Prompting Qwen3 Coder

Prompting Qwen3-Coder

Source: model card https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct + its generation_config.json

• https://github.com/QwenLM/Qwen3-Coder + https://qwenlm.github.io/blog/qwen3-coder/ — snapshot 2026-07-16, verified against a local Ollama 0.32.0 install. If this snapshot is >60 days old and you're doing high-stakes work, re-fetch and reconcile.

Read the first item before anything else. It is the one that invalidates whole categories of prompt work, and it cost one project four controlled experiments to learn the hard way.

1. It is a NON-thinking model. There is no deliberation channel. (the big one)

Model card, verbatim:

"This model supports only non-thinking mode and does not generate <think></think> blocks in its output."

Consequences, and they are large:

• /think, /no_think, and enable_thinking do NOT apply. Those are Qwen3-general features. The official Qwen quickstart has no Qwen3-Coder section at all. Sending think is not a no-op — qwen-code issue #1377 reports a hard 400: "qwen3-coder:30b" does not support thinking.
• ollama show qwen3-coder:30b → Capabilities: completion, tools. No thinking. (Compare gpt-oss:20b, which does list thinking.) That one command is the fastest way to check this for any model.
• Any prompt that asks it to "think step by step", "reason carefully first", or "take your time" is asking for a channel that does not exist. It has no exposed thinking channel and does not emit <think> blocks; it cannot externalize deliberation before the answer. It reasons fine when asked directly (measured — §2, §7), but whatever reasoning happens shows up in the emitted answer tokens themselves, never in a separate hidden pass first.

The practical rule: if you need deliberation to precede the answer, either give it room to reason in its output (see §3), or use a different model. Asking it to "think harder first" targets a channel it does not have — it is ineffective, not a fix.

2. It is OPTIMISED as an agentic patch-writer — but that is not a capability bound

⚠️ CORRECTED 2026-07-16, same day, by measurement. An earlier version of this section said "it is an agentic patch-WRITER, not a critic — this is a routing fact." That was wrong, and it was my inference, never Qwen's claim. Measured: asked directly about a lossy _down migration, qwen3-coder found the bug 5/5 with ~300 tokens of correct reasoning about post-up divergence. It critiques fine. Optimisation target ≠ capability ceiling. Four official sources name what it was tuned for; none states a limit, and the research capture behind this guide explicitly logged "NOT FOUND: whether code review is an intended use." I let absence of evidence harden into evidence of absence. Don't repeat it. Evidence: a controlled silence-probe eval, 2026-07-16 (ablation details in §7 item 2).

Read the positioning as a prior, not a verdict. Every official source positions it the same way, and none of them says "reviewer":

• Release blog: "our most agentic code model to date" — framed around multi-turn environment interaction.
• Model card: "excels in tool calling", "Agentic Coding" with a specially designed function-call format.
• Ollama: "advanced long-horizon reinforcement learning on SWE-Bench", "optimized for repository-scale understanding".

Its headline RL objective rewards producing a working patch — not emitting a critique of someone else's.

Whether one-shot code review is an intended use: NOT FOUND. No vendor says it can't. No vendor says it can. Treat that as absence of evidence — and note that measurement (2026-07-16) shows it can review when asked directly. So "make it a reviewer" is a decision the vendor does not bless, not one it forbids.

Its documented sweet spots: agentic tool-driven coding, code completion, fill-in-the-middle, repo-scale understanding, 358 languages.

3. Prefer tool calling over format: "json" — on vendor grounds, not as a silence fix

The vendor's only documented structured-output path is tool calling (the card documents an OpenAI-compatible tools API; the README ships a dedicated tool parser for SGLang and vLLM). Ollama honours this natively: PARSER qwen3-coder, Capabilities: … tools.

Tool calling is the documented structured-output path; grammar-constrained JSON (format: "json" / GBNF) guidance was NOT FOUND for this model — neither endorsed nor warned against. That is absence of evidence, not evidence of prohibition: it is undocumented for this model, not forbidden.

⚠️ CORRECTED 2026-07-16 by measurement — the mechanism below is REFUTED as a cause of SILENCE. A 2×2 factorial (format:json × brace-shape, n=10 each, qwen3-coder) came back 0/10 in every cell — removing the grammar changed nothing. The real cause was the reviewer prompt's suppression clauses (the same model, asked the same question without that framing, found the bug 5/5). Tool calling is still the vendor's documented path and the advice below still stands on vendor grounds — just do not expect dropping format: "json" to un-silence a quiet seat. Evidence: the same 2026-07-16 probe (§7 item 2).

