AI Brief, 30 August 2026: Tencent's 770B open model, benchmarked by Tencent

Saturday was quiet, and the biggest open-weights release of the week is one this brief missed. On 28 August at 08:48 UTC, Tencent announced and open-sourced Hy4 preview: 770 billion total parameters, 49 billion active per token, a one-million-token context window, Apache 2.0, weights on Hugging Face. The first Upload folder commit on the Hugging Face repository lands at 2026-08-27T14:44:58Z, the day before the announcement. Yesterday's issue went out at 04:53 UTC on 29 August and did not carry it. That is late by a day and worth saying plainly. Hy4 preview is 2.6 times the size of Tencent's own Hy3 from July, and larger than Z.ai's GLM-5.3, which this brief covered on Friday at 753 billion parameters. It is not the month's biggest download: Qwen's 2.4-trillion-parameter Qwen3.8-2.4T-A95B landed on 8 August at 4.89 TB. But that one ships under a bespoke licence, and Hy4 preview ships under Apache 2.0.

The interesting part is not the parameter count, it is the provenance of the numbers next to it. Tencent published a twelve-panel benchmark chart and a fuller appendix table, and Hy4 preview does not take the top position in a single one of the twelve panels. It leads the other open-weights models in five of them. Every competitor figure in that chart is marked with an asterisk, which Tencent's own footnote defines as a result from its own testing rather than the vendor's published number, and in several rows Tencent's measurement of a rival is materially below what that rival claims. Two rows in the appendix are exceptions, carrying official third-party scores for everybody: CritPt, where Hy4 preview places last of the seven current models, and GDPval-AA V2, where it places fifth of seven. No independent evaluation of the model exists yet. Artificial Analysis has not added it to any index.

Separately, and squarely inside the window, Sony Music Publishing and Warner Chappell sued Anthropic in the Northern District of California on 28 August, and named Dario Amodei and Benjamin Mann personally as defendants alongside the company. I read the 48-page complaint rather than the coverage. The publishers build their case on the findings in the earlier Bartz class action, which Anthropic settled for $1.5 billion, and argue that a settlement of that size has become a cost of doing business at a company the complaint says is eyeing a $2 trillion valuation. They seek up to $150,000 per work infringed.

The third item is a governance decision that other projects will cite. Debian's developers voted to permit the responsible use of generative AI in the project, choosing that option out of an eight-option ballot; the result was posted on 29 August. Nothing else of substance shipped. No frontier lab published a model or an announcement in the last 24 hours, arXiv had no Sunday batch, and no serving or training repository cut a stable release.

• Tencent's Hy4 preview: 770B total, 49B active, 1M context, Apache 2.0, 1.56 TB of bf16 weights, announced 28 August and missed by yesterday's issue
• It tops none of the twelve benchmarks in Tencent's own launch chart, and leads the other open-weights models in five
• Every rival figure in that chart is Tencent's own measurement, not the vendor's; on the two official rows Hy4 comes last (CritPt) and fifth of seven (GDPval-AA V2)
• Sony Music Publishing and Warner Chappell sued Anthropic on 28 August in N.D. Cal., naming Amodei and Mann personally and seeking up to $150,000 per work
• Debian voted to allow "responsible use of generative AI", from an eight-option ballot, result posted 29 August
• Nothing else: no lab release, no arXiv batch, no stable release from any serving repo

Tencent's Hy4 preview: 770 billion parameters and no benchmark lead #

What shipped. Weights for tencent/Hy4-preview and an FP8 sibling went up on Hugging Face on 27 August, and Tencent announced them at 08:48 UTC on 28 August. Both are Apache 2.0. The bf16 repository is 1.56 TB across 131 safetensors shards; the FP8 repository is 814 GB. The model is text-only, with no vision input. Serving recipes for vLLM and SGLang both resolve, so day-one support is real rather than aspirational.

