DeepSeek V4 Flash (0731) vs Kimi K3: The 2.8 Trillion Parameter Lie

Four days apart, two labs shipped open weights. One of them is 10 times bigger than the other. The big one wins nearly every benchmark the two share and you cannot run it. Not on hardware you can buy. Kimmy K3 against DeepSeek V4 Flash the 0731 build. Who wins and what the win actually costs? Plus the strange part. The two scoreboards that prove it disagree about a model neither lab built. Start with mass. Hugging Face reports Kimmy K3 at 2.78 trillion parameters. V4 flash 284 billion, 10 times smaller, near enough. And both shipped as open weights. Neither fires all of it. K3 lights up 104 billion per token. Flash lights up 13. Both leave about 96% of themselves asleep. Same shape, one order of magnitude apart. Artificial analysis ran both on one harness. K3 scores 57 first out of 101. Flash scores 50, third of 101. So the ranking is settled and then you read the price list. K3 cost $3 in, 15 out per million tokens. Flash charges 14 cents in, 28 out, 21 times on input, 54 on output for 14% more measured intelligence. And locally, the largest Mac Studio Apple sells 512 GB cannot load K3's smallest build. Flash runs at 34 tokens a second on a MacBook Pro. So, the benchmark winner and the machine winner are two different models. The chronology is worth holding on to. Moonshot opened the K3 API on July 16th and released the weights 11 days later on the 27th. Deepseek's official flash build landed on the 31st at 7 in the morning. UTC Unsloth had a quantized version up by 24 minutes past 12, 4 hours and 54 minutes later. Kimmy K3 is Moonshot's first 3 trillion class open model and the architecture is new. 93 layers, 69 of them running Kimmy delta attention instead of standard attention. 896 experts per layer, 16 of which fire for any given token, plus a 400 million parameter vision encoder called Moonvit. Because K3 reads images natively, [snorts] that vision encoder matters more than it sounds. K3 takes text and images in the same model, holds a million tokens of context, and can emit a million tokens back. V4 Flash is text only, and DeepSeek recommends capping output at 384,000. If your work involves screenshots or diagrams, the comparison ends right here, and you can stop reading benchmark tables. Deepseek's answer is the opposite bet. 284 billion parameters, one shared expert plus 256 routed ones per layer, and only the top six fire per token, 13 billion active. Built to be small, exactly where the serving bill is calculated. The 0731 build changed no architecture at all. Same parameter count, same layout as the April preview. Every gain came from redoing the post training on deep. That took the score from 7.3 to 54.4. Same weights, shape, taught differently. And it worked well enough to embarrass its own flagship. On all nine benchmarks Deepseek published, the 0731 flash build beats Deepseek V4 Pro, a 1.6 trillion parameter model with 49 billion active, 13 billion active parameters beating 49. That is the entire Deepseek thesis printed in their own table. It also ships with a speculative decoding module bolted on called DSpark. That is why HuggingFace reports 304 billion parameters for a 284 billion model. About 20 billion of draft model rides along guessing ahead so the big model can confirm several tokens at once instead of one. One more thing the two share and it's the reason either is runnable at all. Both were trained in low precision on purpose. K3 stores 98% of its weights 4-bit packed. Flash stores 97% the same way. Neither is a big model squeezed down afterwards. They were born compressed. Now the fight on Terminal Bench 2.1. Moonshot's model card puts K3 at 88.3. Deep Seek's card puts Flash at 82.7, 5 1/2 points. Case closed? No. Those two numbers were produced by two different companies grading their own homework on two different harnesses. Deepseek's harness is described in its own footnotes as forthcoming. A polite way of saying nobody outside Deepseek has it. [snorts] But there is a way to check and it's sitting in plain sight in both documents. Each card benchmarks its model against the same two outsiders, GLM 5.2 and Claude Opus 4.8. Neither lab made either of those, so the numbers should match. They don't. On Terminal Bench, Moonshot scores GLM 5.2 at 82.7. Deepseek scores the same model on the same benchmark at 81.0. On Opus, Moonshot says 84.6. Deepseek says 85.0. Same models, different rulers, so measure the gap in a currency. They both quote anchor on Opus. Moonshot puts K3 3.7 points above it. Deepseek puts Flash 2.3 points below it. That makes the gap six points, not 5 1/2. Now anchor on GLM instead. K3 sits 5.6 above it in Moonshot's table. Flash sits 1.7 above it in Deep Seeks. That gap is 3.9 points. Same two models, same benchmark, two