AI Brief, 16 November 2026: a token-efficiency claim, and the index that measures the opposite
Alibaba published Qwen4-Max this morning, dated 16 November; it reached OpenRouter at 02:18 UTC and Hacker News at 02:14 UTC, where it has 209 points. The model is the flagship of a new Qwen4 series: a 2.4-trillion-parameter mixture-of-experts backbone with 71B active per token, a further 328B of what Alibaba calls N-gram Embedding parameters, native vision, and a one-million-token context window. It had been in preview since 11 November. The weights are not released, and the post closes with a promise that smaller Qwen4 models are "coming to open source".
The headline is not a capability score but an efficiency one. On Alibaba's internal Work360 benchmark, token usage falls roughly 24% and 35% at its two highest reasoning settings against Qwen3.8-Max, and on Terminal-Bench 4.0 roughly 48% and 49%. Average cost per Work360 task drops from $4.43 to $1.67, a 62% reduction. API prices are $1.50 per million input tokens and $4.50 per million output, with a 95% cache-read discount.
That claim is unusually checkable, because the one organisation measuring token consumption independently had already added the model. Artificial Analysis listed Qwen4-Max on 15 November and scores it 43 on its Intelligence Index, 38th of 81 — above the median of 41, well below the 65 it records for the leader. On the same page it ranks 65th of 80 on verbosity: the index run consumed 127 million output tokens against a median of 78.6 million. Cost per task came to $3.27, not the $1.67 the launch reports for its own benchmark.
These are not the same measurement: Alibaba's figures are generation-over-generation on a suite it built, while the index is a nine-evaluation composite scored across vendors. A model can use fewer tokens than its predecessor and more than its competitors. But efficiency is what the launch leads with, and the only outside reading of it puts the model among the more verbose systems measured. None of Alibaba's token figures have been independently reproduced.
One detail deserves more credit than this pattern usually gets. The comparison table is published as a single 5,400 by 3,600 pixel image with no numbers in text, the fourth launch in six weeks to ship its scorecard as a picture, a count this brief put at three in five weeks on the 12th. Yet the footnotes cite Artificial Analysis, the Mercor leaderboard, the Vals Index and Arena for six benchmark rows: several numbers in that image are third-party, which is better practice than the format suggests.
Elsewhere, the window's most interesting artifact came from outside the labs entirely: a full-duplex spoken dialogue model that listens and speaks in the same 133-millisecond step, from a lab with no press coverage and 2,098 downloads.
- Qwen4-Max: 2.4T backbone, 71B active, 328B N-gram Embedding, 1M context, API only. Independent Intelligence Index score 43 against a median of 41; 127M output tokens against a 78.6M median.
- Alibaba reports Work360 rising from 47.74 to 52.69 and per-task cost falling $4.43 to $1.67, both self-reported on a benchmark it built.
- Haru-4B, a full-duplex speech model from higurashi-ai, runs at 7.5 Hz with a one-step audio head. Apache-2.0, 321 likes and 2,098 downloads, no latency or quality number in its card.
- Three speech artifacts landed in one window: Haru-4B, Bland's commercial full-duplex model, and
speech-to-text support in
llama.cppnightlies. - LLM reviewers draw about half their stated paper weaknesses from the appendix; humans draw 8.4%. 2,000 papers, code released.
- Voxtral 3 weights still unreleased, day 27, nothing under
mistralaisince 15 October.
The architecture, and what the index measures instead
Qwen4-Max's interest is three changes to the sparse mixture-of-experts recipe, all aimed at long sequences rather than raw capacity.
The attention stack pairs Gated DeltaNet with what Alibaba calls Qwen Sparse Attention. Three DeltaNet layers sit beneath every sparse-attention layer and compress the history into a fixed-size state, so the sparse layer never reads the raw sequence. That layer then uses a lightweight indexer to pick context at micro-block granularity rather than per token, cutting the cost of deciding what to attend to as well as of attending. The pairing works because the halves fail differently: a fixed-size recurrent state loses detail but costs the same at any length, while sparse attention keeps detail but must pay to locate it.
Gated Residual gives each layer four parallel residual branches with learned read and write gates, credited in the post for a pre-training run with no loss spikes at the 2.4T scale. N-gram Embedding adds context-based lookups of local token patterns at the second layer, where those 328B extra parameters live: capacity that costs a table lookup rather than a matrix multiply.
Pre-training evidence is reported as bits per byte, the compression a base model achieves on held-out text, lower being better. Qwen4-Max-Base is 5 to 13% lower than Qwen3.8-Max-Base across knowledge, code, QA and STEM while activating 25% fewer parameters. Alibaba notes against its own interest that two of the five checkpoints compared are post-trained, so those columns are not comparable to a base model.
Forty-three places Qwen4-Max between Muse Wonder and Grok 4.7: creditable for a model priced at a quarter of what the leaders charge, and not a step change. The index's composition matters before anyone leans on it: version 5.0.1 is nine evaluations, two of them, AA-Briefcase v1.1 and AutomationBench-AA, the same ones Alibaba's footnotes cite. The context window it measured is 984k, not the advertised million.
Listening and speaking in the same 133 milliseconds
The window's most substantial open artifact is Haru-4B from higurashi-ai, uploaded 12 November and rising hard since: 321 likes and 2,098 downloads, the latter mattering more because it means people are running it. No publication has covered it.
