Baseten Integrates DeepSeek-V4.1-Flash with 1M-Token Context
The multimodal model enhances AI inference with advanced architecture.
1Introduction to DeepSeek-V4.1-Flash
Baseten announced the integration of DeepSeek-V4.1-Flash into its Model APIs on September 11, 2026. This multimodal model features a 552B-parameter mixture-of-experts (MoE) architecture, utilizing 8B active parameters for prefill and 16B for decode within a 1M-token context window. DeepSeek released the model’s open weights on Hugging Face, licensed under the MIT License, making it accessible for text and image input with text output capabilities.
2Performance Benchmarks
DeepSeek-V4.1-Flash demonstrates superior performance across multiple benchmarks. It scores 90.6 on Terminal-Bench 2.1, 74.2 on DeepSWE v1.1, and 54.8 on AutomationBench. Compared to its predecessors, V4.1-Flash outperforms V4-Pro with fewer parameters, though it still struggles with complex workflows, necessitating human oversight in agent pipelines. Baseten highlights its coding and agentic capabilities, emphasizing its cost efficiency.
3Architecture Innovations
The model employs a 40-layer Transformer with a 20-layer causal encoder and a 20-layer decoder. Its global key-value (KV) cache is projected from the final encoder hidden states, reducing prefill costs. DeepSeek-V4.1-Flash also introduces Compressed Sparse Attention 2 and Hierarchical Sparse Indexer, cutting the global KV cache to 890 bytes per token, a quarter of V4-Flash’s footprint. These innovations enhance cost efficiency and scalability.
4Training and Deployment
DeepSeek trained V4.1-Flash on a 45T-token multimodal corpus, extending its context to 1M tokens. Post-training involved supervised fine-tuning, reinforcement learning, and on-policy distillation. Baseten’s Inference Stack uses NVIDIA Dynamo with KV cache-aware routing, optimizing request handling. The model is available through Baseten’s Model Library, with dedicated deployments for reserved capacity.
5Cost Efficiency and Pricing
Baseten emphasizes V4.1-Flash’s cost efficiency, achieved through lighter prefill and heavier decode steps. New API pricing took effect on September 10, 2026, with off-peak rates at 50% of peak rates. Starting September 14, all deepseek-v4-pro requests will route to V4.1-Flash at its rates, reflecting its superior performance, speed, and runtime efficiency.
6Partnerships and Support
DeepSeek’s official partners, WorkBuddy and OpenCode, fully support V4.1-Flash. Baseten’s Loops training product will soon integrate with the model, enhancing its capabilities. The company also announced the retirement of V4-Flash and V4-Flash-Vision-Exp, with API compatibility temporarily routing to V4.1-Flash.
7Technical Advancements
V4.1-Flash introduces Engram conditional memory with 196B parameters and DSpark speculative decoding. Each MoE layer uses one shared expert and 384 routed experts, with six active per token. The model’s SWA Bounded Replay mechanism reduces persistent KV footprint to one eighth of V4-Flash, significantly lowering storage requirements.
8Implications for AI Development
DeepSeek-V4.1-Flash represents a significant leap in generative AI, combining advanced architecture with cost efficiency. Its 1M-token context window and multimodal capabilities make it a powerful tool for coding agents and complex workflows. Baseten’s integration underscores the growing importance of scalable, efficient AI models in the industry.
Original article: Unite.AI — Baseten Integrates DeepSeek-V4.1-Flash with 1M-Token Context