Two four-billion-parameter models. Different evidence for the shortlist.

Source reading: 14 October 2026. These are Qwen-reported scores from the two model cards, not measurements by Theatre. Each card describes a 4.0-billion-parameter model. Instruct runs without a thinking mode; Thinking produces a reasoning trace before its answer.

The table retains the score values as published. AIME25 is the math row; LiveCodeBench v6 (February–May 2025) is the coding row; IFEval is the instruction-following row. Higher reported scores favour Thinking on these three rows. They do not rank every task.

Output budgets differ in the evidence

The Thinking card gives an output length of 81,920 tokens for its reasoning and coding tasks, including the selected math and coding rows, and 32,768 for other tasks. The Instruct benchmark output budget is not stated in its performance section. This is not a controlled same-budget comparison.

The date is a source-reading date

The benchmark run dates and repetitions are not stated for these rows in either performance section. The LiveCodeBench dataset window is not a run date. The cards do not give uncertainty intervals for these scores.

Speed and cost are separate decisions

These rows supply no matched latency, serving-hardware, quantisation or cost comparison. Parameter count alone does not establish equal memory use or response time. A higher benchmark score does not predict accuracy on the client’s own cases.

We retain Thinking on the shortlist for math and coding because Qwen reports higher AIME25 and LiveCodeBench v6 scores than for Instruct in these two 2507 model cards. The published comparison does not isolate an equal output-token budget, and supplies no matched latency or cost figures. The choice for this project therefore depends on performance on its own cases within its response-time and cost limits.

Use that paragraph with the source links and budget limits. It describes a shortlist decision from published evidence; it does not claim that either model has passed your project’s acceptance cases.

For a commissioned comparison, scope and permission to share the results are agreed before an order. Theatre accepts an order only when it can deliver the agreed packs.

Keep the job and budget the same

Use identical prompts, a fixed set of representative cases and an equal output-token budget. Fix the scoring rules before seeing the answers; count incomplete or truncated answers as well as successes.

Record what the client pays for

Keep the model revision, template, quantisation, hardware and sampling settings with the results. Record repetitions, wall-clock time and input/output token counts. Decide the acceptable response time and cost before selecting a model.

Read the comparison-pack terms

This brief selects facts and adds a reading of Qwen’s model cards; it does not reproduce their full tables, prompts or benchmark datasets. The cards are pinned to the revisions linked below. Both repositories carry Apache License 2.0. Model names identify the sources; Qwen does not endorse this brief.