Dense Qwen3.8-27B now on Nebius Token Factory
On 30 Sep 2026 Qwen said Qwen3.8-27B is on Nebius Token Factory. Weights shipped 14 Aug. Price and speed come from the card and third parties.
The weights have been open since 14 August. On 30 September Qwen announced the hosted API.
On 30 September 2026 at 11:18 UTC, the official @Alibaba_Qwen account on X wrote that Qwen3.8-27B “is now accessible via @nebiustf”. The post talks about agents and deep research and says the 27 billion parameter dense model is ready for multi-step workflows. That is useful if you already run the 27B on your own machines and want a paid API, or if you want a host without downloading the weights. The Qwen/Qwen3.8-27B weights landed on Hugging Face on 14 August 2026. What arrived on 30 September is hosted inference: Nebius Token Factory lists the model in its Q3 2026 release notes and on a public endpoint card. Qwen sells the model; Nebius sells the API. This newsroom did not call the endpoint.
First: what Token Factory is
Inference is using a trained model: you send text (and, here, images) and get a reply. Token Factory is Nebius’s platform for that, with an API at https://api.tokenfactory.nebius.com/v1/. Think of a kitchen cooking from someone else’s recipe: the weights are still Qwen’s, the servers and the invoice are Nebius’s. The comparison stops at quantization, which stores the model’s numbers with less precision to make the server cheaper. The card describes the service as NVFP4 serverless (you pay per call, without renting a machine around the clock), with vision, tool calling (the model asks a program to run a tool), and a 262,000-token context, in region us-central1. Nebius says the region is global and may change. A token, in this piece, is a chunk of text the model reads or writes.
On 9 October 2026 the Token Factory home did not feature the 27B; the model is on the endpoint card.
The dense Qwen3.8-27B and the MoE in the same generation
Qwen3.8 is Qwen’s open generation after 3.5 and 3.6. At the top of the Hugging Face collection, updated 13 August 2026, sits the MoE Qwen3.8-2.4T-A95B: 2.4 trillion parameters in total, 95 billion active. MoE (mixture of experts) turns on only a slice of the parameters per token. The 27B is the dense member: every parameter is in the bill at every step. Qwen calls that size compact next to the MoE. That is the company’s wording, not a hardware measurement.
The Hub card describes a vision-language model (a VLM reads text and images), 64 layers, a native context of 262,144 tokens, extendable to 1,000,000. The Hub labels the safetensors “Model size 28B params”. Thinking, the mode in which the model writes a chain of reasoning before the answer, is on by default, with reasoning_effort at xhigh (default), medium, and low. That is what Qwen states on the model card.
On the same card the company publishes SWE-bench Pro 61.7; Terminal Bench 2.1 73.0; CoWorkBench 70.7; OSWorld-Verified 84.3; GPQA Diamond 89.2. Qwen ran those numbers. CoWorkBench is an in-house suite. The table highlights cuts where the 27B beats chosen rivals, for example SWE-bench Pro against Opus4.6 Max at 53.4. We did not audit those tests.
Price and speed, according to who measured them
On the Nebius endpoint card, extracted 9 October 2026, the approximate price is $0.45 per million input tokens and $3.00 per million output tokens. The card says prices depend on actual usage. Throughput there: 218 tokens per second. The routing key is Qwen/Qwen3.8-27B.
Artificial Analysis, which sells API rankings, listed Nebius (FP4) among nine providers of “Qwen3.8 27B (xhigh)”. On that page’s cut, with 10,000 input tokens, it measured 366.2 output tokens/s and 6.84 seconds to the first answer token. The two speed figures are not the same test. Mastra, a catalogue aggregator, shows the ID nebius/Qwen/Qwen3.8-27B with the same price, pulled from models.dev, and warns that availability and rates may change.
On OpenRouter, in the same day’s extract, Qwen3.8 27B appears with other hosts and without Nebius. The Token Factory provider page on OpenRouter listed 11 models, without the 3.8-27B. Direct inference on Nebius is not the OpenRouter catalogue. Token Factory post-training docs list Qwen3.5-27B and Qwen3.6-27B for full fine-tuning; Qwen3.8-27B was not on the list.
Limits, cost and who it is for
The talk of agents and multi-step research comes from Qwen’s copy on X and from the model card on long-horizon agentic tasks. RUMOR: there is no independent test on Token Factory showing that the 27B holds up on long tool-calling chains.
This helps people who already ran the 27B locally and want another commercial host, or who want a serverless API with vision and tool calling on the key Qwen/Qwen3.8-27B. It does not cover full fine-tuning of this 27B on Token Factory (the 9 October docs did not list it), and it does not turn the model-card benchmarks into independent proof.
SLA, data retention, and whether the endpoint is serverless only or also dedicated, beyond what the card already says, were not checked here. The top of the same family is still the 2.4T MoE, a different size and a different bill.
What we still do not know
We still do not know whether the card’s approximate price holds in real use, whether the 27B joins the post-training list, or whether long tool chains hold up on this endpoint.
What changed on 30 September is the API address, not the model. Anyone who wants to test should start from the Token Factory card and from their own tasks, not from Qwen’s post.
Sources
Official
- @Alibaba_Qwen on X, 30 Sep 2026
- Nebius, Q3 2026 release notes
- Nebius Token Factory, Qwen/Qwen3.8-27B endpoint card
- Nebius, Token Factory
- Hugging Face, Qwen3.8-27B
- GitHub, QwenLM/Qwen3.8
- Hugging Face, Qwen3.8 collection
- Nebius, post-training models
Independent and aggregators