DeepSeek V4.1 Flash 1017 Ultra-fast

DeepSeek introduced V4.1-Flash as the smallest member of a new architecture family, built around an asymmetric Causal Encoder–Decoder design in which a 552B-parameter mixture-of-experts model keeps only about 8B parameters active for input and 16B for output. That split is what makes the model fast: DeepSeek describes it as targeting greater capability, faster inference, and higher throughput while remaining the foundation for scaling toward larger models in the family. The model also handles images natively rather than through a separate vision component, so visual and text inputs share one path through the network rather than being stitched together downstream.

DeepSeek credits new pretraining methods combined with larger-scale reinforcement learning post-training for benchmark results it says place V4.1-Flash ahead of its own flagship models, including V4-Pro, a result the company says was corroborated by tests from multiple outside parties. Efficiency is the other headline: compared with the previous generation, the KV cache requires roughly a quarter of the HBM and an eighth of the SSD storage, which DeepSeek argues directly lowers cache-hit charges that tend to dominate agent workloads. With weights published for open-source deployment and inference support planned with the community, the model suits teams that want frontier-adjacent quality and multimodal input handling at a smaller active footprint and lower memory cost than the tier above it.

Fireworks AIaccounts/fireworks/routers/deepseek-v4p1-flash-1017-ultrafastdeepseek-flash

Open provider page Provider docs

Quick Info

Provider: Fireworks AI
Model key: accounts/fireworks/routers/deepseek-v4p1-flash-1017-ultrafast
Release date: Oct 17, 2026
Last updated: Oct 17, 2026
Knowledge cutoff: 2025-05
Input modalities
Output modalities
Capabilities

Cost

Input token cost: $2.40
Output token cost: $9.60

Limits

Output tokens: 384,000 tokens
Context window: 1,000,000 tokens

Latest news about DeepSeek V4.1 Flash 1017 Ultra-fast

Fireworks AI

Coverage

DeepSeek has introduced DeepSeek-V4.1-Flash, the smallest model in its new architecture family, with native visual understanding and support for multimodal input. It is a 552B-parameter mixture-of-experts model built on a new Causal Encoder-Decoder design, activating just 8B parameters for input and 16B for output, and it ships with new pretraining methods plus larger-scale reinforcement-learning post-training. The announcement highlights a much smaller KV cache than the previous generation, requiring one quarter of the HBM and one eighth of the SSD storage, which DeepSeek says cuts cache-hit charges that often dominate agent costs. V4.1-Flash is live on the DeepSeek API under the deepseek-flash model name, with Hugging Face weights and a technical report published, and partner tools including WorkBuddy, CodeBuddy, and OpenCode listed as supporting it.

Fireworks AI

Coverage

DeepSeek V4.1 Flash 1017 Ultra-fast is a Flash-family variant that serves the October 17, 2026 update of DeepSeek V4.1 Flash through Fireworks AI's accelerated serverless endpoint. The listing describes text and image input alongside configurable reasoning, tool calling, structured outputs, and automatic prefix caching, and notes availability in public preview without a waitlist or session time limit. The same page specifies a 1M-token context window and 1M maximum output, with about 0.8 second latency, 720 tokens per second throughput, and a 95 percent cache hit rate reported for the Fireworks route. It also lists input pricing at $2.40 and output pricing at $9.60 per million tokens, and notes an 11/06/2026 release date for the gateway listing.

Sulat also reports on AI and tech beyond the catalog.Read news.sulat.com