News

DeepSeek V4 Lite Leaks Under NDA - 1M Context, Natively Multimodal, Codenamed Sealion-Lite

DeepSeek's V4 Lite model has leaked through inference provider testing under strict NDAs, revealing a 1M token context window, native multimodal capabilities, and the internal codename sealion-lite.

DeepSeek V4 Lite Leaks Under NDA - 1M Context, Natively Multimodal, Codenamed Sealion-Lite

DeepSeek's next model is already in the hands of at least one inference provider, and it's reportedly far more capable than anything the company currently serves on its public web interface. The model, internally codenamed sealion-lite, is what appears to be DeepSeek V4 Lite - and the early details suggest DeepSeek is about to change the competitive landscape again.

TL;DR

  • DeepSeek V4 Lite (codename "sealion-lite") is under active testing with at least one inference provider under strict NDA
  • The model features a 1M token context window and is natively multimodal - a first for DeepSeek's flagship line
  • Early testers report it's notably better than the current web/app model (DeepSeek V3.2)
  • Separate leaks show the model producing SVG code that beats Claude Opus 4.6, Gemini 3.1, and DeepSeek 3.2

What We Know

The initial leak comes from API infrastructure sources on X claiming direct knowledge of an inference provider actively testing the model. According to these sources, sealion-lite features a 1 million token context window and is natively multimodal - meaning vision capabilities are baked into the base model rather than bolted on as a separate module. The sources describe performance as "much better" than what DeepSeek currently serves through its web and mobile apps.

The NDA detail is significant. DeepSeek has previously released models with minimal advance notice - DeepSeek V3.2 dropped with almost no pre-launch marketing. But active NDA-protected testing with third-party providers suggests a more deliberate rollout this time, possibly reflecting the model's increased capabilities and the market sensitivity around DeepSeek releases.

The SVG Code Leak

Separately, a series of demonstrations that surfaced through unofficial channels last week showed what appears to be DeepSeek V4 Lite producing remarkably efficient SVG code. The model produced a detailed Xbox controller in 54 lines and a multi-element scene of a pelican riding a bicycle in 42 lines - outputs that, according to internal evaluations, outperform DeepSeek 3.2, Claude Opus 4.6, and Gemini 3.1 in code optimization and logical organization.

These SVG demonstrations are a narrow benchmark, but they suggest something broader: the model's spatial reasoning and structured output capabilities have taken a significant step forward.

1M Context and Architecture

The 1M token context window aligns with what DeepSeek has been building toward publicly. On February 11, the company silently upgraded its production model's context window from 128K to 1 million tokens. Community testing showed over 60% accuracy on needle-in-a-haystack retrieval at the full 1M length.

SpecDeepSeek V3.2DeepSeek V4 Lite (Leaked)
Context Window128K (recently 1M)1M (native)
MultimodalNo (text only)Yes (native)
Estimated Parameters~685B MoE~200B (unconfirmed)
Code OptimizationBaselineOutperforms V3.2 in SVG tests
StatusProductionNDA testing

Chinese tech outlet 36kr reports the model has roughly 200 billion parameters and doesn't use the Engram conditional memory system that DeepSeek co-developed with Peking University. If accurate, this positions V4 Lite as a lighter, faster variant ahead of the full V4 flagship - which is rumored to exceed 1 trillion parameters.

Native Multimodal Is the Real Story

DeepSeek's entire V3 generation has been text-only. Every previous DeepSeek model that could process images (like DeepSeek-VL) was a separate model line with a different architecture. The claim that V4 Lite is natively multimodal represents a fundamental shift.

This matters because native multimodality usually means the model was trained from the ground up to understand both text and images in a unified representation space. Bolted-on vision (the approach most models used circa 2024-2025) tends to create a gap between text and vision performance. Native multimodal training, as demonstrated by Gemini 3.1 Pro and GPT-5, closes that gap.

If DeepSeek can deliver native multimodal capabilities in an open-weight model - something no Chinese lab has done at the frontier level - the consequences for the open-source vs. proprietary AI debate are substantial.

What It Does Not Tell You

Several important caveats apply.

First, none of these claims have been officially confirmed by DeepSeek. The company has said nothing publicly about V4 Lite, and the "sealion-lite" codename hasn't appeared in any official documentation or code repository.

Second, the SVG demonstrations, while impressive, test a narrow capability. Creating clean vector graphics code is useful but doesn't tell us how the model performs on the benchmarks that matter most - reasoning, multi-step coding tasks, and long-context synthesis.

Third, the ~200B parameter estimate, if accurate, places this model well below the full V4 flagship in raw capacity. "Lite" presumably means tradeoffs somewhere, and we don't yet know where those tradeoffs land.

Finally, CNBC has already warned that a DeepSeek V4 launch could trigger Nasdaq turbulence similar to the 3% drop that followed V3's release. The market is watching this closely, and the NDA-protected testing phase suggests DeepSeek knows the stakes.


The pattern is familiar: DeepSeek builds quietly, leaks surface, the industry braces, and then the model drops with little ceremony. What's different this time is the scope of ambition. A natively multimodal, million-token-context open model from a Chinese lab that reportedly beats current frontiers on at least some tasks - if even half of the leaked claims hold up, V4 Lite will be the most significant open-source model release of 2026 so far.

Sources:

DeepSeek V4 Lite Leaks Under NDA - 1M Context, Natively Multimodal, Codenamed Sealion-Lite
About the author Senior AI Editor & Investigative Journalist

Elena is a technology journalist with over eight years of experience covering artificial intelligence, machine learning, and the startup ecosystem.