Nov 2, 2026
Four Bits, Three Contracts: An Interactive Guide to MXFP4, NVFP4 and Q4_K
A hands-on companion to the Antshiv Robotics video: decode E2M1, move shared scales, unpack Q4_K bytes, and follow the numbers into a real runtime.
Read post βTagged With
20 posts connected to this tag.
Nov 2, 2026
A hands-on companion to the Antshiv Robotics video: decode E2M1, move shared scales, unpack Q4_K bytes, and follow the numbers into a real runtime.
Read post β
Oct 30, 2026
Listen to three labeled clips as I compare CKE's Kokoro and Qwen speech work: what's working, what's slow, and what still needs proof.
Read post β
Oct 28, 2026
We replayed Cohere's Qwen trace through CKE. The model continued an already-started command, but our tests ran no tool and sent no data. Here is what the evidence does and does not show.
Read post β
Oct 25, 2026
A practical accessibility architecture for a dense Linux agent workspace, with real screenshots, CKE Whisper/Kokoro/Qwen3-TTS PR history, and honest boundaries around named-agent navigation.
Read post β
Oct 22, 2026
A visual kernel-by-kernel comparison from text to sound, including CKE first-block math and missing codec contracts.
Read post β
Oct 20, 2026
Why four-bit quantization matters: follow packed weights through scales, CPU registers, AVX-512, and GEMM, then compare Q4_K, MXFP4, NVFP4, and CKE implementation PRs.
Read post β
Oct 17, 2026
From an early DS4 harness experiment to CKE tokenizer, template, tool-protocol and real harness acceptance work.
Read post βOct 16, 2026
Kernel math / 03A three-tap filter can listen farther in time without gaining another weight. Kokoro uses that idea, but the filter is only one part of making speech. In the first two parts ...
Read post β
Oct 13, 2026
CKE now serves bounded Qwen Code and Gemma4 agent tasks from generated native CPU bundles. Here is what passed, what failed, and why batching remains a separate gate.
Read post β
Oct 11, 2026
A vision model can return plausible text while taking the wrong path through its image prefix. That is the trap I want CKE's tests to catch. In October, the Gemma4 and multimodal work was le...
Read post β
Oct 11, 2026
I want my Linux workstation to read an agent's answer aloud. Whisper already helps me transcribe my own speech; Kokoro is the other direction. I have been following its path through CKE one ...
Read post β
Oct 11, 2026
When I say I want to run several models across my own machines, it is easy to jump straight to a diagram with six GPUs or a cluster of CPUs. The harder question is what one user request actu...
Read post β
Oct 3, 2026
September was the month C-Kernel-Engine started to feel less like a collection of model demonstrations and more like a system I could keep extending. That is not the same as saying it is fin...
Read post β
Sep 29, 2026
Kokoro's phoneme encoder just took its first concrete step into CKE's generated-C lane. The step is deliberately small: one operation β three embedding tables plus a per-token LayerNorm β...
Read post β
Sep 27, 2026
I have been working on speech recognition in C-Kernel-Engine for a while. Whisper can transcribe my own videos. The other direction is becoming more interesting to me now: I want my Linux wo...
Read post β
Sep 27, 2026
I Want To Use My Linux Workstation Without Looking At It I have spent a lot of time building software that lets me run AI models on CPUs. Now I want to use some of that work for a more perso...
Read post βSep 23, 2026
A measured two-CPU investigation of how CKE moved Parakeet and Cohere into generated C, including six complete 42-minute runs, exact cross-host transcripts, performance bottlenecks, and rema...
Read post βSep 13, 2026
Run CKE manually or use the included local coding-agent prompt to inspect Linux, install dependencies, execute a small Gemma model, and preserve the evidence.
Read post βSep 13, 2026
CKE now runs three CPU audio paths on the same 42-minute recording: Whisper is practical and fast, Cohere produces deterministic text and SRT across Intel and AMD, and NVIDIA Parakeet works ...
Read post βAug 12, 2026
Inside the CKE compiler One CKE runtime serves many model families, quantization formats, CPU instruction sets, and execution phases. So how does it decide which exact C kernel runs for each...
Read post βI write about the maths, the kernels, and what happens when I try to make them work. Join me for the investigations, including the bits that go wrong.
Get the illustrated field guide by email, followed by engineering notes as I write them.