How DeepSeek V4 fits on a laptop and what does it mean to us?

Ajouté : 2026-05-18

[00:00:00]DeepSeek V4 is a near-frontier AI model.

[00:00:03]On benchmarks, it trades blows with Claude Opus and GPT-5.4.

[00:00:08]There's a smaller version called Flash that's not far behind. It can code, reason, handle a million tokens of context, and it costs basically nothing to use. You can also run the whole thing on a laptop, fully offline. Let me walk through it. Quick context. DeepSeek is the Chinese lab that made R1 last year.

[00:00:29]They just released V4, two versions, Pro, the big one, and Flash, the smaller one. Pro is for data centers. Flash is the one that matters here. Start with the cost because it's the part anyone can use today. Through DeepSeek's API, Flash costs 14 cents per million input tokens, 28 cents per million output.

[00:00:51]Claude Opus, for comparison, is $5 per million input, 25 per million output.

[00:00:57]Opus's output is almost 90 times more expensive. GPT-4.4 is in the same range as Opus. Flash is cheaper than Google's budget model and Anthropic's budget model. It is the cheapest capable model on the market. What that looks like in practice, people on Reddit report spending 5 cents for 4 hours of heavy coding use, running sub-agents in loops.

[00:01:21]Someone used Flash for everything for a week and spent less than a dollar. A Claude code subscription, that's Anthropic's coding tool that top tier, is $200 a month. You could hammer Flash all day, every day, and your monthly bill might be 10 bucks. But, there's something bigger than the price. You don't control the API. Just this week, Anthropic announced it's putting third-party agent usage behind a separate credit meter. Paying subscribers who've been using Claude with coding tools now have a new limit.

[00:01:54]Some companies already blew through their entire year's AI budget. OpenAI can reject your requests if their filters flag something. You're renting intelligence on someone else's terms.

[00:02:06]They change the rules, you adjust. Which brings us to the other thing Flash can do. You can run it locally. Buy the hardware once, that's it. No subscription, no per token billing, no one deciding what requests are allowed.

[00:02:21]Your data stays on your machine.

[00:02:24]For businesses, that means health data, proprietary code, anything with privacy requirements. None of it ever leaves your control.

[00:02:32]And if you're running AI agents, assistants working on your behalf in the background, checking things, writing code, a local model means you can let it run around the clock without watching a meter tick. You own the capability.

[00:02:46]So, it's cheap and you can own it, but is it actually good?

[00:02:51]Deep Seek's own benchmarks show Flash close to Pro on reasoning, and Pro trading blows with Claude Opus 4.6, GPT 5.4, and Gemini 3.1 Pro on coding and math.

[00:03:03]Their numbers, taken with a grain of salt. What matters more is what people actually using it are finding. On the local Llama subreddit, a developer wired Flash into their coding agent and threw it at a large code base. Over 100 tool calls across multiple runs, editing files, running tests, navigating directories, not a single error. Another person AB tested it against Claude and said the quality difference was minimal.

[00:03:29]There are more reports like that. There are also people who had a bad experience, which I'll get to.

[00:03:35]So, how does a model that performs at this level run on consumer hardware?

[00:03:40]There are three engineering tricks stacked together, and they're what make a laptop, even a word you can use here.

[00:03:47]On paper, V4 Flash is 284 billion parameters. That sounds absurd for a personal machine. A dense model that size at standard precision needs well over 500 GB of memory just for the weights, an eight GPU server rack, not a laptop. First trick, mixture of experts, MoE. The idea isn't new, but DeepSeek has been refining their version for years. Think of a mechanic's rolling cabinet with a thousand drawers. That's the 284 billion parameters, total knowledge. But when you go to fix something, you don't open every drawer.

[00:04:26]You open maybe two, the socket wrench and the hex key. The other 998 stay closed. That's how MoE works. For each token, the model only activates about 13 billion parameters worth of expert modules. The other 271 billion sit in memory, available if needed, but they don't burn compute. So, per token, it acts like a 13 billion parameter model, but draws from a knowledge base 20 times larger. DeepSeek's version uses hundreds of small experts instead of a few big ones, more granular than earlier MoE models like Mixtral. Some experts are shared, always on for basic language.

