The message didn’t matter — until it did
A Tencent- and Huawei-affiliated paper audits latent communication between models, finds prior gains were interface artifacts, and ships a 1.05M-parameter interface where the content demonstrably carries; Japan re-enters twice — an all-post-training Apache release and a ¥1B-gated community license.
Draft-KV’s control experiments show prior latent-communication gains move ≤0.60 pp when the message is swapped — the interface, not the information, was doing the work — then trains a 1,048,584-parameter bridge that more than doubles a frozen 0.5B receiver. Plus: NII’s LLM-jp-4.1 ships as a post-training-only Apache release, and PLaMo’s community license turns out to carry a ¥1B revenue gate. Issue 9, coverage window 28–30 September 2026.
Week of September 28–30, 2026 · Issue 9
The Week in One Sentence
A three-day window with China’s labs quiet produced the field-audit paper of the season — swap the message, watch the gains vanish — OpenAI’s newest generation changed exactly one price (cache reads, halved), Sarvam pulled its own two-day-old cloning skill, and Japan re-entered the issue at both ends of the open-weights spectrum: an Apache-licensed post-training-only national-lab release, and a community license whose ¥1B revenue gate wants its name on everything you build.
The Big Story: Draft-KV — the audit before the method
What happened. (China.) arXiv:2609.34754, submitted September 28 and surfaced via the Tencent org’s paper feed — read here in full (Wu et al. 2026). The setup: “latent communication” passes a sharer model’s internal states directly to a receiver model instead of going through text, and reported results have been striking. The audit first: the authors ran the control experiment the field hadn’t — swap the transmitted message for one from a completely unrelated question. If the states carry information, accuracy should crater. It barely moves: at most 0.60 percentage points across five method–dataset pairs, against system gains as large as 15.44 points. A frozen C2C interface fed deranged messages retained 93–96% of its matched accuracy, and scaling the sharer from 0.6B to 8B raised its own accuracy 29.85 points while system accuracy rose 3.54. The prior methods’ “communication” was mostly the receiver adapting to an interface — not information transfer.
Under the hood. Draft-KV’s fix transmits the right states: the KV tensors formed while the sharer drafts its answer (draft positions, pre-RoPE, visibility-masked — under causal attention, prompt KV is unchanged by the answer, so draft states are where question-plus-reasoning actually lives). Two bias-free linear maps per layer project these into the receiver’s head layout, and — the design choice that matters — the packet lives in separate side memory, never merged into the receiver’s self-attention cache, read through a gated attention branch that reuses the receiver’s frozen projections. Training is progressive (hidden-message reconstruction → answer alignment → answer-text with a hinge guard penalizing packets that would hurt on mismatched inputs), and everything stays frozen except the interface: 1,048,584 trainable parameters — 348× fewer than C2C, with size depending only on the models’ KV widths.
What the evidence says (vendor-run, with the counter-evidence carried). A frozen Qwen2.5-0.5B receiver hits 78.04% on MMLU-Redux with a Qwen3-8B sharer versus 37.45% alone — more than doubling, and exceeding the sharer’s own 75.04. It beats receiver-alone in 35/35 public settings and text-to-text in 33/35 (+5.85 average), transfers to held-out benchmarks (Stage 3 trains on ARC only), and — the honesty signal — deranged packets drop the receiver to 36.40%, below its unaided baseline: in this method the content demonstrably carries the gain. The gaps we carry because the paper doesn’t (there is no explicit limitations section): mismatch harm reaches 11.65 points in one setting; results are point estimates without significance testing; part of the C2C audit relies on a single released checkpoint; and the receiver substrate is prior-generation Qwen2.5.
Why this matters. Two of this publication’s standing theses meet here. First, the audit is §0.6 methodology — carry the counter-evidence — turned into a control experiment, and it found a small field’s headline numbers resting on an artifact. Every multi-model pipeline passing hidden states (agent swarms, router ensembles, speculative stacks) should now demand the deranged-message control before trusting a “communication” result. Second, the parameter story keeps collapsing in one direction: frozen models plus tiny trainable interfaces — SAS’s 33M gates last cycle, 1.05M here — with the side-memory-not-cache design echoing the same decouple-memory-from-computation principle as DeepSeek’s Engram. The Chinese-lab ecosystem isn’t just shipping models; it’s shipping the interconnects between them, and the interconnects are now auditable.
