DeepSeek Releases V4 Flash and V4 Pro with 1M Token Context

DeepSeek released two new open-source models on May 12: V4 Flash, a non-reasoning variant, and V4 Pro, both available simultaneously through four commercial inference providers. Each supports a 1 million token context window and arrives priced aggressively relative to comparable models on the market.

The pricing structure is worth examining in detail. According to Price Per Token, V4 Flash runs at $0.14 to $0.20 per million input tokens depending on the provider, with output at $0.28 to $0.40. V4 Pro sits considerably higher: $1.70 to $2.40 input, $3.40 to $4.80 output. Both models are available through Alibaba, Venice, AtlasCloud, and Parasail, with each provider setting its own rate within that band.

Multi-provider availability on day one is a deliberate operational choice. Single-endpoint launches typically create rate-limit bottlenecks when engineering teams rush to evaluate a new model in parallel. Distributing across four providers from launch sidesteps that friction, even when pricing varies by a small margin between hosts.

The 1M context story

A 1 million token window puts both models in range for workloads that previously required chunking: long-document retrieval, extended code analysis, multi-file reasoning, and full-conversation history injection. For engineers building artificial intelligence pipelines, the practical question is whether V4 Flash handles these tasks at sub-$0.20 input costs, or whether the depth of reasoning in complex queries still pushes teams toward the Pro tier.

Flash ships as a non-reasoning model, a distinction worth unpacking. Reasoning models run extended chain-of-thought passes during inference, which inflates both latency and per-token cost considerably. Flash omits that step, trading reasoning depth for speed and price. That trade-off suits retrieval-augmented generation pipelines, high-volume classification, and summarization tasks where output throughput matters more than multi-step deliberation.

LLM Stats shows that DeepSeek also released V4-Flash-Max and V4-Pro-Max variants in late April, which means the V4 family is already a tiered offering rather than a simple two-model lineup. The May additions extend the stack downward with faster, cheaper options alongside the Pro line, giving teams finer-grained cost-performance tradeoffs without switching providers or model families entirely.

Open weights and the distribution picture

Both models ship as open weights, following DeepSeek's established practice of releasing model parameters publicly alongside commercial API access. This matters for organizations with strict data-residency or privacy requirements who cannot route inference through third-party endpoints. Open weights mean those teams can self-host, while commercial API availability lets smaller teams skip the infrastructure overhead entirely.

AI Release Tracker currently catalogs 156 tracked frontier models, and the release cadence is accelerating. DeepSeek's approach of dropping multiple variants across multiple providers in a single event reflects where the field has moved: releases are no longer singular announcements but rolling updates within a family, and the competitive surface now spans price tiers, hosting flexibility, and context length simultaneously.

The cost gap between the Flash and Pro tiers deserves careful analysis before committing to an architecture. Price Per Token data makes the differential concrete: V4 Flash at $0.14 per million tokens on AtlasCloud versus V4 Pro at $2.40 on Alibaba is a roughly 17x cost multiplier on input alone. Teams that default to the more capable model without profiling their actual task distribution will consistently overpay for reasoning capacity that most of their queries do not require.

Longer term, the 1 million token context window is beginning to look like a commodity feature rather than a differentiator. DeepSeek's open-source delivery of this capability changes the calculation for teams that need long-context artificial intelligence inference without proprietary lock-in. The meaningful competition now runs on quality at Flash-tier price points, not on raw context length.

Heading into the second half of 2026, the question sharpens: if non-reasoning models at $0.14 per million input tokens can cover the majority of enterprise workloads, the business case for reasoning-first architectures shrinks to a narrower set of genuinely hard tasks. Whether that threshold holds as task complexity scales is the benchmark that matters now.

FAQ

What is the difference between DeepSeek V4 Flash and V4 Pro?
V4 Flash is a non-reasoning model built for speed and low cost, priced at $0.14 to $0.20 per million input tokens depending on the provider. V4 Pro includes reasoning capability at $1.70 to $2.40 per million input, making it better suited for complex multi-step tasks where chain-of-thought inference adds measurable value.

Which providers offer DeepSeek V4 as of May 2026?
Both models launched simultaneously through Alibaba, Venice, AtlasCloud, and Parasail, with minor pricing differences between each host.

What does a 1 million token context window enable in practice?
It allows the model to process very long documents, large codebases, or extended conversation histories in a single inference call, eliminating the need to split input into smaller chunks and reconcile partial outputs afterward.

Is DeepSeek V4 available for self-hosting?
Yes. Both V4 Flash and V4 Pro are released as open-weight models, meaning organizations can download and self-host the parameters in addition to using any of the four commercial inference API providers.