Qwen: Qwen3.8 2.4T A95B (batch)

The qwen/qwen3.8-2.4t-a95b:batch entry was last present in the OpenRouter snapshot of 2026-09-21 and absent from the snapshots of 2026-09-23 and 2026-09-24. A live fetch on 2026-09-24 finds it still gone. The model catalog lists the standard qwen/qwen3.8-2.4t-a95b entry and no batch variant of it; …/models/qwen/qwen3.8-2.4t-a95b:batch/endpoints answers 200 with "endpoints":[]; and …/model/qwen/qwen3.8-2.4t-a95b:batch answers 404, Model not found: qwen/qwen3.8-2.4t-a95b:batch. OpenRouter's variant guide reads that pair plainly: "When the resolved catalog entry does not exist … the single-model route returns 404 and the endpoints route returns 200 with "endpoints": []. Treat either as an unavailable model, not as a signal to retry with the base slug." The batch entry is gone from the router.

Qwen3.8 2.4T A95B is not retired. The standard entry lists seven endpoints when fetched on 2026-09-24 — Novita, Alibaba, SiliconFlow, Venice, Modal, DeepInfra and Together — and the weights at Qwen/Qwen3.8-2.4T-A95B were public and not gated on the same day. The standard entry keeps the qwen/qwen3.8-2.4t-a95b-20260812 canonical slug the batch row carried, and the current catalog holds no :batch entry under any Qwen slug. That shows the batch row is absent and that no replacement batch row appears in the catalog; it does not show what OpenRouter did with the variant.

Which provider served the batch entry is not on record. OpenRouter's batch quickstart says "Every request in a batch runs on a single provider, chosen once at submit time", picked from the eligible :batch endpoints after the account's provider allowlist, data policy and BYOK settings, and that "Endpoints at the same price share traffic." So the provider was a per-submission choice, not a fixed property of the row. One metadata fingerprint points somewhere without proving it: of the seven endpoints on the standard entry on 2026-09-24, exactly one has the 1,010,000-token context and 909,000-token completion cap pinned for the batch row when it went missing, and exactly one supports the same nineteen parameters. Both are Together's; every other endpoint's completion cap is less than a third of it. The batch row's input, output and cached-input rates equal those of six of the seven endpoints, Together's among them. Neither OpenRouter nor Together ties Together's endpoint to the historical batch entry, so the match is an inference from metadata. The pinned values are kept in data/vanished/answered/openrouter-models--qwen-qwen3-8-2-4t-a95b-batch.md.

Together, the endpoint that metadata matches, has not withdrawn the model. Its serverless model list, fetched 2026-09-24, still carries Qwen/Qwen3.8-2.4T-A95B. Its changelog records the model arriving on serverless on August 12, 2026, and no deprecation notice since names it: the September 1 notice and the September 15 removal both list the same four other models. Its batch overview still says "Most serverless models support batch processing". Together's endpoint also still sits on OpenRouter's standard entry.

So the reason is unknown. OpenRouter publishes no per-entry reason for a variant leaving its catalog, and neither Qwen nor Together says anything about a batch tier for this model. What can be said is that it was not alone: between the 2026-09-21 and 2026-09-23 snapshots, nine :batch entries left the catalog together, among them minimax/minimax-m3:batch, qwen/qwen3.5-9b:batch and thinkingmachines/inkling:batch. None of the nine had come back by the 2026-09-24 snapshot.

That is why the feeds: binding stays. If qwen/qwen3.8-2.4t-a95b:batch is a catalog entry again, it binds straight back to this page. The timeline dates the first snapshot without the batch entry; it does not record a retirement of Qwen3.8 2.4T A95B, which is live, open-weight and listed on OpenRouter under seven providers.