The REFUTED hypothesis — why grammar was thought to hurt THIS model (an inference from verified facts, never a Qwen claim — priced at ~50% and refuted 0/10 in every cell of a 2×2; retained only because the reasoning is instructive and the vendor facts hold, NOT as a live explanation):

A non-thinking model + a grammar forcing the first token to be { = nowhere to reason. The model's first structural decision is {"findings": [ and then immediately whether to emit ]. Under that constraint the empty/null answer is the shortest, highest-probability, always-valid completion. Emitting real content requires having already solved the problem with zero deliberation tokens spent. Plausible — and measured false. Removing the grammar changed nothing (0/10 → 0/10). Do not carry this mechanism forward as a cause.

The CONFIRMED cause — the reviewer prompt's own suppression clauses — produces a "silent or flooding" signature:

• Suppression clause left in → fast, well-formed, confident empty results. Not a crash. Not malformed. Just nothing. Which means whiff-detection and "did it error?" checks cannot catch it.
• Suppression clause removed / the prompt pushed for output → flooding / confabulation (measured: it flagged a bug-free control diff 9/10 and produced more findings on clean code than on buggy — §7 item 2).

If you see either pole, suspect YOUR PROMPT first — measured. The grammar was the intuitive culprit and it was refuted 0/10 in every cell of a 2×2. Suppression clauses in the reviewer framing were the actual cause. Check what your prompt tells it NOT to do before you touch the output format.

In order of preference:

1. Tool calling — the documented path, and the output mode it was actually RL-trained for. Highest effort.
2. Free prose, parsed leniently — give it room to reason in its output, then extract. Cheapest test.
3. format: "json" — what most harnesses reach for first. Undocumented for this model (not forbidden), so worth moving off toward the documented path — but it is not what silences a seat, so do not chase it as a fix.

4. Official sampling parameters — use them; don't quietly override them

From the model card's Best Practices, corroborated exactly by the shipped generation_config.json (the model's own decode defaults — the strongest primary source):

Recommended output length: 65,536 tokens for most queries.

Ollama's packaging is faithful — ollama show --modelfile qwen3-coder:30b reproduces all four exactly. So if your numbers differ from the table, that is your harness deviating, not Ollama. Check your own code before blaming the model. A harness that sets one global temperature across several models is the usual culprit; make it per-seat.

The greedy-decoding warning — get the scope right. The warning "DO NOT use greedy decoding, as it can lead to performance degradation and endless repetitions" is real and OFFICIAL, but it lives on the Qwen3 general quickstart, not on the Qwen3-Coder card. Two honest caveats: its documented failure mode is endless repetition, not silence, and Qwen's own Coder README recommends do_sample=False (greedy) for FIM — so Qwen is not globally anti-greedy here. Low temperature is hygiene to fix, not an explanation for an empty model. Don't over-read it.

5. Context

• Native 262,144 (256K); extendable to ~1M via YaRN (org-level README only — there is no YaRN section on the 30B model card, so no per-model "when to enable" guidance or stated cost exists. Don't cite one).
• The Ollama default context is version-dependent — do not rely on any default. The tested install (Ollama 0.32.0, 2026-07-16) served 4096 regardless of the model's native window. Current Ollama has moved to VRAM-dependent defaults (roughly 4K / 32K / 256K chosen by available memory), so the number you get now depends on your Ollama version and hardware. Precedence is stable: API param > OLLAMA_CONTEXT_LENGTH > Modelfile PARAMETER > built-in default. Set num_ctx explicitly rather than trusting whatever the default happens to be, or you may silently truncate.
• On constrained hardware, num_ctx is the big memory lever: the native 256K window blows the KV cache to ~45GB and thrashes a 24GB machine.

6. Prompt format

Standard ChatML via tokenizer.apply_chat_template(), [{"role": ..., "content": ...}]. Specials: <|im_start|>, <|im_end|>, <|endoftext|>. Ollama compiles this into RENDERER qwen3-coder, so template corruption is an unlikely failure mode there.

What breaks the template: NOT FOUND — no official "don't do X" list exists.

Untested, no vendor guidance either way: whether a system role helps, and whether splitting instructions and content across messages changes behaviour. If you test it, you're generating new evidence, not applying known guidance — so use a control.