defensible answers, 3.9 or six. The honest version is that Kimmy K3 is somewhere around four to six points ahead on Terminal Bench. And any source quoting you one decimal place is selling a certainty that nobody currently has, which is why the only comparison worth much is the one where a third party ran both. Artificial analysis did on the six benchmarks they publish head-to-head. Browse comp, GDP, Val, GPQA, humanity's last exam, MCP Atlas, Tulle Athalon, Kimmy K3 wins all six. The widest is Tulle Athalon, 73.2 against 47.8. A developer named Filipe ran both on the same prompt in the same harness. One shot a landing page for a $10,000 concept object. His read was that K3 won on polish. Better typography hierarchy, more confident negative space. Flash got the structure, the copy, and the layout logic right and lost on finish. That's what four to six points looks like to a human eye. So K3 is better. That was never really in doubt. The question is, what better costs? And the sticker is brutal. 21 times more for input, 54 times more for output. Except sticker price is a trap. Running that same intelligence index, Flash generated 210 million output tokens. K3 needed 130 million. Flash thinks out loud considerably more. So, price the actual work. 130 million K3 tokens at $15 is $1950. 210 million Flash tokens at 28 is $58.80. 33 times, not 54. Verbosity ate a third of Flash's advantage. It is still 33 times. Put that against 14% more measured intelligence and you have the whole economic argument in one line. You are not buying a better model. You are buying the top 14% of one at roughly 30 times the rate and even the seven-point gap deserves an asterisk. Bench LM puts K3 at 79.9 against Flash's 59.9 then notes that their 90% confidence intervals overlap. A lead you cannot separate from noise is a lead worth spending carefully. Speed goes the same direction. Artificial analysis clocks K3 at 35 output tokens a second with a 3.9 second wait for the first one. Testers put flash between 112 and 118, three times faster on the model that cost a fraction. There's one more line on both price lists. A cash hit drops K3 from $3 to 30, a 10-fold discount. It drops Flash from 14 cents to a third of a cent 50fold. If your workload repeats a long prompt, which most agent workloads do, the gap widens instead of closing. Now the part that decides it for most people watching. Can you run these? Unsloth quantized K3 down to a single bit. The file is 594 GB. You need roughly 610 GB of memory to load it. The largest Mac Studio Apple ships has 512. It cannot open the smallest possible version of this model. The realistic floor for K3 is an 8GPU node with tensor parallelism, which is not a desk, it's a rack. V4 flash at 1 bit is 82 1/2 GB. Unsloth recommends 110 and says Sensible starts at 128. That's a machine you can buy a MacBook Pro or 196 GB workstation card and people are running it. Salvatore Sanfalippo Anti-Rez, the man who wrote Reddus, built a dedicated engine for it in C called DS4 on 128 GB M5 Max at 2bit, 87 tokens a second prefill, 34 tokens a second generating. on an M3 Max 26. There's a line in that repository I keep thinking about. Anti-res discloses that DS4 was built with strong assistance from GPT 5.5, 5.6, and Claude Fable with humans leading the ideas and the debugging. The engine that sets an open model free was written with closed ones. Last difference, and for a business, it might be the only one that matters. Deepseek ships V4 flash under MIT. Do what you like. Moonshot ships K3 under the Kimmy K3 license which artificial analysis labels commercial use restricted. The specifics. If you offer inference or fine-tuning as a service and your revenue crosses $20 million over any 12 months, you need a separate agreement with Moonshot. Past 100 million monthly users, you must display the words Kimmy K3 in your interface. That isn't MIT. And calling both of these open flattens a real difference. Here's a small signal, and I'll give you the half that argues against me, too. Unslaught quantized K3 has been up 5 days with 238 likes. The quantized flash build has been up one day and already has 254. Downloads still favor K3 41,000 to 4,000. So, this is early, but the runnable build is pulling ahead on the metric that cost a click. So the verdict, if you're calling an API and the task is hard, vision, deep agentic work, long research, Kimmy K3 is the better model, and it earns its price on the jobs that need it. For everything else, and for anything you host yourself, V4 Flash is the answer because it's the one that fits. Route by task, not by leaderboard. Default to flash, escalate to K3 when a job actually earns it. and treat every vendor benchmark table as a claim until somebody who didn't build the model reruns it. This week handed us the proof.