model releaseDeepSeek

DeepSeek V4.1 Flash 1017 launches with 552B CED backbone, 1M context, $0.15/$0.60 per 1M tokens

TL;DR

DeepSeek has released V4.1 Flash 1017, an October 17, 2026 snapshot built on its new Causal Encoder-Decoder architecture. It uses a 552B-parameter sparse MoE backbone with a 1M-token context window, native image understanding, and pricing of $0.15 input / $0.60 output per 1M tokens on OpenRouter.

3 min read
0

DeepSeek V4.1 Flash 1017 at a glance

Context window1049K tokens
Input$0.006/1M tokens
Output$0.5/1M tokens

How DeepSeek V4.1 Flash 1017 ranks now

BenchmarkScoreRankSource
AA Intelligence Index32.6 index36 of 124Artificial Analysis, Nov 2026 ↗
Speed (tok/s)155 tok/s32 of 127Artificial Analysis, Nov 2026 ↗
Hallucination Rate (lower is better)66.8%57 of 109Artificial Analysis, Oct 2026 ↗

Rank is against every model in our database with a sourced score on that benchmark, each counted once with its best score.

All DeepSeek V4.1 Flash 1017 specs and scores →

DeepSeek has released DeepSeek V4.1 Flash 1017, an October 17, 2026 snapshot of its V4.1 Flash model. It is the first DeepSeek model built on the company's Causal Encoder-Decoder (CED) architecture. It has a 1.0M-token context window and is priced at $0.15 per 1M input tokens and $0.60 per 1M output tokens, according to its OpenRouter listing.

Key specifications

  • Architecture: Sparse mixture-of-experts on a 552B-parameter backbone, using the CED design
  • Active parameters: 8B on input and 16B on output, an asymmetric split DeepSeek says keeps per-token compute low relative to total size
  • Context window: 1.0M tokens
  • Pricing (DeepSeek endpoint): $0.15 input / $0.60 output per 1M tokens; cache reads $0.003 per 1M tokens
  • API model name: deepseek-flash; OpenRouter ID deepseek/deepseek-v4.1-flash-1017
  • Release date: October 17, 2026
  • Benchmark scores and training cutoff: not yet disclosed in the source material

Native vision and compressed caching

Image understanding is built into the architecture. Visual and text embeddings were trained jointly from the start of pre-training. The earlier experimental V4 Flash Vision Exp (13B active out of 284B total) added vision to an existing text model instead.

DeepSeek also says compressed KV caching cuts cache memory to roughly a quarter of the previous Flash generation. The company claims this significantly lowers costs on agentic workloads. No independent measurements have been published.

Intended use and company claims

DeepSeek positions the model as the cost-efficient tier of the V4.1 family. Its stated targets are:

  • Coding, terminal and computer-use agents
  • Long-horizon tasks that must run across many steps
  • Long-context analysis

DeepSeek claims the model exceeds V4 Pro on performance, speed and task completion time. No benchmark figures accompany the claim, so it cannot be verified yet. V4 Pro 0423 had 1.6T total and 49B active parameters, so the claim covers a model with far fewer active parameters than its predecessor.

Availability

The snapshot is served by DeepSeek and listed on OpenRouter, which reports 99.48% availability over the last three days. It is not currently available through EU in-region routing. OpenRouter also lists generic V4.1 Flash entries at different rates, $0.0219 / $0.6246 and $0.05 / $0.15 per 1M tokens. The source does not say whether those reflect other providers or other configurations. Developers who want a stable target should pin the dated snapshot, since DeepSeek also offers "Flash Latest" aliases that redirect to the newest Flash model.

What this means

The 552B backbone is nearly double the 284B total of V4 Flash 0731, while active parameters are 8B on input and 16B on output versus 13B before. DeepSeek is spending more on stored capacity and less on per-token compute, which suits high-volume agent loops where cost scales with tokens processed.

The cache-memory reduction may matter as much as the headline price. Long-running agents spend much of their budget re-reading accumulated context, so a cache a quarter the size could lower serving costs for those workloads. DeepSeek's cache-read price of $0.003 per 1M tokens points the same way.

Treat the V4 Pro comparison cautiously until DeepSeek publishes benchmark data. The asymmetric input/output activation also makes direct comparison with dense or symmetric MoE models harder, so teams should test on their own agent workloads before migrating.

Related Articles

model release

Ant Group releases Ling-3.2-flash: 560B-parameter multimodal model with 1M-token context

Ant Group released Ling-3.2-flash on November 5, a native multimodal model that accepts text, images, and video. It has 560 billion total parameters, activates 25 billion per inference step, and supports a one-million-token context window.

model release

Thinking Machines Lab launches Inkling 2 Small, a 594B-parameter sparse MoE model, on Tinker

Thinking Machines Lab introduced Inkling 2 Small on November 5, 2026. The sparse Mixture-of-Experts model has 594 billion total parameters and 17 billion active parameters, and is available through Tinker with closed weights at launch.

model release

Google releases Gemini Omni 1.2 Flash video model: 15-second clips up to 4K with native audio

Google released Gemini Omni 1.2 Flash on November 5, 2026, a native multimodal video generation and editing model that succeeds Gemini Omni 1.1 Flash. It generates clips up to 15 seconds with native audio at resolutions from 360p to 4K. OpenRouter lists it at $1.50 input and $9 output per 1M with a 131K context window.

model release

Google releases Gemini Omni 1.1 Flash, a generally available video model with 4K output and native audio

Google released Gemini Omni 1.1 Flash on Nov 4, 2026, a native multimodal video generation and editing model that is the generally available successor to Gemini Omni Flash Preview. It outputs clips from 360p drafts up to 4K with native audio, and is listed on OpenRouter at $1.50 input / $9 output per 1M tokens with a 131K context window.

Comments

Loading...