Qwen — Alibaba Model Studio batch¶

qwen:batch is Alibaba Model Studio's batch tier for the Qwen family: 50% off standard, OpenAI-shaped end to end, with two things the OpenAI venue does not have — a region, and a completion window you choose.

Unverified live

No batch has yet been submitted through this driver, and there is no receipt in receipts/ for qwen:batch. The request dialect, region base URLs, window validation and status mapping are exercised network-free against a fake client; the 50% spread is a published number here, not a settled one. Groq answered 403 not_available_for_plan, Mistral 402 enable billing, and Gemini was gated until billing was switched on — the first sub-cent live run is what verifies a driver, and this one has not had it.

Install and configure¶

pip install "offpeak[qwen]"        # an alias of the openai extra
export DASHSCOPE_API_KEY=sk-...    # or ALIBABA_API_KEY; both names are in use

The driver refuses to build a client without one of those keys rather than letting the openai SDK fall back to OPENAI_API_KEY.

It is opt-in — not in default_venues().

import offpeak
from offpeak.venues import QwenBatch

jobs = [offpeak.job("qwen3.7-max", f"Summarize:\n\n{d}", max_tokens=512) for d in docs]
results = offpeak.run(
    jobs, deadline="2d", venues=[QwenBatch(region="intl", completion_window="48h")]
)

Region¶

Model Studio is two deployments, priced and provisioned separately. A key is issued for one and does not work at the other.

`region`	Base URL
`"intl"` (default)	`https://dashscope-intl.aliyuncs.com/compatible-mode/v1` — Singapore
`"cn"`	`https://dashscope.aliyuncs.com/compatible-mode/v1` — Beijing

The bundled sheet carries the international rows in USD. The Beijing region is priced separately, in its own currency, and runs its promotions per region, so a region="cn" run settles against a rate that is not that region's. Override with prices.register_price() if you run there.

The window¶

completion_window accepts an integer with an h or d unit anywhere in the documented 24h–336h range — "24h", "72h", "14d". The driver refuses anything outside it before the upload, where the venue would refuse it after.

The price does not vary with the window: the docs publish one batch rate and one window range and say nothing about the two interacting. That is the same term structure Groq publishes — a longer window buys completion probability, not price — and it is read the same way here. If Alibaba ever prices the curve, the driver's docstring says where that assumption lives.

What is on the sheet¶

Model	Standard	Batch (50%)
`qwen3.7-max`	$2.50 / $7.50	$1.25 / $3.75
`qwen3.8-max`	$2.00 / $6.00	$1.00 / $3.00

USD per 1M tokens, input / output, international region, from alibabacloud.com/help/en/model-studio/model-pricing — qwen3.7-max first read 2026-08-21 and confirmed 2026-08-30, qwen3.8-max read 2026-08-30. The page marks qwen3.7-max "Limited-time 50% off" with no date it runs through; a PromoNote needs one, so there is none, and the number may step up unannounced.

Only the flagship rows are carried. The plus and flash families are tiered by context length and priced differently in thinking and non-thinking mode, and the sheet has neither dimension — a single number for them would be wrong for most requests, so they resolve to None rather than a guess.

What was checked, and what was not¶

Read from the batch interface docs on 2026-08-30: batch "costs are only 50% of real-time calls"; files upload with purpose="batch"; the batch is created with endpoint matching the url on every JSONL line (/v1/chat/completions for text); the two base URLs above; completion_window "Range: 24h-336h"; the key is DASHSCOPE_API_KEY.

Not checked: the Singapore model list beyond the four the batch page names (qwen-max, qwen-plus, qwen-flash, qwen-turbo), and whether the versioned ids on the pricing page are batchable there. A batch on an unsupported model fails at the venue, after the upload, and run() rescues it through the sync fallback at list.

Routing¶

supports() claims Model Studio's own spelling — qwen-max, qwen3.7-max, qwen-plus — and deliberately not the qwen/… namespace, which is how another catalogue (Groq's) spells the open-weight models it serves. A bare qwen prefix would route a Groq-spelled id to Alibaba.

max_tokens is passed through as the caller spelled it; the max_completion_tokens rewrite the OpenAI driver applies to its newer families does not touch Qwen ids.