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 finished. I spent much of the month finding places where a model appeared to work until a real recording, a real tool call, or an independent numerical comparison asked a harder question.

I wanted CKE to run models on CPUs through kernels, circuits, kernel maps, generated C, and evidence I could inspect. September added another question: can the same architecture keep working when the model is not just another decoder? We exercised it with Qwen3.8 Flash, Whisper, Parakeet, Cohere Transcribe, vision, a small training loop, and the early Kokoro text-to-speech graph. Serving became more than a port that returns tokens, helped substantially by an outside contributor.

This is a retrospective on work merged in September 2026. I reviewed CKE at commit 8ef9b28a6 on October 3, but I have kept October's follow-on work out of the September claims.

What moved during the month

The sequence matters more than a count of pull requests. Early in the month, the Qwen3.8 Flash and Whisper fixes exposed where a plausible output was not enough. Then native generated audio paths arrived for Parakeet and Cohere. Later we connected more training, vision, serving, and Kokoro boundaries. All of that was accompanied by more explicit tests for missing evidence and wrong artifacts.

The compiler had to meet a stranger model

Qwen3.8 Flash Next (#456) made the challenge concrete. Its circuit needed explicit hyper-connections, per-layer embeddings, sparse attention, and mixed expert contracts. We added circuit and kernel-map declarations and then chased numerical drift with X-Ray. A short Q4 trajectory could be checked; that did not certify BF16, long context, or every chat shape. I prefer that narrower sentence because a new model can expose a missing compiler capability without invalidating the whole architecture.

The important movement was not “CKE knows Qwen3.8 by name.” It was that more of the model's requirements were expressed as circuit edges, exact providers, layouts, and execution-plan rules that another model could reuse. The bring-up trail records both the repaired reductions and the comparisons still unavailable. That is how I want model support to mature: declare, generate, compare, localize, then retain the regression.

Audio stopped being just a kernel demo

Whisper became practical for my own video work, but a real recording also found a graph-wiring problem and an audio-tail problem. #455 hardened long-audio execution; #476 and #482 reduced repeat work across windows; #483 made worker teardown fail closed. This was useful software work because I could transcribe a recording, inspect the result, and catch a regression before using it to edit a video.

Parakeet and Cohere then tested a different question: could the generated model own more of the audio graph instead of Python reconstructing it around native kernels? For Parakeet, #530 generated the TDT state progression and a standalone WAV host. The 42:22 recording completed on Ryzen in 17 windows, and the transcript matched the retained CKE reference after trimming a trailing newline. That is strong execution-equivalence evidence, not an independently aligned human word-error-rate result. Cohere's separate frontend, encoder, decoder, and long-audio work followed. I am not combining those into a universal accuracy or speed ranking.

Kokoro was the other end of the audio loop: speech out, rather than words in. By September 30, the generated path had a bounded phoneme encoder, duration and alignment, text expansion, shared prosody LSTM, and connected F0/noise prosody on pinned input. The shared ALBERT encoder work in #602 and the connected prosody work through #634 are substantial. But at the September cutoff there was no complete acoustic decoder, waveform, CKE WAV, or speech session. October has moved some of those boundaries; that is a follow-up article, not a reason to rewrite September as if it already had TTS.

Training and vision made the evidence more demanding

I care about backpropagation because I do not want CKE to stop at inference. #527 restored a scoped v8 generated FP32 training workflow: forward and backward, native AdamW, checkpoint/resume, export to inference, and checks that the loaded runtime was actually the one we built. The two-layer numerical certification and four-layer English-text workflow are useful foundations. #563 added two-layer composition parity and a dtype ledger. This does not mean arbitrary large-model training, BF16 training parity, or every optimizer is done.

Vision brought a different discipline. A generated encoder can execute and still fail the task a user cares about. September's independent evidence and vision-evaluation hardening separated execution coverage from OCR field correctness and kept private evaluation details out of public reports. OCR is one vision task, not a blanket certificate for image understanding. The result is less impressive as a headline and much more useful when deciding what to trust.