What I would test. Replicate the deranged-message control on any latent-communication claim you’re building on — it is a one-day experiment — then test Draft-KV’s failure mode: the 11.65-point mismatch-harm case suggests the interface can actively mislead when the sharer is wrong; measure how often a wrong draft poisons the receiver versus merely failing to help.
Elsewhere in AI Engineering
No second China item met the bar — the sweep ran and the window was quiet: all five major orgs silent September 28–30, no V4.1 siblings, no GLM-5.5, no Qwen4. Two adjacent community papers surfaced on the same Tencent feed (visual state reconstruction; RAG reranking) and were deprioritized unread beyond titles. Continuity housekeeping: the Issue 6 v6.1 corrected claim re-verified unchanged (ExtractBench rows still on the V4.1-Flash card — no regression); SAS’s license absent for a fifth consecutive check with its repo now at 63 stars and still zero reproduction activity; the hosted Qwen3.8-Flash card returned 401 for the ninth time; and the Zhenwu V900 primary spec sheet remains unfound, with the search backend down until October 13 limiting this cycle’s rechecks — all carried, all dated.
Sarvam dropped its own voice-cloning skill — two days after the API (India) — Worth Watching
What happened. (India. Labelled update on Issue 8’s coverage of the Dubbing and Voice Cloning APIs.) On September 28, Sarvam’s official skills repo merged two maintenance PRs: a sync with docs.sarvam.ai (“fix deprecations, add missing coverage”), then a pass that “live-verify the recent additions, fix 6 real bugs, drop voice-cloning” — the voice-cloning skill, one of those recent additions, was removed in the same commit that live-verified the rest. That is two days after the Voice Cloning API’s changelog entry (September 26; Issue 8). The same day, the cookbook merged a Voice Coding Assistant example whose review fix switched “from the deprecated sarvam-m model to sarvam-105b via the SDK.”
Why this matters. The drop reads two ways, and we label both: either the skill failed its own live-verification gate (a quality signal about the example, not necessarily the API), or the cloning surface itself is being reworked — Issue 8 flagged that the API shipped with no consent-verification terms visible, and a vendor pulling its own cloning skill two days post-launch is at minimum a governance-adjacent signal; the commit message does not say which. The sarvam-m detail is cleaner: the repo’s own wording calls the 24B open-weights model (May 2025’s hybrid) deprecated, with an official example migrating to the hosted 105B — if you serve those weights, the direction of travel is now documented, and we are watching for a formal retirement notice before treating dates as knowable. The meta-signal rhymes with the platform’s month: “live-verify the recent additions, fix 6 real bugs” is example code held to an execution bar — the agents-need-CI thesis applied to the vendor’s own repo.
Go deeper: skills PR #22 · cookbook PR #41
GPT-6.1 Sol priced the cache, again (US) — Worth Testing
What happened. September 29, OpenAI announced GPT-6.1 Sol — “near-Astra intelligence for a fifth of the price,” in the announcement’s own framing (the page declines our fetcher, as always; the rate card, read today, is the primary here). The arithmetic is exact: gpt-6.1-sol at $2.00 / $10.00 per million (input/output, short context ≤272K) versus gpt-6-astra’s $10 / $50 — one-fifth on both lanes; long context (>272K) runs $4 / $15; batch and flex halve it; “fast” doubles it. And the generation-over-generation change is a single column: gpt-6.1-sol is identical to gpt-6-sol except cached input, halved to $0.10. The card now also defines the short/long cutoff numerically (272K), carries an Astra-only “Ultrafast” tier ($60 in / $300 out, short), and dates GPT-5.6 Sol’s promo “at least through November 21, 2026.”
Why it matters. When a model generation’s only pricing delta is cheaper cache reads, the vendor is telling you where it thinks the workloads live: long stable prefixes, re-read agent contexts — the same economics this publication has tracked since DeepSeek’s $0.003 cache floor. “Near-Astra” is a vendor claim with zero independent placement yet (Real-SWE, re-read this morning, has no 6.1 entry); what is verifiable today is the price structure. If the claim survives the instruments, the mid-tier just became the default agent build — until then, pin your Astra spend and watch the leaderboards, not the launch posts.