For scale: Tencent's previous flagship, Hy3, shipped in July at 295B total and 21B active with a 256k context, and its Hugging Face repository is 598 GB. Hy4 preview is 2.6 times the parameters, 2.3 times the active parameters and four times the context.

The architecture is where the engineering is. The backbone is 78 layers. Exactly one of them, layer 0, uses a dense feed-forward network; the other 77 are mixture-of-experts with 256 routed experts and 1 shared expert, of which each token activates the top 8 routed experts plus the shared one. A separate multi-token-prediction layer, 10B parameters with 0.7B active, is built in for speculative decoding. The model card credits DeepSeek and GLM as the inspiration, and the attention is Gated DeepSeek Sparse Attention combined with IndexCache for cross-layer index reuse.

Those two mechanisms are what make a million-token context tractable, and they are worth unpacking because config.json states both exactly.

The first is multi-head latent attention. Rather than caching a key and a value vector per head per token, the model caches one compressed latent plus a small rotary component. Let $d_c$ be the latent rank and $d_r$ the rotary dimension. Hy4 sets $d_c=512$ and $d_r=64$ , so the cache holds

$$d_c+d_r=512+64=576$$

values per token per layer. Decompressed, each of the $n_h=64$ heads would need $d_{qk}=192+64=256$ key dimensions and $d_v=256$ value dimensions, or $n_h(d_{qk}+d_v)=64\times 512=\mathrm{32,768}$ values per token per layer. The ratio is

$$\frac{n_h(d_{qk}+d_v)}{d_c+d_r}=\frac{\mathrm{32,768}}{576}\approx 57$$

Concretely, in bf16 across all 78 layers that is 87.8 KiB of cache per token instead of 5.11 MB. Fill the full million-token window and you are holding about 94 GB of KV cache rather than 5.4 TB. The first number fits on a node; the second does not exist.

The second mechanism is the sparse index. Each query attends to at most index_topk = 2048 keys no matter how long the sequence is, so attention cost grows as $O(nk)$ with $k=2048$ rather than $O(n^2)$ . At the full 1,048,576-token context that is $\mathrm{1,048,576}/2048=512$ times fewer query-key pairs than dense attention would compute.

Picking those 2048 keys is itself work, done by a 32-head indexer, and this is where IndexCache earns its place. The indexer_types array in config.json has 78 entries, and only 21 of them read full. The other 57 read shared: those layers reuse the index their nearest preceding full layer computed instead of recomputing it. The pattern is layers 0 and 1, then every fourth layer thereafter.

So roughly three quarters of the indexer's work is skipped. The residual path is also non-standard: hc_mult is 4, meaning identity hyper-connections carry four parallel residual streams rather than one, which is the kind of change that only shows up when you read the config.

The benchmarks, and who measured them. Tencent published twelve headline panels and a much fuller appendix. Hy4 preview does not hold the top position in any of the twelve. It does beat every other open-weights model in five of them: SWE Atlas Refactoring (53.3), PostTrainBench (35.6), OneMillionBench with tools (65.4), BioMysteryBench (71.3) and HorizonMath pass@4 (8.8). In the other seven, GLM 5.3 or Kimi K3 is ahead. In all twelve, Claude Opus 5 or GPT 5.6 Sol takes the top slot.

The appendix footnote is the sentence that reframes the whole chart: results marked with an asterisk are from Tencent's own testing, and every competitor figure in the twelve headline panels carries one. Where Tencent's run and the vendor's published number both appear, they often disagree, and usually not in the vendor's favour. On Terminal-Bench 2.1, DeepSeek V4 Pro claims 87.9 and Tencent measured 80.3; Kimi K3 claims 88.3 and Tencent measured 85.7. On DeepSWE, GPT 5.6 Sol claims 72.7 and Tencent measured 68.9. The chart uses the lower number in each case.

