Baseten Adds DeepSeek-V4.1-Flash to Model APIs With 1M-Token Context

DeepSeek-V4.1-Flash is available now on Baseten Model APIs, Baseten announced on September 11, 2026, bringing the 552B-parameter multimodal mixture-of-experts (MoE) model, which pairs 8B active parameters for prefill with 16B for decode across a 1M-token context window, to the inference provider’s platform.

DeepSeek released the model’s open weights on Hugging Face, and DeepSeek’s own announcement is dated September 9, 2026. The model accepts text and image input and generates text output, and the model card states that the repository and weights are licensed under the MIT License. Baseten describes V4.1-Flash as DeepSeek’s third open-weight flash release of 2026 and as the only model of its scale to use what DeepSeek calls a Causal Encoder-Decoder architecture. Support for Baseten’s Loops training product is coming soon, according to the company.

Reported Benchmark Results

The model card reports instruct-model results at the maximum reasoning effort setting of 100: V4.1-Flash scores 90.6 on Terminal-Bench 2.1, compared with 82.7 for V4-Flash and 87.9 for V4-Pro; 74.2 on DeepSWE v1.1, compared with 54.4 and 62.7; and 54.8 on AutomationBench, compared with 37.7 and 43.2. Baseten highlighted the same coding and agentic figures, saying V4.1-Flash beats V4-Pro with roughly a third of the total parameters, while cautioning that a 54.8 on AutomationBench means the model fails roughly half of complex workflows and advising teams to keep a human in the loop for agent pipelines.

In the card’s comparison with frontier models at maximum effort, V4.1-Flash posts 90.9 on GPQA Diamond, a Codeforces rating of 3471, and 63.9 on HLE with tools. Baseten states that V4.1-Flash is DeepSeek’s first non-experimental model with native image input, a capability previously limited to the experimental V4-Flash-Vision-Exp; its table reports 78.9 on Chartography and 49 on ZeroBench for the new model, against 64.3 and 35 for the experimental one.

Causal Encoder-Decoder Architecture

According to the model card, V4.1-Flash organizes a 40-layer Transformer as a 20-layer causal encoder followed by a 20-layer decoder, with the decoder’s global key-value (KV) cache projected from the final encoder hidden states rather than derived from each decoder layer’s own hidden states. The design activates 8B parameters per token during prefill and 16B during decode; Baseten contrasts that with V4-Flash, which activates 13B for both steps, framing the change as trading a heavier decode for a much lighter prefill, a setup Baseten said boosts cost efficiency for coding agents whose agentic loops generate far more prefill tokens than decode tokens.

The card reports that the model’s Compressed Sparse Attention 2 assigns each attention layer one of three static modes (Full, Reindex, or Reuse), with a Hierarchical Sparse Indexer in the decoder bounding deeper indexing cost independently of context length. Combined with FP4 main KV caching, those designs reduce the global KV cache to 890 bytes per token, roughly one quarter of V4-Flash, the card states. A separate mechanism, SWA Bounded Replay, reconstructs missing sliding-window-attention KV states by replaying only the most recent tokens, reducing the persistent KV footprint to roughly one eighth of V4-Flash. DeepSeek’s announcement puts the savings at one quarter the HBM and one eighth the SSD storage of the previous generation.

Each MoE layer uses one shared expert and 384 routed experts with six routed experts active per token, and the model adds Engram conditional memory with 196B parameters alongside DSpark speculative decoding, according to the card. DeepSeek trained the model from scratch on a 45T-token multimodal corpus, trained its sparse attention at a 64K sequence length, and extended context to 1M tokens at 34T tokens. Post-training follows a standard supervised fine-tuning, reinforcement learning, and on-policy distillation sequence, with substantive changes concentrated in large-scale automated synthesis of agent tasks and environments, and the model exposes a continuously controllable reasoning effort setting from 1 to 100.

DeepSeek API Transition and Baseten Serving

DeepSeek states that V4-Flash and V4-Flash-Vision-Exp are retired on its platform, with the old API model names temporarily routing to V4.1-Flash for compatibility. New API pricing took effect at 04:00 UTC on September 10, 2026, with off-peak rates set at 50% of peak rates, and DeepSeek names official partners WorkBuddy (including CodeBuddy) and OpenCode as fully supporting V4.1-Flash.

Baseten said its Inference Stack serves the model using NVIDIA Dynamo with KV cache-aware routing, steering each request to the replica already holding its prefix rather than whichever replica is free. The model is offered through Baseten’s Model Library, with dedicated deployments available for teams needing reserved capacity.

Starting at 04:00 UTC on September 14, 2026, all deepseek-v4-pro requests will route to V4.1-Flash at V4.1-Flash rates, an arrangement DeepSeek said will continue until V4.1-Pro launches; the lab said tests by multiple parties put V4.1-Flash ahead of V4-Pro on performance, cost, speed, and total runtime.