Go deeper: pricing · HN discussion
Sonnet 5.5 shipped five breaking changes — it’s migration week (US) — Worth Reading
What happened. September 28, per the platform changelog: claude-sonnet-5-5, live across the Claude API, Bedrock, Google Cloud, and Foundry — “about 30% faster and up to 30% cheaper for most workloads” over Sonnet 5, in the announcement’s workload framing (per-token rates were not restated in what we read, and none are inferred here). The operational content is the five breaking changes: turning thinking off now requires thinking: {"type": "between_tools"} instead of "disabled"; tool_choice any/tool return 400; thinking blocks are tied to the model and conversation; the computer_20251124 tool is rejected on the Claude API and Google Cloud; and the advisor tool refuses Opus 4.8, Opus 4.7, and Sonnet 5 as advisors. Plus account-bound thinking blocks: foreign-account blocks are silently dropped and the request still succeeds.
Why it matters. This is the second 5.5-generation launch in six days to ship breaking changes (Opus 5.5 made thinking undeletable on September 22), and thinking: "disabled" now means three different things across the current lineup — a 400 on Opus 5.5, between_tools on Sonnet 5.5, plain off on earlier models. Multi-model fleets need per-model request builders, not shared ones. And it is migration week cross-vendor: Google retired gemini-omni-flash-preview today — the deprecations page, updated this morning, no longer lists it; gemini-omni-1.1-flash stands alone — and antigravity-preview-05-2026, its listing now showing a May 19 release date, stops October 5. Re-pin, re-test, and re-price in one pass.
Go deeper: changelog · announcement
Eleven v4 made voice cloning stronger the week an Indian vendor pulled theirs (US) — Worth Watching
What happened. September 28, ElevenLabs’ blog listing announced Eleven v4, “our most emotive model” — “emotional nuance, faster response times, and stronger voice cloning to text-to-speech,” available via ElevenAgents, ElevenCreative, and the API, with an Eleven v4 Turbo variant alongside. Everything here is listing-depth: the post itself did not resolve for our fetcher, and no pricing or latency figures are cited. (Labelled context: one week earlier, the same team’s Interspeech paper — arXiv:2609.25411, September 21 — showed learnable null embeddings beating fixed ones across speaker similarity, stability, and expressiveness in CFG-based speech synthesis. Whether that technique ships in v4 is stated nowhere we can read; the timing is the only link, and we label it as such.)
Why it matters. The same week, an Indian vendor pulled its own voice-cloning skill two days after launch (Sarvam, in the item above) while ElevenLabs shipped “stronger voice cloning” as a headline feature — two markets moving in opposite directions on the same capability, and the consent-verification gap this publication flagged in Issue 8 remains unpublished on both sides. For builders, a v4-plus-Turbo pair spanning emotive and latency tiers keeps variant choice the deciding axis — pin which one your latency budget actually needs before the invoice chooses for you.
Go deeper: blog listing · paper
LLM-jp-4.1: a national lab ships a post-training-only release — bases untouched (Japan) — Worth Reading
What happened. (Japan. Dating labelled: the repos were created September 15–24 but absent from this desk’s September 27 sweep; the technical blog is dated September 28 and the cards’ last modification is September 28 08:07 UTC — the series became publicly verifiable September 28, inside this window.) NII’s LLM-jp released LLM-jp-4.1: 8B and 33B dense plus a 32B-A3B MoE (128 experts, 8 activated, 3.83B active — 65,536-token context across the board), each as a -thinking variant with day-one GGUFs, plus the SFT and DPO datasets. Apache-2.0 (card license section links the canonical text; no per-repo LICENSE file ships — noted). The headline engineering decision: the base models are unchanged from 4.0 — every gain in 4.1 is post-training.
