284 Billion Parameters, 13 Billion Moving
April 24, 2026. DeepSeek drops V4-Flash, a 284-billion-parameter mixture-of-experts model under an MIT licence, and anybody running language models on hardware they own ought to pay attention. I've spent a week running it hard. What I'm seeing rewrites the arithmetic of local deployment, and I don't say that about many releases.
Bottom line up front: V4-Flash is the first model I've tested where a machine sized for a single organization is genuinely competitive with cloud inference on real production work. That claim only lands if you understand what a mixture-of-experts architecture does, so that's where I'll start.
284 Billion Parameters. 13 Billion Active. Here's Why That Matters.
V4-Flash is sparse by design. The raw parameter count hits 284 billion, but in an MoE model not every parameter fires on every token. V4-Flash activates roughly 13 billion parameters per token, about 4.6 percent of its total weight.
Memory bandwidth's the real bottleneck in inference, not raw compute. Dense models like Llama 3.3 70B haul all 70 billion parameters across the bus on every single pass. V4-Flash only moves the experts it lit up. That's a different machine's worth of throughput out of identical silicon, and it's measurable rather than theoretical.
The VRAM Reality and What FP8 Does
V4-Flash at full FP16 precision wants roughly 282 gigabytes for the weights alone. That's the honest number and it rules out a lot of hardware. No amount of enthusiasm changes it.
FP8's what closes the gap. At FP8 the weights land around 170 gigabytes, so 192GB of VRAM holds the whole model with about 22GB left over for the KV cache. That's the part that changed this year. You're running the full FP8 model rather than a heavily compressed INT4 build, and the difference in output quality isn't subtle.
Here's the throughput arithmetic, and it's the same arithmetic for any card you're weighing. V4-Flash activates roughly 13 billion parameters per token, and at FP8 those active weights come to about 13GB. The RTX PRO 6000 Blackwell Server Edition moves 1,597 GB/s, so the memory-bandwidth roofline sits near 1,597 / 13, about 123 tokens per second in theory. Real single-stream inference lands in the 60 to 90 tokens per second range once you've paid for the KV cache, attention overhead, and expert routing. Those are estimates grounded in the bandwidth math, not vendor marketing, and I'd rather hand you the division than the conclusion.
FP8's a precision step down from FP16, but it's a far lighter touch than the aggressive INT4 quantization a tighter-VRAM machine used to force on you. For summarization, classification, structured extraction, and code generation, the quality holds. And the speed you get back for it isn't nothing.
None of that's a property of any particular build. It's a property of memory bandwidth and active weight bytes, which is why the numbers travel. Point them at whatever silicon you're considering and they'll tell you roughly what it'll do before anybody quotes you a thing.
One Million Tokens. That's Not Marketing Language.
V4-Flash carries a one-million-token context window. I've run 500K-token contexts and watched the model hold coherence across the whole span. A million tokens is roughly 750,000 words. A 1,500-page document. An entire codebase. A multi-year email archive dropped into a single prompt.
Most production work lives in the 4K to 32K range, so the full million sits largely untapped right now, and I'll say that rather than pretend otherwise. But the implication for an organization is real: you load whole documents, whole codebases, whole conversation histories, and the chunking strategies and summarization workarounds go in the bin. A law firm drops a 50-page contract in whole. A clinic loads a patient's full chart. A tribal government loads an entire grant application with every appendix attached and asks the model to check it against program requirements.
Here's where local and cloud diverge in a way no benchmark score captures. Cloud providers cap context windows or charge a premium for long-context use, because that's their meter and it's supposed to run. Running V4-Flash on your own hardware means the million-token window costs you nothing extra per use. The model lives on your GPUs. Context is a function of your VRAM, not somebody else's billing tier.
What Full FP16 Costs You in Memory
Running V4-Flash at FP16 rather than FP8 is a memory question with a clean answer: about 282GB of weights, so 384GB of VRAM covers it with room for the KV cache. Aggregate bandwidth scales with the memory you add, too, which shows up in sustained throughput and in how many people can be talking to the thing at once. Early testing puts single-user generation at 90 to 140 tokens per second at that capacity, and you're never trading precision for fit.
Whether any of that's worth buying isn't a question I can answer from here. The tidy decision tree people publish for this ("if X then buy Y") is mostly a sales instrument wearing a lab coat. What decides it is your concurrency, your document lengths, and which department's workflow turns out to be the real load, and I don't know any of those until I've sat next to your people and watched them work. Neither do you, entirely. That's just how organizations work, and no insult is intended.
The Sovereignty Argument
This part rarely gets quantified. It should.
When V4-Flash runs on hardware you own, your data doesn't leave your premises. No API call to DeepSeek's servers. No metadata logs. No audit trail living in somebody else's system. The model's MIT-licensed, which means you own the weights outright: modify them, redistribute them under MIT terms, run them for as long as you like without a licensing fee. That's yours. Nobody can revoke it in a terms-of-service update.
Cloud API providers can't offer that guarantee, and to be fair to them they've never claimed to. OpenAI, Anthropic, DeepSeek's own cloud product, all of them reserve some right to use submitted queries for model improvement in their terms. Many explicitly retain data. Even a contractual opt-out doesn't change the fact that your data transited their infrastructure, and for a regulated organization that transit is the liability, not the retention.
So local deployment is a control conversation, not just a speed or cost one. If your threat model includes data exfiltration, inference interception, or a data-residency rule with teeth, local's the only option. V4-Flash is what makes it the only option that also performs.
A medical practice running clinical documentation through DeepSeek's cloud API is sending patient data to servers it doesn't control. Running the same model locally means the patient data never leaves the building. The model file sits on an NVMe drive in your rack. Inference happens on your GPUs. The response comes back over your own network. DeepSeek the organization is involved in exactly zero steps, which is the entire point.
What V4-Flash Replaces
Before April 2026, local deployment meant picking between size and speed. Llama 3.3 70B is quick but caps at a 128K context window and doesn't carry the reasoning capacity of a 284B-parameter system. V3 had 671 billion parameters with 37 billion active, but wanted 350GB or more of VRAM at full precision, which priced it out of most rooms.
V4-Flash fills a gap that didn't exist until now. The knowledge capacity of a 284B-parameter system. The inference speed of a 13B one. A context window ten times larger than most of what it competes with. It sits alongside Llama as a genuinely different capability class, not a marginal upgrade, and that's the first time I've written that sentence about an open release.
That middle ground is where production work lives.