China’s silicon story graduated from claims to SKUs — receipts pending

Alibaba’s Zhenwu V900 is a real chip with a real CEO quote and no published spec sheet; EVIE makes document retrieval the second contested niche in two weeks.

AI Engineering
Hardware
Open Weights
China
India
US

Alibaba unveils the Zhenwu V900 at Apsara — “China’s most powerful AI chip,” 3× the M890, with no absolute specs yet — plus a 5–10T-parameter model plan; Tencent fully open-sources EVIE, the ViDoRe-V3-#1 document retriever, under Apache-2.0. Issue 7, coverage window 21–23 September 2026.

Author

the dimsum editorial

Published

September 23, 2026

Week of September 21–23, 2026 · Issue 7

The Week in One Sentence

China’s window produced silicon instead of models — Alibaba’s Zhenwu V900 arrived as a named SKU with a 3×-over-M890 claim and no spec sheet behind it, and Tencent’s Apache-2.0 EVIE opened document retrieval as the second contested niche in two weeks — while the US launch storm (GPT-6 Sol and Luna, Opus 5.5, Grok 4.7) fought on price, cache economics, and undeletable reasoning rather than capability, and India’s Sarvam quietly finished its keyterm-biasing surface on streaming speech-to-text.

The Big Story: the Zhenwu V900 — a named chip, a bold claim, and no spec sheet

What happened. (China.) At Alibaba’s annual Apsara conference in Hangzhou on Tuesday September 22, CEO Eddie Wu unveiled the Zhenwu V900, calling it “the most powerful AI chip in China today” — three times the performance of the previous-generation Zhenwu M890 — and said it powers both frontier training and inference across Alibaba’s data centers for the company and its cloud customers (AP). Availability timing was not stated. The same keynote carried a supply admission worth more than the headline: global shortages “are currently limiting the speed at which we can scale our compute infrastructure,” with Alibaba “mobilizing every resource” to keep up — plus a plan to train a 5–10-trillion-parameter model (the current flagship, Qwen3.8-Max, is 2.4T) and a projection of over 20 gigawatts of computing by 2032.

What the evidence says — layered, because it must be. The unveiling, the executive quotes, and the 3×-over-M890 framing come from the AP wire and are solid as attributed statements. What does not yet exist anywhere we can verify: an absolute spec sheet. Specialist press has published figures — 216 GB memory, 1,200 GB/s interconnect, native FP8 and FP4 support — but none of that appears in the wire copy or in anything we could read from Alibaba’s own channels this cycle (the Alibaba Cloud blog’s September 22 “full-stack AI strategy” post exists but wasn’t machine-readable). Per our HIGH-cost discipline, those specs stay labeled press-reported-unconfirmed and out of the verified ledger. The 3× claim itself is unfalsifiable as stated — no absolute number, no M890 baseline given.

Why this matters. This publication has tracked China’s silicon story as claims since Issue 1 — Zhipu’s ~100,000-domestic-chip serving assertion (still unaudited, still carried), Huawei’s Ascend roadmap, the bandwidth-vs-compute argument that sparse-attention, memory-heavy architectures presuppose. The Zhenwu V900 is a different category of evidence: a named SKU, from the country’s largest cloud vendor, positioned explicitly for the training-inference stack its own Qwen models run on. And the two data points rhyme: the one press spec with thesis weight is native FP4 — the same precision lane DeepSeek’s V4.1-Flash just made load-bearing for KV cache (Issue 6). If vendor silicon and vendor models are converging on 4-bit lanes and memory-first architectures in the same month, that is the infrastructure thesis compounding. Receipts pending at every layer: no spec sheet, no availability, no independent benchmark — and, on the other side of the ledger, Zhipu’s chip-count claim reaches its eighth consecutive ledger check still unaudited.

What I would test. Nothing yet — that is the point. The moment an official spec sheet or cloud-instance pricing lands: compare the V900’s claimed memory bandwidth per FP4-FLOP against H20/H200 class parts at Chinese-market pricing, and watch whether Qwen serving prices move on V900 capacity (the cache-hit price lines from Issues 4–6 are the baseline). Until then, this story is a watch item with a name attached.