Technical signal. Per the technical blog (Japanese, read this cycle): SFT data grew ~5.4B → ~12B tokens (~3× total learning), rebuilt on Nemotron-derived STEM and agentic subsets with all responses regenerated by gpt-oss-120b — OSAID-compliant subsets only, with Olmo-3 SFT data explicitly refused over licensing terms despite the AIME gains it brought. Tool calling ships through a modified OpenAI-Harmony chat template (parallel calls via <|end|>; the card warns plain openai-harmony tokenization doesn’t work). A developer-message brevity instruction roughly halves final-answer tokens. Alignment is judge-DPO with no RL anywhere: gpt-oss-120b scoring four sampled candidates into preference pairs — 305,831 pairs for the 8B, 269,287 for the MoE, 89,023 for the 33B (β=0.5, two epochs). Vendor-run results with their own counter-evidence carried: MT-Bench roughly flat-to-down (attributed to the conciseness push); swallow-eval has the 33B beating Olmo-3.1-32B-Think and Qwen3-32B at roughly gpt-oss-120b level but behind Qwen3.8-27B, Gemma-4-31B, and Muse-Glimmer-30B; tool calling is the strength — BFCL-nejumi multi-turn has the 33B above Qwen3-32B and Qwen3.5-9B; GPQA/AIME improved yet trail open-weight peers. The blog also reports its own self-reference filter backfiring — rewrites of “ChatGPT” to “LLM-jp-4” lifted self-recognition from 8/24 to 23/24 while causing OpenAI/LLM-jp confusion, v2 pending.
Why it matters. “Same base, gains entirely post-training” is the exact claim GLM made for 5.3-over-5.2 (Issue 3) — now executed by a national lab, with the training data public and the judge model named. The governance internals are the memorable part: OSAID discipline that refuses a benchmark-winning dataset on licensing grounds, and honest reporting of an alignment fix that misfired. For engineers, a fully-Apache Japanese family with multi-turn tool-calling strength and day-one GGUFs is directly deployable — and every benchmark above is vendor-run; the natural first check is whether NII’s own Open Japanese LLM Leaderboard adds the 4.1s (an NII-affiliated instrument — label it accordingly; it hasn’t yet).
Go deeper: model collection · tech blog · cookbook
The PLaMo Community License, read at last — a ¥1B gate and a naming requirement (Japan) — Worth Reading
What happened. (Japan. Labelled update on Issue 8’s HIGH+U flag.) Issue 8 covered PLaMo-3-610M-Fin-Instruct with its license flagged “Worth Verifying” — the terms sat on a JavaScript-only page no desk fetcher could render. This cycle they were read from the page’s own shipped web bundle (the strings the site renders), verbatim. The grant is non-exclusive, worldwide, non-transferable, royalty-free — conditioned on “the USER’s or the USER’s affiliated company’s annual revenue or sales must not exceed 1 billion yen” (≈US$7M); larger users route to a contact form. Recipients must visibly display “Built with PLaMo”, and — the unusual clause — derived models must include “PLaMo” in their names. Output belongs to the user, but models trained on the output are themselves DERIVED MODELS under the same terms, extending the license through distillation generations. PFN may modify the agreement with notice; Japanese law, Tokyo District Court. (Method note: extracted from the vendor’s shipped bundle rather than a rendered page — human review should confirm in a browser before legal reliance.)
Technical signal. The serving story matured in the same window: vLLM PR #56575 — --reasoning-parser plamo3 and --tool-call-parser plamo3, the nested control markers the 610M emits — merged into vLLM mainline on September 29. PLaMo-3 serving no longer depends on the vendor’s fork; the 610M’s downloads rose 107 → 277 over the week, still with zero benchmark tables published.
Why it matters. Put the three gates side by side: Kimi K3’s $20M MaaS-revenue trigger (Issue 1), GLM-5.3’s $10B aggregate-MaaS gate (Issue 4), and now PLaMo’s ¥1B (~$7M) any-revenue ceiling — three orders of magnitude apart, and PLaMo’s is the strictest while gating all use, not a deployment mode. In the same window, NII’s 4.1 shipped Apache-2.0 (item above): Japan’s two national-scale open releases of the week sit at opposite ends of the licensing spectrum — one permissive to the point of anonymity, one that wants its name on everything you build. For the 610M finance model specifically, the practical reading: any company above ~$7M revenue — most companies that would deploy a finance model — is not licensed under these terms.
