There is a time-consuming part, though, that is performed manually by our (human) team: implement the logic of the model in C++ and assembly code in a super-optimized way, co-designed for each specific hardware card.

This can take months.

We hope to accelerate the process with AI agents, but we're not there yet.

This is our main use case.

It should be possible to apply the same methodology (digging deep into the hardware details to understand all its little characteristics, and rethinking the inference stack around that).

You can ask it to implement an algorithm; we provide suggested prompts you can test.

Also, this tech preview is really about the speed of the inference engine (not the model itself) so I'm glad you got 3.4k tok/s!

Our view is that MTP / speculative decoding could help getting a X multiplier (X = 2 to 6) on the tokens per second speed we currently achieve.

We are a bit greedy, we want to stack optimizations on top of each other to get the maximum speed possible.

It involves additional compute to verify the predicted tokens during the forward pass (it's like a small batch), which should be totally doable for dense models, and will be more tricky for MoEs because it could mean activating more experts and thus more active parameters.

About model performance, we plan to support the latest frontier models (this tech preview is about the speed of the engine)

DTP is something we built for our roadmap in order to get to extremely high speeds (like 10k+ tokens/s). When the budget is under 10 µs per layer, any little overhead matters.

For 1k to 5k tokens/s, regular TP still works because we are able to optimize the inter-GPU all-reduce collectives at under 3 µs, which allows to continue streaming model weights in shared memory, registers and caches while GPUs exchange data.

The last section of the article lays out the scaling laws that apply when porting this approach to another model. In a nutshell, DeepSeek V4 Pro with 49B active params is close to the upper bound.

Also worth noting that our results are currently for standard datacenter GPUs. On consumer hardware, though the same low-level optimization approach applies, the bandwidth limitations will cap the achievable speed.

To answer your questions:

- yes, we rewrite the whole model code (while keeping the same logic) in CUDA/HIP and assembly, in order to optimize by hand for each GPU type. It's quite tedious for sure, but I guess this is the price to pay to get this kind of results.

- the batching question is a great one. In agentic systems, there is probably a trade-off between sequential thinking/iterations vs parallel exploration of multiple solutions. Also, there could just be multiple independent tasks running in parallel, depending on the use case.

We plan to support a small amount of batching, but it quickly becomes a trade-off vs speed. Pick one for your use case, I guess.

Also to consider: because we answer requests much faster, we are also able to process lots of them without needing high batches - and scaling on multiple nodes is possible.

- open sourcing: maybe, maybe not. I'm still undecided on this. We are a small startup and I'm told that giving our IP away might be shooting ourselves in the feet. On the other side, I think it could be of great benefit to the community and for us... we'll see

Our process has been, and will continue to be, a sequence of (tedious) R&D experiments where the GPU never behaves as expected when pushed to its limits in ways no-one really tested before (I still have nightmares of the L3 cache cross-IOD bottlenecks on MI300X).

IMHO, we did solve the multi-GPU memory bandwidth scaling problem, and thus the linear scaling of the size of the model towards infinity. But the main difficulties will come from keeping the speed, with steady and continuous memory streaming, while implementing the much more complex architecture of modern frontier MoEs (attention compression tricks, hash layers, routing logic, etc.)

- model size: 2B is just for this preview (it was faster to implement), our article explains how we expect to support large frontier MoE at 1,000 to 5,000 tokens/s

- reaching 500 tok/s, or even up to ~1,000 tok/s, on a consumer GPU card is possible with existing inference engines like vLLM. But there is a ceiling.

The hard part comes we you try to be faster than that: these frameworks won't scale higher just by adding GPUs or using faster GPUs. There is a "glass ceiling" due to microseconds lost everywhere in the stack (grid syncs, inter-GPU comms, kernel launches, CPU sampling, etc.).

All our work at Kog is about removing these bottlenecks.

Monokernel deep dive (GPU Engineering): http://blog.kog.ai/building-a-single-kernel-latency-optimize...

Delayed Tensor Parallelism (research): http://blog.kog.ai/delayed-tensor-parallelism-for-faster-tra...

To try the speed on the playground: http://playground.kog.ai

Sorry for the confusion