[00:05:07]Others are routed, specialized. And they solve the load balancing problem without the usual quality penalty. That granularity and balance is what keeps the active compute so low. Second trick, attention.

[00:05:23]In a transformer, every token checks in with every other token. The problem, it's quadratic. Double context, quadruple work. At a million tokens, that's a trillion comparisons. Plus the KV cache, the model's short-term memory, also grows with context. At a million tokens, that cache would eat about 80 GB on on own.

[00:05:45]DeepSeek's fix is hybrid attention. Two modes. The first compresses every four tokens into one summary entry. Then a tiny scanner picks only the most relevant ones for full attention.

[00:05:59]Most of the context gets skipped. The second mode is more aggressive, compresses every 128 tokens into one entry.

[00:06:08]A million tokens becomes about 8,000 entries, and 8,000 squared is nothing.

[00:06:14]So, you just run attention on all of them.

[00:06:16]Both modes keep a sliding window of recent tokens uncompressed, so recent context stays sharp. Net effect, at a million tokens, the KV cache drops to about 7% of the previous generation.

[00:06:30]Instead of 80 GB, maybe 5 to 10. So, you've got MoE slashing the active compute, and hybrid attention slashing the memory. Combined, about a tenth of the compute and less than a tenth of the memory of the previous generation at long context.

[00:06:46]But, the model is still physically large. The download from DeepSeek is about 160 GB.

[00:06:53]To get this into memory on a single machine, you need one more piece, quantization.

[00:06:59]Think of it like rounding. Instead of storing every number to 16 decimal places, you round to four or two. You lose a little fidelity, you save a ton of space. That's how you go from a 160 GB download to something that fits in a real computer's memory.

[00:07:16]DeepSeek trained parts of V4 in four-bit precision from the start, unusual. Most models compress after training. They baked it in. At standard four-bit quantization, V4 flash needs a machine with 128 to 192 GB of memory. That's a high-end Mac Studio, around $6,000. Not cheap, but you can order one and put it on your desk. People have pushed further. Salvatore San Filippo, he created Redis, the database that runs behind a huge chunk of the internet, built a custom 2-bit quantized version that fits in about 70 GB. That's within reach of a MacBook Pro with 128 gigs of RAM.

[00:07:59]His engine now supports Nvidia and AMD GPUs, too.

[00:08:04]Nvidia's Digits workstation ships with 128 gigs of unified memory. Consumer adjacent hardware where 128 gigs is becoming normal. So, the path is at 4-bit, a $6,000 Mac Studio. At 2-bit, a high-end MacBook Pro or an Nvidia Digits box. Not impulse buys, but things you can get at retail. And that's how a model with 284 billion parameters ends up running on your desk completely offline. Now, I want to be fair about what this actually is. The real-world reports are a mix. Alongside the people getting great results, there are developers who run into real problems.

[00:08:45]Flash sometimes doesn't follow instructions as reliably as Claude, especially on precision tasks. There are complaints about weird symbols, changing things it shouldn't. One person couldn't get it to follow basic formatting rules for a customer service chatbot. Just refused.

[00:09:03]The pattern, Flash is excellent when the task is narrow and well-defined, sloppy when instructions are open-ended or require sustained discipline. So, it's not a drop-in replacement for Claude Opus. Not yet. But, a lot of people have landed on a workflow. Flash as a worker for well-defined implementation tasks, a smarter model for planning and review.

[00:09:25]In that role, Flash is absurdly cost-effective.

[00:09:29]What's worth watching isn't any one benchmark or price cut. It's the direction. A year and a half ago, running a state-of-the-art AI model locally was a fantasy. Six months ago, it was theoretically possible, but practically absurd. Now, there's a model you can download, quantize, and run on a desk, and it codes, reasons, and handles context well enough that real developers are using it for real work.

[00:09:56]The gaps are real, but the gap between frontier and local just got a lot smaller.

[00:10:02]If the pattern holds, the next version closes it further.

[00:10:06]A few things we touched on could each be their own episode. The open-source AI landscape right now, Kimi, Qwen, GLM, all these labs racing, and why the best models keep coming out for free.

[00:10:18]Why frontier models are so expensive, the token economics behind those prices, and whether they can last. Or how coding agents actually work under the hood.

[00:10:28]What happens when you give AI tools and let it run.

[00:10:32]If one of those is what you want next, tell me.