7. Debug checklist — empty or garbage output

Run these in order. Each is cheap and rules out a whole class:

1. ollama show <model> → does it list thinking? If not, delete every "think first" instruction and stop trying to enable a mode that isn't there.

2. Does your prompt give it explicit permission to return nothing? This is the highest-yield check — do it before the grammar. A placebo-controlled ablation (7 arms x 2 diffs x n=10, 2026-07-16) isolated one clause as a binary switch:

   "If the diff looks clean, return an empty findings array. Do not manufacture findings."

   Removing that sentence alone: 0/10 → 10/10 speaks. Removing "do not invent issues", "no style nitpicks", or the "other reviewers are looking at this too" framing: 0/10 each — inert. A same-meaning paraphrase (placebo): 0/10 — so it is that clause, not prompt perturbation.

   ⚠ But do NOT just delete it. Removing it made the model flag a bug-free control diff in 9/10 runs and produce MORE findings on clean code (30) than on buggy code (23) — it pattern-matches the shape of a risky construct rather than detecting the bug. The clause is load-bearing FP suppression. For this model, on this diff class, it is silence or flood — there is no dial. Never ablate it without a bug-free twin diff in the same batch; that twin is the only thing that catches this.

3. Are you sending format: "json" / a grammar? → worth trying free prose + lenient parse on vendor grounds (§3) — but measured NOT to be a silence cause. Do not expect it to fix an empty seat.

4. Do your sampling params match §4? → if not, that's your harness, not the model.

5. Is num_ctx set explicitly? → the Ollama default is version-dependent (4096 on the tested 0.32.0; newer builds pick a VRAM-dependent default) and can truncate silently — don't rely on it (§5).

6. Positive control — run this the moment output looks uniform: ask the model a plain question about the input ("which files changed?") and ask it directly about the suspected bug, with no role framing. If it answers both, the model is fine and your scaffold is the bug. This one check overturned two confident wrong conclusions in a single session.

7. Ruled out already, don't re-derive: thinking toggles (§1), YaRN/context extension (§5), Ollama-vs-vendor param mismatch (§4 — there isn't one), and ollama#10976 "empty output" (needs tools + think; if you send neither, it's not your bug).

8. Still empty after the positive control shows it can answer? → now it's scaffold, not model. If the positive control also comes back empty, only then consider the wrong-model conclusion (§2) — and note that conclusion has been drawn wrongly before.

8. The meta-lesson (why this skill exists)

The originating project — a local-model review council — ran four controlled prompt experiments (120 runs, predefined rubrics) trying to make this model contribute as a review seat. All four were rejected, and the conclusion drawn at the time — "the model can't think, so it can't review" — turned out to be wrong.

Read the card and run ollama show before you design an experiment. The real capability the model lacks is an exposed deliberation channel (no <think> blocks, no room to reason before the answer) — NOT the ability to reason at all. That distinction is the whole lesson: prompting it to "think step by step" asks for a channel that doesn't exist, but the model still reasons fine when you give it room in its output or ask it directly.

And the sequel matters more. All four rejected experiments ADDED targeting instructions. None removed the prompt's suppression clauses. 0-for-4 on adding says nothing about removing — and it was the suppression clauses, not a missing reasoning capability, that produced the silence. The 2026-07-16 probe refuted the "wrong model" conclusion outright: it finds the bug 5/5 when asked without the reviewer framing.

Three lessons, each paid for:

1. Optimisation target ≠ capability ceiling. Don't turn "tuned for X" into "can't do Y."
2. Run a positive control before concluding a model can't. Ask it directly, no framing.
3. Predefine the metric and the threshold — never the conclusion. A rule that bundles its own inference will fire confidently and be wrong. (Three-for-three in one session.)

9. Known gaps — stated, not filled

• No vendor guidance on empty responses, over-refusal, or under-flagging. §3's mechanism is an inference.
• No statement on whether code review is an intended use.
• No presence_penalty published for Qwen3-Coder. Don't invent one.
• No official position on format: json / GBNF for this model — so §3's advice is an evidence-backed hypothesis, not a vendor-blessed rule. Labelled deliberately.
• Other variants exist (verified in the official list): Qwen3-Coder-Next(-Base/-FP8/-GGUF), 480B-A35B-Instruct(-FP8), 30B-A3B-Instruct-FP8. This skill is scoped to 30B-A3B-Instruct; do not assume it transfers, especially to Next.