That is not necessarily bad faith, and Tencent documents its harnesses in more detail than most: Terminal-Bench 2.1 was run through the Claude Code harness with a 500-turn cap and a 12-hour per-trial timeout, SWE-Atlas with a 256-turn budget and a network allowlist added "to prevent hacking behaviors", BioMysteryBench with answers judged by a competitor's model, Kimi K3. Tencent also discloses that on high-setting runs, Claude Opus 5 hit truncation rates of 28.76% on HorizonMath and 17.17% on BrokenArXiv, which is candid, and which undercuts the HorizonMath panel where Hy4's 8.8 beats Claude's 5.3. Publishing that caveat is to Tencent's credit; it is also a reason not to read that particular win as a win.

The two appendix rows carrying no asterisks at all are the ones to trust most, because every number in them is an official score rather than anyone's reproduction.

The other unasterisked row is GDPval-AA V2, an Elo score, where Hy4 preview's 1678 sits above DeepSeek V4 Pro (1580) and Kimi K3 (1675) and below Qwen 3.8 Max (1717), GPT 5.6 Sol (1711), GLM 5.3 (1763) and Claude Opus 5 (1831). Fifth of seven.

Caveats. Every benchmark figure above is Tencent's, including the appendix. There is no independent measurement of Hy4 preview at all: it does not appear anywhere in Artificial Analysis's changelog, whose most recent entries as of 29 August cover GLM-5.3-Flash and Qwen3.8-Flash-Next. Tencent also restates its own Hy3 baseline, noting in the appendix that some Hy3 scores differ from previously reported results after updates to the harness, judge model and anti-hacking measures, so the generation-over-generation jumps are measured against a moved goalpost. And Tencent is unusually blunt about the model's state: the card lists spending longer than necessary on reasoning and over-verifying its own work as known issues, and calls this an early version shipped deliberately early.

The first hands-on. Simon Willison posted his notes at 23:53 UTC on 29 August, which is the only independent look at the model published inside this brief's window. He read the chat template and found only two reasoning settings, high (the default) and no_think, ran his pelican-on-a-bicycle prompt through OpenRouter, and noted the reasoning trace's clipped, ungrammatical English. It is an impression rather than an evaluation, and he does not present it as more than that.

Sony and Warner Chappell sue Anthropic, and name its founders #

What was filed. On 28 August, roughly forty music publishing entities led by Sony Music Publishing (US) LLC and Warner Chappell Music, Inc. filed a copyright complaint in the Northern District of California, case number 5:26-cv-09217, with a jury demanded. The defendants are Anthropic PBC, Dario Amodei and Benjamin Mann. The complaint runs 48 pages with two exhibits, and the full text is on CourtListener, which is where I read it rather than relying on the coverage.

The theory. Rather than argue training-data fair use from scratch, the publishers build on findings already made against Anthropic. They cite the Bartz court's characterisation of Anthropic's conduct as piracy at massive scale, and its findings that Anthropic torrented at least five million pirated books from Library Genesis in June 2021 and at least two million more from Pirate Library Mirror in July 2022. Anthropic settled Bartz for $1.5 billion. The publishers' argument is that the settlement functioned as a cost of doing business, and they point to reporting that Anthropic is pursuing a $2 trillion valuation in an October IPO to make the point about scale.

The distinctive move is the personal liability claim. The complaint alleges that Mann himself used BitTorrent to download the LibGen corpus and directed other employees to do the same for PiLiMi, and that Amodei authorised and directed it. It also presses a distribution theory alongside reproduction, on the grounds that BitTorrent uploads every piece it downloads, so each torrented work is alleged to infringe two exclusive rights rather than one. Separate counts cover the outputs Claude generates and the removal or alteration of copyright management information.

What is sought. Statutory damages of up to $150,000 per work infringed under 17 U.S.C. § 504(c), or actual damages and profits at the publishers' election; up to $25,000 per violation for the CMI claims under § 1203(c)(3)(B); and a permanent injunction that would bind Amodei and Mann by name alongside the company. The works listed in the complaint's opening include "Ain't No Mountain High Enough", "Eye of the Tiger", "Livin' On a Prayer" and "Hallelujah", with the full schedule in Exhibit A.

