GPT-OSS vs Qwen for short writing: which draft would we use?

Sam could wait for a blue backpack or switch to a green one. The shop could send green within two working days of confirmation. It couldn't promise when the parcel would arrive.

That was the catch in a fictional customer email we gave GPT-OSS 120B and Qwen 3.8 27B. We wanted a reply that Sam could understand without committing the shop to a delivery date. GPT-OSS finished all three attempts. Qwen left two answers blank and abandoned the third halfway through the offer.

We'd take a slightly clumsy email over no email. But Qwen did eventually complete the job after we changed one setting. Leaving that out would make for a cleaner verdict and a worse comparison.

We tested the models through Groq on 17 October 2026. Alongside the customer email, we asked for an event announcement and the actions from a meeting. Each model got three independent attempts per brief. The results below include the failures, followed by the separate checks we ran to understand them.

GPT-OSS gave us an email to edit

The original notes said blue was out of stock. More stock was expected on 5 November, but that date wasn't confirmed. Green was available at the same £42 price. Sam needed to reply by 28 October; otherwise the order would stay open.

We set a 60–80-word limit. GPT-OSS's replies were 80, 76 and 83 words long. The middle one is the version we'd start editing:

Hi Sam,

Your blue backpack is out of stock. We expect delivery on 5 November, but the date is not confirmed. You can wait for blue or switch to the green version, which is in stock and can be dispatched within two working days after you confirm the change (not a guaranteed delivery date). The price remains £42 either way. Please let us know your choice by 28 October; otherwise we will keep your order open.

The parenthesis makes the message harder to read. “Not a guaranteed delivery date” sounds like a reminder to the person composing it, and it sits inside the longest sentence. We'd give the arrival-date restriction a sentence of its own. The rest gets Sam through the choice without threatening to cancel an unanswered order.

We'd also clarify what the shop expects on 5 November. In the source, that's incoming stock. A customer reading “We expect delivery” might think it means their parcel. Keeping a date provisional is only part of the job; the email also needs to say whose delivery we're talking about.

Here's our edit of the source. This is an editorial example, not another model response:

Hi Sam,

We're out of blue backpacks. We expect more on 5 November, though that date may change. You can wait or choose the green one at the same £42 price. We'll dispatch green within two working days of you confirming; we can't promise an arrival date. Please let us know by 28 October. If we don't hear from you, we'll keep your order open.

The 83-word GPT-OSS version would need trimming to meet the brief. Three words don't sound like much until the destination has a fixed limit. We'd check the count ourselves before dropping any of these replies into a support template.

Qwen used its answer budget before answering

Qwen had no complete email under the main test settings. One reply stopped after 30 words, as it began describing the green option. The two blank responses reported spending all 2,048 allowed completion tokens on reasoning.

That allowance covered reasoning and final text together. We'd asked for a short email, but the model could use the entire budget before writing it. The failure was visible in the response data: both blank runs ended because they reached the limit.

For the main comparison we'd selected medium reasoning for both models. It was a common setting we could request, though a matching label doesn't mean equal computation. We repeated the first email request with twice the allowance, 4,096 tokens each. GPT-OSS returned the same finished 80-word reply. Qwen again returned no final text, with 4,096 reported reasoning tokens.

Next we restored Qwen's original cap and switched reasoning to none. It produced a complete 88-word email. The stock date stayed unconfirmed. Green would dispatch after Sam agreed to the change, and the order would remain open without a reply. We could finally assess the writing.

Its “We anticipate a delivery” opening was more formal than the shop needed. It also spent words repeating the restrictions on upgrades and delivery dates. Those restrictions came from the source, but the reply still exceeded our ceiling by eight words. We'd shorten it before using it.

For someone reproducing this through Groq, the relevant request fields were:

{
  "model": "qwen/qwen3.8-27b",
  "reasoning_effort": "none",
  "max_completion_tokens": 2048
}

These are the settings from the successful diagnostic, not a complete API request; the source email and instructions still go in the messages. Groq's API reference documents none as this model's default. Our medium-mode test had changed that default.

For this short-writing job, we'd try the default before raising the token cap again. We wouldn't carry Qwen's blank-answer result into a claim about every configuration or about its consumer chat app. The mode change gave us direct evidence that the same brief could complete.

Qwen was clearer about the workshop's places

The announcement was for Night Photography for Beginners: 25 October 2026, 18:00–20:00, in the Mill Room at Riverside Arts Centre. There were 36 places at £18 each, including printed notes. Attendees could bring a camera or a smartphone. Tripods were optional.

Old planning details remained in the source: a proposed 24 October date and £15 price, plus an unscheduled second session. The room could physically hold 60 people. All six announcements used the confirmed date and price and avoided offering the room's full capacity as tickets. Every reply was within the requested 80–100 words.

GPT-OSS nevertheless managed this sentence in its second attempt:

This hands‑on workshop will be held across 36 public places, offering practical tips for low‑light shooting.

There was one venue. The 36 places were seats.

We'd used “public places” in the source, which was awkward wording on our part. Qwen understood the available-ticket meaning every time. GPT-OSS did twice, then attached the number to where the workshop would happen. In the final announcement we'd write “36 places available”.

GPT-OSS also added “All skill levels are welcome” to the beginners' event. Qwen's first reply offered help capturing “the city's evening light”. The notes hadn't specified a city or an outdoor session. We'd remove both additions. The supplied details already gave the models enough to write an inviting announcement.

The booking deadline produced a less conspicuous change. GPT-OSS wrote “Book before 20 October 2026” in one reply; the source said booking closed on 20 October. We'd restore the closing-date wording so readers didn't interpret the deadline as the day before.

Qwen's treatment of the places was more consistent across these attempts. Neither model produced copy we'd publish without editing.

Both handled the meeting actions

Lina was to send two layouts by 22 October. Omar needed to check the signup form with a screen reader, with no deadline agreed. Priya would send a test newsletter by 26 October, but only after Omar approved the form.

We asked for exactly four bullets, each no longer than 25 words. The fourth concerned an unresolved move from monthly to fortnightly publication. The source also mentioned 1,200 subscribers, which was background rather than an assignment.

Every response kept the missing deadline visible and left the publication-frequency discussion unresolved. Priya's task still depended on Omar's approval. GPT-OSS's first answer said “after Omar approves the form”; Qwen's said “only after Omar approves the form”. All six replies met the bullet limits, and none invented work from the subscriber figure.

We'd use either model's list as a starting point here. For the emails, GPT-OSS was more reliable at finishing under our medium-reasoning settings. Qwen needed the mode change before we had a complete draft, and it still missed the word limit. For the announcement, Qwen was better at keeping the seats distinct from the venue. Those are the differences we'd actually use to choose between them for these jobs.

The main requests used Groq's hosted models, openai/gpt-oss-120b and qwen/qwen3.8-27b, temperature 0.2, medium reasoning, a 2,048-token cap and seeds 1701–1703. The larger-cap pair and Qwen's non-reasoning request were separate diagnostics. No browsing or corrective follow-up was included in the 18 main responses.