URLs, ports) in live config; long jobs in the background with a log; build only needed targets; fix in place; result-first replies in the user's language.
Per tool batch: at most one short note on the result that needs it (repeated call, same error, empty searches, probe loops, timeouts, writes outside scope, skill churn, stalls) plus a STATUS line — goal | scope | calls since last change | read / missing / changed | next move. Pressure rises with inaction: next → overdue → STALLED.
Think cadence: thinking opens at decision points — a user message, an error, a corrective note, every 8th turn; other turns open with the think block closed, so the model acts directly. Computed from the history alone, so it also works when the harness drops past reasoning.
Modes from the request: a question → answer from evidence, no building; "plan first, ask before starting" → present the plan, no changes before approval; the approval reply clears it.
Hand-back: 36 calls without a change (18 on a question) → the reply is pre-started as a report (done, blocker, options) and control returns to the user.
.arrow project memory: .arrow/memory (decisions, commands, traps) and .arrow/plans (goal, - [ ] steps, verify) at the project root — created on first contact, ticked as work lands, re-read after a context summary.
Prompt cache: every note and STATUS line is computed from earlier messages only, so past renders never change and the prefix cache holds.

Switches (chat_template_kwargs, default in brackets): reasoning_effort [medium] · enable_thinking [true] · think_cadence [true] · act_streak [8] · status_line [true] · deliver_at / deliver_at_question [36 / 18, 0 = off] · arrow_memory [true; false = off, "always" = also without file tools] · decisive [true; false drops the agent rules] · terse [true] · tool_call_format [xml].

Reasoning level for Nex

Nex was trained on its own chat template, where the reasoning level is set by how the reply opens: a bare <think> means adaptive (medium), <think> + newline means always think (high). The -nex file opens replies the Nex way; the generic file always opens with <think> + newline. Reasoning level passed by the harness → Nex mode:

none → no thinking
low / minimal → adaptive, with a "keep thinking brief" instruction
medium (default) → adaptive: the model decides whether and how much to think
high / xhigh / max → always think

Changes since arrow.3.1

Think cadence: thinking at decision points only; routine turns act directly.
STATUS line after every tool batch, with rising pressure on inaction.
Scope: project root anchor; approvals and writes outside it are flagged.
Question and plan-first modes, read from the request; approval clears the plan gate.
Hand-back report after a long stall; escalating notes instead of repeated ones.
Tool timeouts and probe loops recognized; long jobs moved to the background.
.arrow rhythm: plan ticks, plan close, memory updates before answering.
English-only template text; replies follow the user's language.

Used in

License

Apache-2.0. Derived from froggeric/Qwen-Fixed-Chat-Templates and peculiar-ragdoll/Qwen-Sharp-Chat-Templates, both Apache-2.0.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Collection including gbuzhf/Qwen-Sharp-Arrow-Chat-Templates

Various

Collection

ICE Quantization Method & Arrow Chat-Template • 2 items • Updated 14 days ago