Elsewhere in AI Engineering

Tencent fully open-sources EVIE — retrieval becomes the second niche war (China) — Worth Testing

What happened. (China.) EVIE-8B (created September 4) was updated September 22 with the milestone announcement: weights, inference pipelines, evaluation suites, training code, and distillation recipes are now fully open-sourced — under Apache-2.0 (license field read verbatim), with a 4.5B student model alongside the 8B teacher. EVIE is a visual document retrieval embedder (ColQwen3.5 backbone, bidirectional full attention, late-interaction multi-vectors at 4096-D per token with MaxSim), trained on 775,635 document-query pairs; the released checkpoint is a weight-space mix (α=0.40) of two training arms — an unusual, now-reproducible detail. Vendor benchmarks put it #1 on ViDoRe V3 at 66.75 nDCG@10 (92.18 nDCG@5 on V1, 74.23 on V2; 79.51 nDCG@10 macro across 138 tasks). No paper yet — the card says it’s coming, so benchmark provenance is card-only for now.

Why it matters. One week after the GLM-OCR/WeVisDoc parsing race (Issue 6), here is the adjacent niche — document retrieval — going fully open from the third major Chinese ecosystem, permissively licensed, with training code rather than just weights. The pattern is now unmistakable: China’s labs are shipping the entire enterprise document stack (parse → embed → retrieve) as open, commercially-usable components while the frontier-model race runs on headlines. For RAG pipeline builders, EVIE’s ViDoRe position is externally verifiable on the leaderboard — that’s the first check worth running, before any benchmark trust is extended to the card.

Keyterm prompting reaches Sarvam’s streaming STT (India) — Worth Reading

What happened. (India. Labelled update: the REST/Batch half of this feature shipped September 9 — pre-window and unpublished then; the streaming half is this window’s news.) September 22, per the docs changelog: the keyterms parameter now works on both streaming speech-to-text endpoints — /speech-to-text/ws (legacy) and /speech-to-text-realtime/ws — when using model="saaras:v4", matching the existing REST and Batch behavior: “a JSON-encoded list of up to 50 terms (64 characters each)” passed as a query parameter when opening the connection.

Why this matters. Domain-term biasing at recognition time is the difference between a voice agent that hears “Bhartiya” and one that doesn’t — and until Tuesday it was a batch/REST affordance on a platform whose voice agents run on streams. The feature is now uniform across the request surface, which is the quiet prerequisite for using it in production telephony and realtime pipelines instead of post-processing transcripts. Limits to know before relying on it: 50 terms × 64 characters is a small prior; the changelog says nothing about how biasing interacts with the realtime endpoint’s mid-stream reconfiguration; and the effect on recognition of terms outside the list is unmeasured from here — a vendor feature with no independent eval, labelled as such.

Go deeper: changelog

GPT-6 Sol and Luna: the cheap tier got halved (US) — Worth Testing

What happened. September 22, OpenAI announced GPT-6 Sol and GPT-6 Luna (the announcement page declines our fetcher — dating pinned via its HN thread and corroborating coverage; the body stays unread, so this item runs on the rate card and the community, labeled accordingly). The rate card, read today, completes the three-tier GPT-6 lineup: gpt-6-luna at $0.10 input / $0.01 cached / $0.50 output per million (cache writes $0.125), gpt-6-sol at $2.00 / $0.20 / $10.00, joining gpt-6-astra at $10 / $1 / $50 — long-context tiers at 2× input and 1.5× output, batch and flex at half, a 2× “fast” tier, and an EU-residency note (Sol/Luna: Standard processing only).

Why it matters. The community’s framing is the right one: Luna is half the price of GPT-5.6 Luna ($0.10/$0.50 vs $0.20/$1.20, community-cited) — Simon Willison calls that “a really big deal” — and the sharpest pushback is equally right: one user’s 30-day coding-workload computation had Luna at ~$184 where the cheapest Chinese models ran ~$40, because cache-read pricing, not list price, decides real agent workloads. Against DeepSeek’s $0.003 cache-hit floor (Issue 4), Luna’s $0.01 reads are the actual contest. Sentiment that Luna is “a price cut, not an improvement” over 5.6 is unverified from here — no independent instrument has scored the pair yet, and this desk publishes no benchmark figure it couldn’t read from a primary.

