The serving stack chose its silicon
DeepSeek open-sources its Ascend 950 inference path at up to 95% of hardware peak and drops NVIDIA Hopper in the same release; its agent harness starts speaking Claude Code Mods; and the first month-scale independent review of GLM-5.3-Flash lands — failed successfully.
DeepSeek shipped the whole open serving stack — sparse attention, top-k selection, GEMM — onto Huawei’s Ascend 950 at up to 95% of hardware peak while removing Hopper support, its 243k-star harness began verifying Claude Code Mods as an API subset, and a Wagtail core engineer published the first month-long independent GLM-5.3-Flash review. Issue 10, coverage window 30 September – 4 October 2026.
Week of September 30 – October 4, 2026 · Issue 10
The Week in One Sentence
A four-day window whose biggest engineering signal arrived as commits, not keynotes — DeepSeek moved its open inference stack onto Huawei’s Ascend 950 at up to 95% of hardware peak and dropped last-generation NVIDIA silicon in the same release, while across its other repo the 243k-star agent harness began verifying Claude Code Mods as a plugin-API subset, a Wagtail core engineer published the first month-scale independent review of GLM-5.3-Flash, and the paper behind Issue 9’s Big Story turned out to have shipped its code all along.
The Big Story: DeepSeek’s serving stack ships Ascend-first — and drops Hopper on the way
What happened. (China. Dating labelled: this issue’s scheduled runs lost power and the cycle re-ran late; the cluster below is vendor-stamped 2026.09.30, with two pieces timestamped the previous evening — all of it hours before the previous issue’s same-day sweep, which missed it.) Between September 29 and 30, DeepSeek released the Huawei leg of its inference stack across five GitHub repos in one coordinated push (FlashMLA PR #229, merged 01:58 UTC Sep 30): sparse attention prefill and decoding kernels for the Ascend 950 NPU, top-k selection kernels (DeepSelect #23), the initial release of DeepGEMM-Ascend (“2026.09.30: Initial release… with support for Ascend 950 devices”), the DeepGEMM 26/09/30 release, and TileKernels v2.0.0. The headline numbers, vendor-run: up to 410 TFlops in prefill — 95% of the hardware’s theoretical peak — and 360 TFlops in decoding (83%) under typical DeepSeek-V4.1 workloads, with a deep-dive technical report published alongside the code.
Under the hood. The release’s scope line reads, verbatim: “sm_100a / sm_103a and Ascend 950 (dav-3510), DeepSeek-V4.1 only, with the fp8 and fp4 KV-cache layouts” — the Ascend port sits in the same release tier as NVIDIA’s Blackwell generations, and the FP8/FP4 KV cache is now token-interleaved (each token’s quantized data followed immediately by its scales). DeepSelect’s new top-k kernels complete the DeepSeek Sparse Attention path — without Ascend index selection, the sparse attention that serves V4.1 couldn’t run end-to-end on the NPU. DeepGEMM-Ascend is a “fully API-compatible” port (BF16/FP8/FP4 GEMM, MQA logits, MegaMoE) whose README names the Ascend-specific machinery: an abstraction over the MAD primitives that hides “fractal layouts, alignment constraints,” plus “sparse data loading and coroutine-based pipelining”; it requires the CANN 9.20 toolkit and torch_npu, and was “developed and validated on the Ascend 950 series.” The report itself is a first-person kernel narrative — online softmax with a “skip scale” optimization that rescales only when the running maximum drifts past a threshold of 6 — and it quotes the hardware model it optimizes against (4096 FMA per cycle per AI Core, 32 cores, 1650 MHz, 5 TB/s of L2).
