Strata 0.2.14 now reads prompts using GPU kernels it wrote itself, shedding its last hard dependency on llama.cpp. The change delivers mostly single-digit speed gains across hardware, with the standout being a 256-token prompt on AMD's Strix Halo dropping from 2,010 ms to 408 ms. Skeptics point out that quality claims remain qualitative, documentation contradicts itself, and one audit found roughly 19% of downloads were just the demo trailer. Strata runs only three models in exchange for speed, and the recommended setup keeps a stable OpenAI-compatible serving layer while treating the runtime itself as replaceable.

Strata 0.2.14 ships its own prompt-kernels, one step from dropping llama.cpp entirely

The open-source local AI engine that runs a 125-billion-parameter model on a gaming PC now reads prompts with kernels it wrote itself, and the last hard dependency on llama.cpp is now optional.

Strata 0.2.14 has been released today. Both the app and terminal has been bumped to 0.2.14, while the engine packages moved to 0.1.60. Keep in mind that the both Intel Arc and DGX Spark engines remains at version 0.1.59.

"Nothing to do, update as usual," the release notes say. No new settings. There's a genuinely low-key mood here. And yet the one headline change is exactly the part worth paying attention to.

At first the change is easy to shrug off. Strata has always leaned on llama.cpp's kernels for prompt reading, the step that turns words into numbers the model can chew. Now those kernels are Strata's own, on NVIDIA and AMD cards, on Linux and Windows. -DSTRATA_MMQ=ON still builds the old route if you insist, but it isn't the default anymore.

That's the moment Strata removes its last real dependency on llama.cpp's prompt code. It's an overfit engine built on llama.cpp's formats, not a branch of llama.cpp's source. That framing tells you everything. It carries GGUF files, the quantization formats, the kernel lineage, but not the breadth.

Strata built it, then stopped needing llama.cpp

To see why that matters, a quick walk through what Strata actually does. The headline model, Qwen3.8-Flash-Next, has 125 billion parameters but only uses about 6 billion per token. That's a Mixture-of-Experts design, and it means the full thing doesn't fit in a gaming GPU. General-purpose engines grind to a halt once the VRAM fills up.

Strata spreads the load across the whole machine. VRAM holds the hot stuff, RAM holds the bulk of the experts with the CPU churning the rest, and an SSD holds a lookup table read in tiny rows per token. A draft head guesses the next few tokens, the engine checks them in one pass, and Strata says that lands you the same answer 1.6 to 1.8 times sooner.

Strata publishes per-card numbers, which is unusual and actually useful. The showstopper is short prompts on AMD's Strix Halo. A 256-token prompt dropped from 2,010 ms to 408 ms. That's roughly five times faster just to begin answering.

On an RTX 5090 with a 32K prompt, output climbed to 7,870 tokens/s from 6,851. A 20,000-token prefill on an RTX 4090 came down from roughly 6,150 ms to 5,651. These are all Strata's own numbers, on specific machines with a specific setup. Take them on faith accordingly.

Head here for the release, and here for the full notes before you install.