Caveats. This is a complaint, which is one side's account and has been tested by nobody. Anthropic has not answered and no schedule has been set. The Bartz findings the publishers lean on are real and on the record, but whether they carry over to musical compositions, and whether two founders can be held personally liable for corporate acquisition decisions, are exactly the questions the case exists to decide.

Debian settles its LLM policy, and the answer is "responsibly" #

Debian's developers voted on a General Resolution covering the use of large language models in the project. Voting ran from 15 August to 23:59:59 UTC on 28 August, and the result was posted on 29 August: the winner is Choice 5, "Responsible Use of Generative AI".

The ballot itself is the interesting artifact. It carried eight options, spanning from "No LLM contributions to Debian via Social Contract", which would have amended the Social Contract to forbid them outright, through conditional-acceptance positions, to "Debian is created by humans" and one arguing the climate cost alone is disqualifying. The proposal that won neither endorses nor prohibits generative AI tools, and instead holds that contributions must meet the same standards of quality, correctness, maintainability and legal compliance however they were produced, with the contributor remaining responsible for understanding, reviewing and testing anything a tool produced.

That is a liability-allocation rule rather than a technology rule, and it is the shape most large projects will probably converge on, because it is the only one that does not require detecting whether a patch was machine-written. The anti-LLM proposals raised licensing uncertainty as their strongest argument, which the winning text does not resolve so much as push onto the individual contributor.

I could not retrieve the numeric tally: debian.org/vote/2026/vote_002_results returns 404 and the August mailing-list archives returned nothing to the fetches I made, so the option ordering below first place is not something I can state. LWN's comment thread asserts that the two strongest anti-LLM proposals finished below "None of the above", but that is a reader's claim in a comment, not a result I verified.

Also notable #

• Anthropic is reported to be changing Claude Code's weekly limits. Per the @ClaudeDevs account on X, relayed by Techmeme at 18:05 UTC on 29 August, weekly limits rise a "permanent" 25% on 14 September for most plans. Because a temporary 50% boost is currently in force, that works out to roughly a 17% reduction against today's ceiling. Unconfirmed: x.com is unreadable from this environment, so I could not open the original post and am relying on the aggregator's summary of it.
• Cursor responded to this week's OpenAI shutoff. Co-founder Michael Truell said OpenAI accounts for about 5% of Cursor's traffic and that Cursor had trusted OpenAI to remain neutral, per The Information via Techmeme at 14:15 UTC. The Information is paywalled here and I did not read the article, so the 5% figure is theirs and not one I could check.
• No stable release from any serving or training repository. llama.cpp cut four tags inside the window, b10686 through b10689, the last at 04:02 UTC today, but all four carry prerelease: true and are per-commit CI builds. vLLM's v0.28.0 is dated 26 August, not 29 as its Hacker News submission time suggests, and was already available before the previous two issues.
• Sources I could not verify. mistral.ai/rss.xml returned an empty body, api-docs.deepseek.com/updates served a 432-byte client-rendered shell, and qwen.ai/research remains unreadable. Those three labs are unverified for this window, not confirmed silent. Two Hugging Face repositories that would otherwise have been worth dating, google/timesfm-3.0-pytorch and nvidia/Nemotron-3-Diarization-preview, are now gated and their commit histories are not readable without a token.

What to watch #

• Whether any independent evaluator publishes a Hy4 preview score. Artificial Analysis carries nothing for it yet, and until one appears every number about this model comes from Tencent.
• Whether Anthropic answers the music publishers' complaint by moving to dismiss the claims against Amodei and Mann personally. That motion, if it comes, is the first real test of whether founder liability survives.
• Whether anyone reproduces the IndexCache result independently. Reusing one sparse index across four layers is a large claimed saving, and the model card asserts it without publishing an ablation.
• arXiv's Monday batch announces at 00:00 UTC and will carry Thursday and Friday submissions, the first new papers since 28 August.