DeepSeek · current

DeepSeek Flash (V4.1-Flash)

DeepSeek Flash is DeepSeek’s current recommended API model. The API id is deepseek-flash. The pricing page lists the model version as DeepSeek-V4.1-Flash. Legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp still route here and bill at Flash prices.

Status: Current
Price: $0.15 in, $0.60 out (off-peak, cache miss)US dollars per million tokens (word pieces) in and out, unless noted
Released: 10 Sep 2026 (announced)

Training

DeepSeek’s 10 Sep 2026 announce describes a 552B-parameter MoE with a Causal Encoder–Decoder architecture (8B active parameters for input, 16B for output) and larger-scale RL post-training. Full dataset mix and FLOPs belong in the tech report on Hugging Face when cited. Knowledge cutoff is not published on the pricing page.

Inference

Thinking mode is on by default; non-thinking is available. Native vision is supported on Flash. Tool calls, JSON output, the Responses API, and an Anthropic-compatible API are supported. FIM completion is non-thinking only (beta).

Context

The documented context window is 1 million tokens. Max output is 384K tokens on the pricing page.

Methods

Dual thinking and non-thinking modes. Effort levels are low, high, and max in the docs. Beyond “larger-scale RL post-training,” the exact train-time RL recipe is not published on the announce page.

Benchmarking

The announce claims benchmark results ahead of flagship models including DeepSeek-V4-Pro. Specific shared leaderboard numbers (HLE, Arena, and similar) are not copied here from that announce extract. Prefer the DeepSeek tech report or Hugging Face materials when you need dated figures.

Hardware

The announce claims KV cache use of about 1/4 HBM and 1/8 SSD storage versus the prior generation. Training cluster size is not published. Open weights: DeepSeek-V4.1-Flash on Hugging Face

Access and pricing

Base URL is https://api.deepseek.com (OpenAI format) or https://api.deepseek.com/anthropic. Peak and off-peak pricing (off-peak is half of peak). Per 1M tokens on the pricing page:

	Off-peak	Peak
Input cache hit	$0.003	$0.006
Input cache miss	$0.15	$0.30
Output	$0.60	$1.20

Peak hours (docs): 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday, excluding Chinese public holidays. Concurrency limit listed: 2500.