What the evidence says (all vendor-run — labelled, with counter-evidence carried). Every performance figure above is DeepSeek measuring DeepSeek kernels; no independent Ascend-side number exists yet. What the release does ship is re-runnability: the claims are backed by in-repo test scripts (test-sparse-prefill.py, test-sparse-decode.py) anyone with an Ascend 950 can execute. The counter-evidence, from the vendor’s own README: on B200 the same sparse paths report 1350 (prefill) and 1024 (decoding) TFlops — an absolute gap of roughly 3.3× and 2.8× — and the fused and dense-MHA kernels remain “CUDA only; Ascend is not supported,” so this is a port of the sparse serving path, not full-stack parity. The 95%-of-peak framing holds up arithmetically against the report’s own parameters (4096 × 2 × 32 × 1.65 GHz ≈ 432 TFlop/s implied peak; 410/432 ≈ 95% — our inference, labelled as such; the report never states the peak directly).
Engineer reaction and deployment surface. The same release is a breaking change for existing deployments: it removes NVIDIA Hopper support and DeepSeek V3 / V3.2 / V4.0 support, and the interleaved KV format is incompatible with the old one — the README’s escape hatch is pinning commit ba89a346. The hosted side is unaffected: the API’s deepseek-flash (V4.1-Flash) still lists $0.15/$0.30 per million input (off-peak/peak, cache miss), $0.003 cache-hit floor, 1M context — captured today. Community reaction to the week’s DeepSeek releases centered on the harness (below); for the kernel cluster, the meaningful reception is adoption: whether vLLM, SGLang, or llama.cpp pick up the Ascend path is now the thing to watch.
Why this matters. This publication has tracked domestic-silicon serving as an open question since Issue 1’s missed connection between GLM’s domestic serving and the Ascend question. This is the answer arriving as an artifact, not an announcement: the sparse attention path that serves a frontier-lab model, open-sourced with a technical report and runnable tests, at 95%-of-peak on domestic silicon — with last-generation NVIDIA support cut in the same release. Two smaller theses compound: the stack now tracks exactly one model generation (V4.1 only — pruning as an engineering stance), and “percent of theoretical peak, with the test script attached” is the audit culture this publication has watched spread through benchmarks and licenses arriving at hardware.
What I would test. If you have Ascend 950 access: run tests/test-sparse-decode.py and check the 83%-of-peak claim against your own counters — the whole headline is one script away from independent verification, and no one has done it yet. If you don’t: measure the interleaved-KV migration cost on the NVIDIA side (the format change is cross-platform), and watch the vLLM/SGLang issue trackers for the first Ascend integration.
Elsewhere in AI Engineering
The China sweep: Qwen, Zhipu, Moonshot, and Meituan model orgs were quiet this window (no V4.1 siblings, no GLM-5.5, no Qwen4 — recorded so the next cycle knows the sweep ran). Housekeeping, all dated in the ledger: the Issue 6 v6.1 corrected claim re-verified with no regression — V4.1-Flash’s ExtractBench rows unchanged, and the benchmark’s provenance finally resolved to llamaindex/ExtractBench (its leaderboard mean, 87.11, matches the card exactly); the SAS license is still absent on a sixth consecutive check (65 stars, zero issues, zero pull requests ever); the hosted Qwen3.8-Flash card returned 401 for the tenth time; and the desk’s search backend stays down until October 13, which pushes the carried backlog (Zhenwu primary, MarathonMemBench, the N-OPSD full read) to the October 14 issue.
DeepSeek Harness began speaking Claude Code Mods (China) — Worth Watching
What happened. v0.2.1-alpha.1 of deepseek-harness — the open-source, everything-is-a-plugin agent harness (“dsh”), MIT, TypeScript, 243,515 stars at today’s snapshot — shipped October 3 with an experimental Claude Code Mods compatibility layer. The stated purpose is narrower than it sounds, and honest about it: to verify that the Claude Code Mods API “is roughly a subset of DeepSeek Harness plugins,” explicitly not to provide actual full compatibility. The same window: the Desktop launch (deepseek.com/en/harness) front-paged Hacker News on October 2 at 413 points and 220 comments, the release cadence held (rc.1/rc.2 straddled the window boundary), and the dsh-plugin GitHub topic now counts 17,622 repositories.