Go deeper: pricing · HN discussion

Claude Opus 5.5 cut prices and made thinking mandatory (US) — Worth Reading

What happened. September 22, per the platform changelog: claude-opus-5-5, for long-running agentic coding and knowledge work — 1M context, 128K output, always-on adaptive thinking, at $4 / $20 per MTok, down from Opus 5’s $5 / $25 (20% both lanes), across the Claude API, Bedrock, Google Cloud, and Microsoft Foundry. Breaking changes: thinking: disabled now returns 400 — reasoning can’t be turned off, only re-depthed via the effort parameter; tool_choice any/tool return 400 (use auto with strict tool use); computer use migrates to the computer_toolset_20260801 toolset (Bedrock’s computer_20251124 keeps working, per the changelog). In beta: inline tool definitions (inline-tools-2026-09-15) — tools defined mid-conversation in a system message without invalidating the prompt cache, with MCP toolsets via mcp-client-2026-09-15.

Why it matters. Undeletable thinking is a billing decision as much as a behavior one: you always pay for reasoning tokens now, and control only their depth. And the inline-tools beta is this publication’s cache thesis arriving as API surface — the same week OpenAI’s rate card started pricing cache writes explicitly, Anthropic shipped a mechanism whose stated purpose is not resetting the prefix bill. It is also migration week across town: Google’s gemini-omni-flash-preview stops September 30 (seven days from publication; replacement gemini-omni-1.1-flash) and antigravity-preview-05-2026 stops October 5 — anyone re-testing harnesses this week should re-pin everything at once.

Go deeper: changelog

Grok 4.7: same price, harder benchmarks — and a competitor leads its own table (US) — Worth Watching

What happened. September 21, xAI shipped Grok 4.7 — flagship for coding and knowledge work, “a new, larger base model,” RL weighted toward multi-hour problems, native Grok Bot harness understanding. Read the pricing carefully: the body says “served at the same price and speed as Grok 4.6,” and the rate card, read today, confirms identical rows — $2 / $6 per M below 200K prompt tokens, $4 / $12 above, 500K context; the headline “twice as fast, at half the price of comparable models” is a GPT-5.6-Sol-class comparison, not a 4.6 one. A “Grok 4.7 Fast” variant runs 2× speed at 2× price (Cursor and Grok Build only). Closed weights.

Technical signal. The vendor table, labeled as such1: CursorBench 4.0 46.3 (Grok 4.6: 40.4; GPT-5.6 Sol: 41.7; Fable 5.1: 51.8 — the leader in xAI’s own table); DeepSWE v1.1 71.0* at high effort (GPT-5.6 Sol: 72.7); Terminal-Bench 4.0 at 37.6 vs 4.6’s 20.3 — the hard generation of that benchmark (Issue 4 documented a 90.6-on-2.1 model scoring 31.2 on 4.0), so the near-doubling is the table’s most meaningful row; GDPval 1695 Elo (Fable 5.1: 1735); LatchBio 62.4 (top score); HackerBench 3.3% dual-use pass-through. Artificial Analysis (search-cited, not our read) has 4.7 at the lowest cost-per-task, ~$2.73 at high reasoning.

Why it matters. A same-price flagship refresh whose own comparator table shows a competitor leading two of three headline agentic benches is the launch storm in miniature: prices holding or falling, leadership fragmented by workload. The independent placements are where 4.7 actually gets situated — Real-SWE’s table, re-read today, has no 4.7 entry yet.

Go deeper: announcement · pricing

No Japan items this issue: the scheduled September 23 Japan desk run was missed, and the editorial fallback ships three regions rather than filler. Japan coverage returns with the September 27 issue.

Research Radar

No China paper in this three-day window met the read-beyond-abstract bar — the window’s artifacts are a chip announcement and a model open-sourcing without its paper (EVIE’s is “upcoming,” and is on the watch list). The V4.1-Flash tech report (Issue 6) remains the desk’s most recent full paper read.

Repo Radar

tencent/EVIE-8B (China)

  • What’s actually in it: 8.4B-param BF16 teacher (~16.8 GB, unsharded), 4.5B student, and — per the September 22 announcement — the full training and distillation recipe in the GitHub repo (repo tree not opened this cycle; card-level inspection).
  • What you can run: late-interaction retrieval with MaxSim scoring over multi-vector embeddings; the weight-mix checkpoint recipe is in the open training code, which is rarer than the weights.
  • What’s missing: the paper (benchmark provenance is card-only until it lands); teacher-vs-student scores aren’t broken out on the card.
  • Why open it: the retrieval half of the open document stack — pairs with the parsing models from Issue 6 into a fully permissive Chinese-lab RAG pipeline.

