DeepSeek is China's most advanced AI Lab. DeepSeek-V4 was recently released on the market with free source code and free LLM model weights a few weeks back on GitHub for all to download. That allowed the corporations of the world to run DeepSeek-V4 on their own hardware. DeepSeek-V4 can also be run using the hardware owned by DeepSeek and is about 20 - 50 times cheaper to use than running on the advanced LLM models and hardware offered by the American Frontier AI Labs. DeepSeek-V4 has 1.6 trillion parameters, but DeepSeek figured out a very clever way to only have 49 billion parameters active at any one time by just turning on the "neurons" needed at any one instance. That substantially reduced the hardware requirements needed to run DeepSeek-V4.
Then, on June 27, 2026, DeepSeek added a new architectural enhancement to DeepSeek-V4 called DSpark to very dramatically speed up "inference" runs on the DeepSeek-V4 model. Inference is when a model runs and actually takes in money from end users. People feed the LLM some input tokens in a prompt, and the LLM then spits out the answer as a series of output tokens. Remember, a token is about 2/3 of a word. Customers then pay the LLM provider for the number of input and output tokens. Figuring out the 1.6 trillion parameter weights in an LLM is called "training", and that training costs lots of money and electricity to conduct. Once the trillions of parameter weights have been determined, they do not change, and the training costs then end. Now, running the trained LLM is what makes the real money and is called "inference".
However, all the Frontier AI Labs around the world have now discovered that running very large LLM models in Production has now become their largest bottleneck to making money and justifying to investors the trillions of dollars needed to fund Advanced AI and all that it requires in hardware and software. You see, it takes a good deal of hardware to run inference on a 1.6 trillion-parameter LLM, and that hardware can easily get overwhelmed by the number of input requests coming into an AI datacenter in real time. Think of a 1.6 trillion-parameter LLM as a huge prokaryotic cell that has to contain everything needed to keep the cell alive and running. To overcome this industry-wide problem, DeepSeek came up with this new idea called DSpark. We keep the huge 1.6 trillion-parameter LLM prokaryotic cell but then change its internal architecture by having it use many embedded "helper" LLM models that are much smaller and have far fewer parameters. These small "helper" models are like mitochondria. They run much faster and with much less hardware than the BIG 1.6 trillion-parameter model. Their job is to run quickly and then "guess" the next 10 or so output tokens. Those output tokens are then sent to the BIG 1.6 trillion parameter model to be checked. If the BIG model likes the tokens, it keeps them; otherwise, it truncates the string of 10 tokens when it finds the first token that it does not like. Having the BIG model only check the tokens is about 6 times faster than having the BIG model figure out the next output token on its own. This makes DeepSeek-V4 DSpark run 6 times faster, and it can run 6 times the load on the same hardware.
Here is a nice YouTube video that explains it all:
DeepSeek’s New AI Breakthrough Just Broke AI’s Limits
https://www.youtube.com/watch?v=V7GBRPf7Zy8
Here is a link to the DeepSeek paper about DSpark:
DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation
https://github.com/deepseek-ai/DeepSpec/blob/main/DSpark_paper.pdf
You have to click on the "..." on the upper-right of that page to download the paper.
All of this DeepSeek software is free to download from GitHub! Again, it seems that China is trying to win the Global AI Race by taking the profit motive away from the American Frontier AI Labs.
So What is a Eukaryotic Architecture in Biology and why was it so Important in the Evolution of Carbon-Based Life on our Planet?
Softwarephysics has long advocated for a biological approach to the generation and running of software. On that basis, one might question what the big deal is about DeepSeek's new DSpark Eukaryotic Architecture. Well, the reason for being excited about this innovative advance in AI is that it seems to be recapitulating one of the most dramatic advances in the evolution of carbon-based life on the Earth. For more on that see The Rise of Complexity in Living Things and Software. In fact, the eukaryotic architectural change to carbon-based life may represent a very significant Filter in the origin of Intelligence for any galaxy in our Universe. Here is a very interesting SpaceTime YouTube video on the subject that suggested it as a possible Filter that may have been very difficult to overcome for most simple prokaryotic carbon-based life in our galaxy:
Is There A Simple Solution To The Fermi Paradox?
https://www.youtube.com/watch?v=abvzkSJEhKk
The above video discusses the huge complexity differences between the simple prokaryotic cell structure of bacteria and archaea and the vastly more complicated eukaryotic cell architecture that is common to all higher forms of carbon-based life on the planet. The video explains the commonly held thought that an ancient prokaryotic bacterium that had developed a tolerance to oxygen and had actually developed a way to metabolize organic molecules using oxygen as an oxidizing agent had invaded a much larger prokaryotic archaea cell in a parasitic manner and then took up residence within it. These two cell types then developed a symbiotic relationship in which the parasitic bacterium finally became a mitochondrial organelle that supplied vast amounts of free energy for the host archaean cell.
Figure 1 – The prokaryotic cell architecture of the bacteria and archaea is very simple and designed for rapid replication. Prokaryotic cells do not have a nucleus enclosing their DNA. Eukaryotic cells, on the other hand, store their DNA on chromosomes that are isolated in a cellular nucleus. Eukaryotic cells also have a very complex internal structure with a large number of organelles, or subroutine functions, that compartmentalize the functions of life within the eukaryotic cells.
Figure 2 – Not only are eukaryotic cells much more complicated than prokaryotic cells, but they are also HUGE!
The question is if simple prokaryotic cells arose nearly four billion years ago, just after the Earth's crust solidified, why did it then take several billion years for the more complex eukaryotic cell architecture to arise? Perhaps this was the only time for this to ever happen in our galaxy. That indeed would be some kind of Filter!
Figure 3 - Mitochondria are like little parasitic bacteria that at one time invaded some prokaryotic archaeon cells about 2 billion years ago, and went on to form a strong parasitic/symbiotic relationship with their archaeon hosts. Mitochondria have their own genes stored on bacterial DNA in a large loop, just like all other bacteria. Each eukaryotic cell contains several hundred mitochondria, which self-replicate before the eukaryotic cell divides. Half of the mitochondria go into each daughter cell after a division of the eukaryotic cell. The eukaryotic host cell provides the mitochondria with a source of food, and the mitochondria then metabolize that food using the Krebs cycle and an electron transport chain to pump H+ protons uphill to the outside of their internal membranes. As the H+ protons fall back down they release stored energy to turn ADP into ATP for later use as a fuel.
Figure 4 - The new DSpark architecture of DeepSeek-V4 operates in a very similar manner to the large number of mitochondria found in eukaryotic cells. Given an input prompt of tokens ABC , the model executes one step to generate the next token D , which serves as the anchor for the drafting phase. Using D as the input, DSpark employs a heavy parallel backbone and a lightweight sequential head to generate draft tokens EFGH along with their corresponding confidence scores 1 – 4. The Hardware-Aware Prefix Scheduler then evaluates these scores to retain the prefix EFG and drop the low-confidence token H . Finally, the target model verifies the scheduled prefix in parallel. As illustrated, E and F are accepted while G is rejected, prompting the model to generate a corrected token G* to complete the current round.Click to enlarge.
By using smaller and less-complicated "helper" LLM models as virtual mitochondria, DeepSeek-V4 DSpark is able to speed up the inference of model input prompts by a factor of six and allow current AI hardware configurations to handle up to 6 times the load without a hardware upgrade. This again highlights the advantages of taking a biological approach to advance the effectiveness of both hardware and software.
Comments are welcome at scj33345@gmail.com.
To see all posts on softwarephysics in reverse order, go to:
https://softwarephysics.blogspot.com/.
Regards,
Steve Johnston