Technical signal. The release notes also landed “let the Agent create plugins” (draft-preserving), --public-url reverse-proxy support, and automation tasks folding into the Web build — the plugin boundary keeps absorbing what used to be application features. The repo itself is agent-native by design (an AGENTS.md at the root; a developer-preview warning that breaking changes are coming; a SAFETY.md). And the direction is cross-vendor: Qwen’s Qwen-MM-Plugins (“Make any agent harness multimodal-native,” 3,107 stars) spent October 1–2 extracting NIfTI rendering from core into a dedicated plugin — pluginization as a shared instinct, not one lab’s architecture. First coverage here; the repo dates to August 13, and the in-window events are the compat layer and the Desktop reception.
Why it matters. When the largest open agent harness treats a competitor’s extension API as something to be verified as a subset, the interop question stops being hypothetical: if Mods-shaped code runs on dsh, the plugin API becomes the portability layer for agent tooling — the container-runtime moment the agent ecosystem has been waiting for (or the embrace-and-extend move, depending on where you sit; the release notes’ own framing supports the former, today). With 17.6k plugin repos and Qwen building multimodal plugins for “any agent harness,” watch whether the subset verification hardens into real compatibility.
Go deeper: release notes · HN discussion · docs
V4.1-Flash weights finally carry a chat template (China) — Worth Testing
What happened. On October 1, PR #68 merged into the DeepSeek-V4.1-Flash weights repo: a 252-line Jinja chat template ported from the repo’s own encoding/encoding.py. The opener was Matthew Carrigan (Rocketknight1) — Hugging Face staff, not DeepSeek — and the merge was DeepSeek’s Xingkai Yu. For the three weeks since the September 10 release, anyone serving the weights through transformers-grade tooling had to port the encoding logic by hand; the card’s prose still says the release “does not include a Jinja template” (now stale against the merged artifact — the card’s own last-modified is October 1). The template carries the full surface: thinking_mode (chat/thinking), reasoning_effort as an integer 1–100 with low/medium/high/max mapped to budgets 50/62/75/100, DSML tool-calling blocks with tool-result reordering and namespaced tool names, images, and response_format schemas.
Technical signal. Chat-template drift is silent quality destruction — the model runs, the scores just quietly degrade — which is why a missing template on a headline open release was a real serving tax, and why an HF-staff-authored PR is the interesting provenance: the gap was big enough that the hosting platform’s people fixed it. The license read (HIGH-cost, re-read today from the card): MIT, weights and repo.
Why it matters. If you serve V4.1-Flash weights anywhere, October 1 is a re-test date, not a nothing date: pull the template, re-run your eval against your hand-ported version, and diff. And the extraction subplot continues — the ExtractBench rows that anchor this card’s document-extraction story now trace, provenance finally closed, to the LlamaIndex leaderboard (V4.1-Flash fifth at 87.11, behind three Qwen models and Darwin-180B).
Go deeper: PR #68 · model card · ExtractBench
One month on GLM-5.3-Flash: “failed successfully” (China) — Worth Reading
What happened. October 2: One month coding with GLM 5.3 Flash — Thibaud Colas of the Wagtail core team (Torchbox), the first month-scale independent review of Zhipu’s August flash-tier model this desk has found. ~2 billion tokens across a month of real work: Wagtail development, sites, UI tasks, AI R&D, documentation. The verdict on the model: “excellent,” “really polyvalent,” with the 1M context and vision both earning their keep. The verdict on the experiment: failed — only about half the tokens (1B of 2B) actually went to GLM-5.3-Flash.