My Read on the Week

1. The silicon claims-to-SKUs transition is the quarter’s real story. Issue 1 asked whether Chinese silicon could serve these architectures at production quality; the answer is arriving as products rather than proofs — Alibaba’s named chip, Huawei’s recent unveilings (per AP’s own context), Zhipu’s still-unaudited serving claim. Each layer is separately unverifiable today, which is precisely why our ledger keeps them in different tiers: an attributed CEO quote is not a spec sheet is not an audit. The engineering question underneath hasn’t changed since Issue 1 — memory bandwidth per token — and FP4-lane convergence between model design (MXFP4 KV) and chip claims is the first structural signal that the stack is co-designing.

2. The open document stack is assembling itself, niche by niche. Parsing (GLM-OCR, MIT; WeVisDoc, Apache-2.0) last week; retrieval (EVIE, Apache-2.0) this week — all permissive, all small enough to deploy, all from different Chinese ecosystems. This is the open-weights dividend compounding where enterprises actually spend: boring document plumbing. Two weeks makes a pattern; a month makes a stack.

3. The launch storm’s real battlefield is the invoice. The US window shipped three flagship refreshes in 48 hours, and their durable news was economic: Luna’s $0.01 cache reads landing against DeepSeek’s $0.003 cache-hit floor (Issue 4); Opus 5.5 making reasoning undeletable — a billing decision wearing a behavior costume — the same week it shipped cache-preserving inline tools; Grok 4.7 holding price while its own comparator table shows a competitor leading. Every desk’s sharpest item this window metered something; the capability race has moved to the cache line.

What I’d Watch Next

  1. The Zhenwu V900 spec sheet and availability — the moment an Alibaba primary publishes absolute numbers, the 3×-M890 claim becomes checkable and the press-reported 216 GB / FP4 figures resolve one way or the other. HIGH-cost, unverified, top of the list.
  2. EVIE’s paper and the ViDoRe leaderboard — externally verify the #1 position, and read the paper when it lands (the weight-mix training detail deserves methodology).
  3. Migration week, US side — gemini-omni-flash-preview stops September 30 and antigravity-preview-05-2026 stops October 5: anyone running agent harnesses should re-pin and re-baseline in one pass this week, before the calendar does it for them.

One Thing I’d Experiment With This Week

Assemble the open document stack end-to-end and measure where it breaks. Every component of a Chinese-lab RAG pipeline is now permissively licensed: parse with GLM-OCR or WeVisDoc (Issue 6), embed and retrieve with EVIE, reason with any MIT Apache-licensed model in our ledger. Nobody has published the stack working together.

  • Setup: a fixed corpus of 500 mixed-quality documents — clean digital, degraded scans, multi-language — through the full pipeline: parse → chunk → EVIE multi-vector index → query → generation. Same corpus through a Western-baseline stack (your current parser + a standard bi-encoder retriever) as control.
  • Conditions: two retrieval regimes — clean parses only vs. including degraded-document output (where WeVisDoc’s 69.08 real-degraded track and EVIE’s leaderboard position should either shine or collide); log parse-failure rates per document class.
  • Measure: end-to-end answer accuracy, retrieval recall@10 per document class, total pipeline cost per 1,000 documents — and licensing friction (should be zero everywhere; verify each license field, per the house rule). Decision output: the first published performance-and-cost table for the fully-open China document stack — and the degradation curve, which is where every parsing benchmark says pipelines actually fail.

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/2026-09-23/.

Footnotes

  1. Three vendor launch tables in 48 hours (xAI’s Grok 4.7 comparators, OpenAI’s GPT-6 tiers, Anthropic’s Opus 5.5 entry) each pick their own comparator generations — Fable 5.1 / GPT-5.6 Sol here, Opus 5 vs 5.5 there — and none is independently reproduced. Terminal-Bench score labels carry their generation (2.1 vs 4.0 are different tests; Issue 4’s cliff example), per EDITORIAL.md §6.↩︎

Citation

BibTeX citation:
@online{dimsum_editorial2026,
  author = {dimsum editorial, the},
  title = {China’s Silicon Story Graduated from Claims to {SKUs} —
    Receipts Pending},
  date = {2026-09-23},
  url = {https://news.llm101.in/issues/2026-09-23/},
  langid = {en}
}
For attribution, please cite this work as:
dimsum editorial, the. 2026. “China’s Silicon Story Graduated from Claims to SKUs — Receipts Pending.” September 23. https://news.llm101.in/issues/2026-09-23/.