DeepSeek has uploaded the weights for its V4.1-Flash model to HuggingFace and made it available via Baseten’s Model APIs. The model has 552 B total parameters but only activates 8 B during the prefill step and 16 B for decoding, which makes it cheaper to run than earlier versions. It supports a 1 M token context window and can take text and images as input. Baseten’s serving stack routes requests to GPUs that already hold the needed prefix, further speeding up inference for coding agents.
Why it matters
Developers building coding assistants can lower GPU costs and run longer prompts without extra memory.