Go deeper: license page (requires JavaScript) · vLLM PR #56575
Research Radar
The window’s research story leads the issue as the Big Story — Draft-KV, read in full (paper-notes in the research packet). No artifact release exists yet (bare paper); the code watch is on the ledger.
Repo Radar
No new China repo merited entry this cycle. The SAS repo was rechecked (fifth license check: still no LICENSE, zero issues, zero pull requests) — an artifact this cited with no license and no reproduction attempt is quietly becoming its own datapoint about uptake speed in this ecosystem.
My Read on the Week
(Synthesis and interpretation — not reported fact. Written after a cross-item pass spanning all four desks’ items.)
1. The audit became the field’s unit of trust. Draft-KV’s deranged-message control (does the claimed channel carry the claimed information?) is the same method as Issue 8’s found-on-the-card ExtractBench tables and Issue 5’s SAS failure tables — and this window shows vendors adopting it inward: Sarvam’s skill drop came from a live-verification pass on its own examples, and NII’s blog reports its own self-reference filter misfiring rather than hiding it. When labs start auditing themselves in public and papers start their method sections with the control experiment, “no control shown” becomes the new “no benchmark shown” — our default posture for unverified.
2. The bases froze; everything above them ships. Draft-KV’s 1,048,584-parameter interface between frozen models, SAS’s 33M gates, GLM’s post-training-only flagship claim (Issue 3), LLM-jp-4.1 with its bases deliberately untouched from 4.0 — and on the pricing side, GPT-6.1’s only per-token change is cheaper cache reads. The same idea is compounding at every layer: stop paying to move the base, pay only for the delta — the interface, the post-training, the cached prefix. This publication has tracked the prefix economy since DeepSeek’s $0.003 cache floor; this is the year it became the product.
3. Licensing is now a spectrum with named endpoints — and the consent gap runs orthogonal to it. One window, one country, two national-scale releases: Apache-2.0 with public training data at NII, and a ¥1B-capped, naming-requiring, output-inheriting community license at PFN — against the field’s existing gates ($20M MaaS at Kimi, $10B at GLM). Meanwhile ElevenLabs shipped “stronger voice cloning” the same week an Indian vendor pulled its own cloning skill, and consent-verification terms remain unpublished on both sides. The license and the consent policy are the two documents that now decide deployment as much as any benchmark — and only one of them gets benchmarked.
What I’d Watch Next
- Draft-KV’s artifact and replication — a bare paper today; the audit begs independent replication on other latent-communication methods, and on agent-to-agent protocols next.
- The October 13 search reset — the deferred backlog clears in one cycle: Zhenwu primary, ExtractBench provenance, MarathonMemBench, AA numeric reads, and the browser-render confirmation of the PLaMo license text.
- LLM-jp-4.1’s first independent placements — whether NII’s own Open Japanese LLM Leaderboard adds the 4.1s (NII-affiliated — label), and the first third-party check of the multi-turn tool-calling claim.
One Thing I’d Experiment With This Week
Run the deranged-message control on your own stack. If you operate any multi-model pipeline — a router feeding a reasoner, a drafter feeding a verifier, an ensemble voting on answers — you likely have at least one implicit “communication” path whose value you’ve never isolated.
- Setup: pick one hand-off in your pipeline. Instrument it so the downstream model’s input can be silently replaced with the upstream signal from an unrelated request of the same shape (the derangement). Freeze everything else.
- Conditions: production traffic replay for one hour in three arms — true hand-off, deranged hand-off, and no hand-off (downstream alone).
- Measure: outcome-quality delta per arm, latency, and cost. Decision output: one number — how much of your hand-off’s apparent value survives derangement. If nearly all of it does (the C2C pattern), your “smart pipeline” is an interface artifact and a simpler design would score the same; if it collapses (the Draft-KV pattern), your channel is real and worth optimizing — and now you have the harness to optimize it against. Either way, you’ve applied the week’s most transferable idea to your own system.
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/2026-09-30/.
References
Citation
@online{dimsum_editorial2026,
author = {dimsum editorial, the},
title = {The Message Didn’t Matter — Until It Did},
date = {2026-09-30},
url = {https://news.llm101.in/issues/2026-09-30/},
langid = {en}
}