A contributor helped turn serving into an actual boundary

CKE also started meeting the world outside its command-line runner. An outside contributor helped shape that serving layer. The arc runs from a native session behind HTTP/SSE in #370, through streaming and response tests in #451 and #473, to structured tool-call work in #495. That contribution mattered, alongside the compiler and server work around it.

The September #594 PR shows why agent serving is more than returning tokens. Qwen3.5's publisher template could render correctly while its imported tool delimiters became different BPE token IDs. Separately, a native worker could finish while an unconsumed HTTP stream kept the single-flight slot occupied, causing subsequent requests to get 429 Session busy. The fix bound special-token handling to imported metadata and the worker lease to request ownership. Its pinned 0.8B bundle matched 327/327 prompt token IDs in one tested mode and 325/325 in another; Responses and Chat Completions both completed a native read_file tool call and continuation. The Qwen Code harness still repeated tool calls until its limit, so I will not turn that into a claim that a coding-agent task was completed.

Three streams toward practical use

By the end of September I could see three streams that need to meet, not three separate product announcements. Vision asks whether a generated path can accept an image and produce a result whose task quality we can independently judge. TTS asks whether a circuit can carry text through the acoustic graph to actual samples a person can hear. Serving asks whether the right model, tokenizer, prompt, tools, and request lifetime are bound together when a real client calls. Each stream improves a different contract CKE will need before I can rely on it daily.

For vision, September's independent comparison work and evaluation hardening were less about claiming a new model and more about refusing to call execution a successful reading of an image. The next useful evidence is a clearly scoped task such as OCR field extraction or a particular image question, evaluated against retained independent answers. A passing OCR field test would still not certify arbitrary photo description.

For TTS, the September Kokoro circuits turned more of the phoneme encoder, duration, text expansion, and prosody branches into generated execution. The missing output was the one I would actually listen to: waveform samples. That is why the September status is connected features, not “CKE speaks.” The next gate is a generated acoustic decoder and waveform path, then text preparation, repeated speech requests, playback, and listening-quality evaluation. The current Kokoro runbook can be ahead of this September snapshot; I am keeping the dates separate.

For serving, #594's token-ID and worker-lease fixes made the first text-and-tool path more trustworthy. The next question is whether that path survives a complete user session: model identity, cancellation, context limits, tool result, repeat request, and an honest error when something is unsupported. The September server is single-flight, not a multi-node scheduler or a general multimodal gateway. A local model can still be useful through it, as long as the application treats that limit as real.

I want that loop for my own work: transcribe a recording, inspect an image or screen when needed, ask a local model to act through bounded tools, and hear the important result while I continue working on Linux. Whisper already helps with the first step. Vision, serving, and Kokoro need to earn their parts through the same generated-C and regression discipline. Using an external speech output or coordinator while CKE catches up is an honest integration choice, not a claim that CKE already does everything.

The less visible progress: refusing false passes

We tightened nightly and regression evidence while adding models. #467 made the nightly verdict follow the actual gate rather than merely publishing an artifact; #515 made idle-nightly scheduling capacity-aware; #518 kept kernel evidence distinct. That will not make regressions impossible. It makes it harder for an old report, a missing checkpoint, or a skipped test to masquerade as a current pass.

My September takeaway is simple: the models are teaching the architecture where its contracts are incomplete. Audio forced execution ownership and window policy into the open. Vision forced a separation between runtime success and task success. Training forced artifact identity and gradients to be checked. The serving work forced us to care about token IDs, request ownership, and real tool round trips. That is progress I can build on, even when the next model still finds a new edge.

Scope: September 2026 merged work. Sources are linked at each claim. Numerical and task results refer to their stated model, machine, artifact, and test; they are not general performance guarantees.