Thirteen months ago, DeepSeek V3 wiped $600 billion off Nvidia's market cap in a single session and forced every AI lab in Silicon Valley to explain why their models cost ten times more. Now V4 is imminent -- and the technical details leaking out suggest DeepSeek is not just iterating. It is rearchitecting.
Here is what we know about the four innovations under the hood, what an early V4 Lite leak reveals about the model family's practical capabilities, and why this matters for anyone building on top of AI tooling today.
MODEL1: Tiered Memory That Cuts GPU Costs by 40%
The headline architecture change in V4 is called MODEL1. Instead of cramming everything into expensive GPU VRAM the way current transformer models do, MODEL1 introduces tiered KV cache storage that intelligently distributes data across GPU, CPU RAM, and disk.
Frequently accessed tokens stay in fast GPU memory. Everything else gets pushed to cheaper tiers. The result: a reported 40% reduction in memory usage without compressing or discarding information. For context windows beyond one million tokens, this is the difference between "theoretically possible" and "actually affordable to serve."
The practical implication is straightforward. If you are running local models or paying per-token API costs, a 40% memory reduction translates directly into lower infrastructure bills -- or the ability to handle longer, more complex prompts on the same hardware you already own.
Sparse FP8: 1.8x Faster Inference With a Precision Trick
V4's second architectural bet is sparse FP8 decoding. The idea: not every token in a response requires the same mathematical precision. Critical tokens -- the ones that carry meaning, logic, and factual accuracy -- get calculated at full precision. Less important tokens (filler words, formatting, boilerplate syntax) run at lower-precision FP8.
DeepSeek's published research suggests this hybrid approach achieves a 1.8x inference speedup with minimal accuracy loss. For high-traffic applications -- customer-facing chatbots, API-heavy developer tools, batch document processing -- that is a meaningful reduction in both latency and cost per query.
This is a direct continuation of work from DeepSeek's Sparse Attention paper (DSA), which already demonstrated that you can cut long-context costs by roughly 50% if you are willing to be selective about where you spend compute.
Engram Memory: Separating What the Model Knows From How It Thinks
The third innovation, Engram memory modules, might be the most interesting from an architecture standpoint. Rather than forcing the model to "re-derive" basic programming knowledge on every inference pass -- wasting GPU cycles remembering what a for loop is -- Engram separates static knowledge (API signatures, language syntax, factual databases) from dynamic reasoning (the actual problem-solving logic).
Static knowledge sits in CPU RAM. Dynamic reasoning runs on the GPU. The result, per Kilo Code's analysis, is an estimated 30% VRAM reduction for local development workloads. The GPU gets freed up for the hard parts: understanding your codebase, reasoning about your specific architecture, solving the actual problem you asked about.
For developers running models locally on consumer hardware -- an increasingly popular setup as open-source AI democratizes access -- this could meaningfully expand what is possible on a single GPU.
V4 Lite Already Leaked, and It Is Generating SVGs
While the full V4 model has not officially launched, a lighter variant called V4 Lite surfaced through unofficial channels this past weekend. The standout capability: generating production-quality Scalable Vector Graphics code with remarkable efficiency.
In one demonstration, V4 Lite produced a detailed Xbox controller rendering in 54 lines of SVG code. A multi-element scene required just 42 lines. Internal evaluations suggest V4 Lite generates more optimized vector code than DeepSeek 3.2, Claude Opus 4.6, and Gemini 3.1.
This matters beyond the novelty factor. SVG generation is a proxy for spatial reasoning and structured code output -- skills that transfer directly to UI component generation, diagram creation, and design automation. If V4 Lite handles SVGs this well in a leaked, unoptimized state, the full V4 model's capabilities across coding tasks could be substantial. Leaked (but unverified) benchmarks claim 90% on HumanEval and over 80% on SWE-bench.
The Market Context: $650 Billion in AI Spending Meets a $6 Million Challenger
Wall Street is watching closely. Bridgewater estimates Big Tech will spend roughly $650 billion on AI infrastructure in 2026 alone. Meta just signed a 6-gigawatt GPU deal with AMD. OpenAI has restructured its compute spending to $600 billion through 2030.
DeepSeek built V3 for a reported $6 million.
That asymmetry is what makes each DeepSeek release a market event. If V4 matches or exceeds current-generation models from Anthropic and OpenAI at a fraction of the cost, the entire capital allocation thesis behind Big Tech's AI spending comes under pressure. CNBC is already warning that the Nasdaq could replay its 3% single-day drop from January 2025.
But there is a flip side. DeepSeek's own market share in open-source models fell from 50% to under 25% over the past year as competition from Qwen, Kimi K2, and InternLM intensified. V4 is not just a product launch -- it is a bid to reclaim technical leadership in a market that moved fast while DeepSeek was building.
The Pricing Math Behind Open-Weight vs. Hosted
Leaked API pricing for V4 suggests roughly $0.27 per million tokens. Compare that to hosted models from US labs, where frontier-tier pricing can run 20-40x higher depending on the provider and model class.
For teams already comfortable with open-weight deployment -- running models on their own infrastructure or through providers like Together AI and Fireworks -- V4 could push the cost-per-query for production AI workloads into genuinely trivial territory. We are approaching a world where the model itself is effectively free, and the real cost is the engineering effort to integrate it well.
That shift rewards teams that invest in AI integration architecture rather than just subscribing to the most expensive API.
What To Watch This Week
The V4 launch window is measured in days, not weeks. Three things will determine whether this is another DeepSeek market shock or a more measured evolution:
-
Independent benchmark verification. The leaked numbers are impressive but unconfirmed. Watch for results from LMSYS Chatbot Arena and independent evaluators, not just DeepSeek's own claims.
-
Open-weight availability. DeepSeek's power move has always been open-sourcing their models. If V4 ships with open weights, expect every major inference provider to have it running within 48 hours.
-
The Lite variant's role. A lighter, specialized model alongside a full-power flagship suggests DeepSeek is thinking about deployment tiers -- matching the right model to the right task rather than forcing one giant model to do everything.
The AI model landscape is moving into a phase where architectural efficiency matters more than raw parameter counts. DeepSeek V4 is the sharpest test yet of whether that thesis holds up in production.