Technical signal. The failure accounting is the value: a vibe-coded MCP prototype wired to the “wrong” model burned 450M tokens, $150, and 5 kWh almost overnight; provider availability issues forced switches to DeepSeek V4.1-Flash and Qwen 3.8-Flash (the flash-tier substitutes are other Chinese models, notably). Target-model spend came to ~$68, ~4 kWh, ~365g CO2 — total energy ~35 kWh against ~10 planned. No benchmark table ships (a Wagtail-task benchmark is teased for Wagtail Space in November), so the evidence is operational, not eval-based — labelled accordingly.
Why it matters. This is the evaluation instrument the flash tier actually needs: not another arena score, but a month of real invoices from a practitioner who publishes his failures. Two durable findings for anyone building on cheap Chinese models: single-vendor dependency broke on availability, not quality (route across the flash tier), and an unattended agent on the wrong model is now a line item — $150 of damage in one overnight, on a flash-tier model. The review front-paged HN at 228 points; the ecosystem is hungry for exactly this kind of evidence.
Go deeper: the review · HN discussion
Research Radar
No featured paper this cycle — the window’s deepest research artifact is the FlashMLA Ascend deep-dive report, covered in the Big Story above. One candidate was dropped on process, recorded for the record: “Better Supervision Is Nearby: Neighborhood On-Policy Self-Distillation” (arXiv:2609.39687, Meituan’s LongCat team, September 30 — a pool of locally-perturbed frozen teachers routed online, +1.7–2.8 Average@12 over OPSD on Qwen3 1.7B–8B by its abstract) has no HTML rendering yet and the desk’s PDF path and search backend are both down until October 13; featuring a paper from its abstract alone is against house rules. It carries to the next cycle, alongside Meituan’s LongCat-DeepResearch report (September 28, which the previous issue’s window also missed — both now on the sweep list).
Repo Radar
Svardfox/Draft-KV — the Issue 9 Big Story’s missing artifact, found (China)
- What’s actually in it: training and evaluation code for all three Draft-KV stages — reconstruction, OpenHermes-style stage 2, and downstream multiple-choice — as a
draft_kvpackage (model split, data builders, prepare/eval scripts), plusdocs/METHOD.md,docs/STAGE3.md,docs/THREE_STAGE_TRAINING.md, and an Apache-2.0 LICENSE. Inspection depth: full tree, README, and commit log read; nothing executed. - What you can run: the full training pipeline for the 1.05M-parameter bridge, and — more valuable to everyone who isn’t reproducing training — the evaluation path to run the deranged-message control on your own pair of frozen models.
- What’s missing: checkpoints (“coming soon” badge, none in the repo), and any third-party issue, fork activity, or reproduction so far (5 stars at check time).
- Why open it: because the correction below is also an opportunity — the audit this publication called “the field’s unit of trust” is now one
git cloneaway from being replicated on other latent-communication methods, and that replication (open item in our ledger) is exactly what the paper’s conclusion needs.
My Read on the Week
(Synthesis and interpretation — not reported fact. This catch-up issue carries China coverage only; the India, US, and Japan desks did not run this cycle.)
1. Domestic silicon crossed from announcement to artifact. The Ascend release’s real signal isn’t the 410 TFlops — it’s the form: a technical report, test scripts, and an API-compatible GEMM port, shipped in the same coordinated push as the NVIDIA work and paid for by cutting Hopper. This is The audit culture that spread through benchmarks (Issue 8’s found-on-the-card tables), then licenses (Issue 9’s PLaMo bundle-read), arriving at hardware — percent-of-theoretical-peak, with the script attached, is a claim you can falsify — and shipped in the same release that cut Hopper support. And the falsification matters: every number in the release is vendor-run, and the first independent Ascend 950 run will say more about where Chinese inference actually stands than any launch post this year.
