← Portada | Mercado

Baseten Puts DeepSeek-V4.1-Flash on Model APIs With 1M-Token Context

2026-09-11 · Fuente:Unite AI

Baseten Puts DeepSeek-V4.1-Flash on Model APIs With 1M-Token Context

The 552B-parameter multimodal MoE model, with asymmetric 8B/16B active parameters and a 1M-token window, is live on Baseten two days after DeepSeek's open-weights release.

Baseten has added DeepSeek-V4.1-Flash to its Model APIs, with the inference provider announcing availability on September 11, 2026. The listing puts a 552B-parameter multimodal model on Baseten's hosted platform just two days after DeepSeek's own announcement of the release, which carried a September 9, 2026 date.

Architecturally, DeepSeek-V4.1-Flash is a mixture-of-experts (MoE) model with an asymmetric activation profile: 8B parameters are active during prefill, while 16B are active during decode. Prefill and decode are the two phases of inference — processing the input prompt and generating output tokens, respectively — so the split allocates more active capacity to generation. All of this runs against a 1M-token context window, with per-token compute drawing on the smaller active parameter counts rather than the full 552B-parameter weight set. The 1M-token window gives the model room for extended inputs — long documents, large codebases, or lengthy conversation histories — within a single context.

On the input-output side, the model accepts both text and image input and generates text output. DeepSeek has published the model's open weights on Hugging Face, meaning the same checkpoint now exists both as an open release for teams that want to serve it themselves and as a managed endpoint on Baseten for those who would rather consume it through the provider's API.

The two-day gap between DeepSeek's September 9 announcement and Baseten's September 11 rollout shows how quickly newly released open-weight models are reaching third-party inference platforms. For Baseten customers, DeepSeek-V4.1-Flash arrives as a text-and-image-input, text-output option combining a 1M-token context window, a 552B-parameter MoE backbone, and asymmetric 8B/16B active-parameter operation, available now through Model APIs.

Etiquetas: Baseten · DeepSeek · DeepSeek-V4.1-Flash · Mixture of Experts · Inference APIs · Open Weights


Lecturas relacionadas


萬安算交所 AIXX · AIXX · Exchange verificado · Ir al mercado