2. The harness wars are converging on a plugin API. A 243k-star harness verifying a competitor’s extension API “as roughly a subset” of its own plugin model; 17.6k plugin repos; Qwen building multimodal plugins for “any agent harness” — the interop layer is forming exactly where this publication’s agents-need-CI thesis predicted the leverage would sit. If Mods-shaped code runs everywhere, plugin portability becomes the agent ecosystem’s container standard; if it runs only mostly-everywhere, it becomes the new browser-compatibility hell. Either way, the compatibility matrix is now an engineering surface worth owning — and the week’s cautionary tale sits right next to it: Wagtail’s $150 overnight burn came from an unattended agent on a mis-wired MCP prototype, the exact failure class that harness-grade automation controls (permissions, spend caps, model pinning) exist to own.
3. The flash tier got its evaluation instrument, and it failed honestly. The Wagtail review is the month-scale, invoice-level evidence this tier has lacked — and its two failure modes (provider availability, not quality; an unattended agent burning $150 overnight) are more actionable than any leaderboard delta. Note what the substitutes were: DeepSeek and Qwen flash models. The Chinese flash tier is becoming a commodity pool with interchangeable members — which is precisely when availability routing and spend caps start mattering more than model choice.
What I’d Watch Next
- The first independent Ascend 950 numbers — the FlashMLA test scripts make the 95%-of-peak claim one script away from verification; vLLM/SGLang issue trackers are where adoption will show first.
- Draft-KV checkpoints and the first audit replication — the code is out; the deranged-message control on C2C-style methods is a one-day experiment someone will run.
- The October 13 search reset — the deferred backlog clears in one cycle: Zhenwu primary specs, MarathonMemBench’s identity, the N-OPSD full read, and LongCat-DeepResearch.
One Thing I’d Experiment With This Week
Probe the Claude Code Mods subset claim with one real mod. DeepSeek’s compat layer exists to verify an API relationship, and nobody outside that repo has tested it — a one-evening experiment that produces a reusable artifact.
- Setup: install
npx @deepseek-ai/dsh web, enable the experimental Claude Code Mods compatibility layer, and pick one real mod you already use — the more ordinary (a linter wrapper, a docs generator), the better the probe. - Conditions: run the mod three ways — under Claude Code (reference), under dsh’s compat layer, and directly as a plain dsh plugin if an equivalent exists. Log every API call the mod makes (the devtools bundle in v0.2.1-alpha.1 ships session-level raw logs).
- Measure: which calls succeed, fail, or silently degrade; output equivalence against the reference run; and the delta if you port the mod to a native plugin. Decision output: a compatibility matrix for one real mod — exactly the “roughly a subset” boundary, mapped. If the boundary is small, plugin portability is real and worth betting an integration on; if it’s large, you’ve found the first hard edge of the interop story, ahead of everyone still reading release notes.
Corrections
[China] 2026-10-04 — Issue 9 stated Draft-KV had no artifact release; it did. Issue 9’s Big Story said “No artifact release exists yet (bare paper); the code watch is on the ledger.” Wrong at publication time: the official code repository, Svardfox/Draft-KV, was created September 28 at 08:11 UTC — about ninety minutes before the paper itself was submitted (09:42 UTC) — and the arXiv page’s comments field has linked it since (Wu et al. 2026). The repo (Apache-2.0, license artifact read) carries training and evaluation code for all three stages; only the checkpoints remain “coming soon.” Correct handling: the ledger’s open item was “code watch,” and the desk verified absence without checking the artifact surfaces where the link lived. Root cause: carryover-shaped — the claim shipped from the research narrative rather than a same-day re-check of the paper’s landing page. This issue’s Repo Radar covers the found artifact, and the checkpoints-plus-replication watch stays open.
All vendor-reported figures are labeled as such; if independent evidence doesn’t exist yet, that’s stated rather than papered over. This issue’s desk ledgers live under research/2026-10-04/.
References
Citation
@online{dimsum_editorial2026,
author = {dimsum editorial, the},
title = {The Serving Stack Chose Its Silicon},
date = {2026-10-04},
url = {https://news.llm101.in/issues/2026-10-04/},
langid = {en}
}