It is a full-duplex spoken dialogue model: it listens and speaks at once instead of taking turns. The conventional pipeline is a chain — voice-activity detection decides the user has stopped, recognition transcribes, a language model answers, synthesis speaks. Each stage waits for the last, and the first is a guess about whether a pause ends a sentence.
Haru-4B collapses that chain. Every 133 ms it consumes one frame of the user's audio and emits one frame of its own audio plus one text token. The user's audio latent, the model's previous text token and its previous audio latent are summed into a single position of the language model, so listening and speaking share one sequence and neither waits.
The frame rate is where the design becomes concrete. The codec is causal and runs at 24 kHz with 3,200 samples per latent, so each frame covers
where the numerator is the samples the codec packs into one latent and the denominator the sampling rate. That 133 ms is the entire compute budget per step: one forward pass of a 32-layer, 2,560-hidden Qwen3.5-4B backbone plus the audio head, or the conversation falls behind. Audio is a 64-dimensional continuous latent produced by a 4-layer MLP with MeanFlow sampling in one step from the model's hidden state. There is no separate audio transformer and no discrete audio tokens, which is what makes the budget achievable: a multi-step diffusion head would need every sampling step to finish inside 133 ms.
The interface follows the same shape, a loop rather than a request:
from haru import DuplexSession, HaruModel
from haru.audio import load_audio, prepare_reference
model = HaruModel.from_pretrained("higurashi-ai/Haru-4B")
voice = prepare_reference(load_audio("voice.wav")) # 3s-3min reference sets the voice
session = DuplexSession(model, system_prompt=card, reference=voice)
frame = session.start() # first 133 ms: frame.audio and frame.text together
while not frame.done:
frame = session.step(user_audio) # 3200 samples in at 24 kHz, 3200 out
Two caveats. The weights are Apache-2.0 by both tag and card, but the model derives from Alibaba's Qwen3.5-4B and carries Microsoft's VibeVoice acoustic tokenizer, so upstream terms govern what a downstream user can do. And despite the name it is 5.27B parameters in a single 10.5 GB bfloat16 file. The authors describe it as a research base model that holds a conversation but is not a reliable source of facts.
It did not arrive alone, which is the better reason to take the week seriously. Bland launched Tofu-1, a
commercial full-duplex speech-to-speech model for calls, on 15 November, and llama.cpp nightlies on 15
and 16 November added multimodal support for two speech-to-text models, putting recognition inside the
runtime most local inference is built on. Two very small text-to-speech models are also getting real use at
opposite ends of the attention curve: tiny-voca, 4.54M
parameters in 18.2 MB, trained from scratch on a CC BY 4.0 dataset of 19.1 hours, has 85 likes and 40
downloads; kitten-tts-2-mini, 82M parameters as
82.7 MB of int8 ONNX across 16 languages, has 20 likes and 1,271 downloads. The ratios run opposite
ways, and downloads are the better signal.
What a machine reviewer complains about
A paper announced on 16 November, submitted on the 13th, measures something every reader who submits to a conference has a stake in. AI Reviews Are Biased Towards the Appendix, by Saptarshi Ghosh, Mohammad Saim and Tianyu Jiang, takes 2,000 papers from ACL, EMNLP, COLM and ICLR and asks not whether machine reviews agree with human ones but which part of the paper they criticise. The answer is lopsided.
Recent GPT and Claude models draw about 50% or more of their stated weaknesses from the appendix. Humans draw 8.4%. The split is not only in proportion but in kind: human weaknesses concern presentation and the core idea, while machine ones reach deep for fine-grained details not central to the contribution. The behaviour held across API, coding-agent and web access, making it a property of the models rather than of one harness. The code is released.
An appendix is where authors put what did not fit. It is now the most heavily weighted part of a submission for any reviewer using a model, an incentive nobody designed.
Also notable
-
vLLM 0.34.0 was tagged at 04:27 UTC on 16 November, the stable cut of the 0.34 line after four release candidates. The tag carries no highlights body, so the candidate notes are the only changelog.
-
Meituan's LongCat-2.5 drew coverage on 15 November for being trained entirely on Chinese accelerators, its self-hosting guide describing 194 Huawei Ascend chips and no Nvidia guidance at all. The MIT-licensed weights went up on 4 November, so the artifact predates the write-up; the hardware detail is new, and it sits against the checkpoint-forensics argument covered here on the 13th, which read six Chinese models as initialised on export-controlled Nvidia parts. Meituan's claim is self-reported.
-
OpenAI launched a standing science publication on 15 November, opening with a lattice formulation result in particle physics. Runway released Praxis-1 on 14 November, an open-weights video model repurposed as a robot controller.
-
A single-author paper announced on 16 November, arXiv:2611.17958, reports that injecting an off-topic concept into a reasoning model's activations overrides its belief about what it was asked, after which it rationalises the injected answer. No code.
-
Anthropic's IPO is reported as roughly seven times oversubscribed at up to $90.4B, secondhand; the filing was not traced.
What to watch
- Whether an independent token count is published for the benchmarks Alibaba used. Efficiency is the launch's centrepiece and the only outside reading points the other way. The decisive number is output tokens per task on a suite Alibaba did not build.
- Whether the promised smaller Qwen4 models ship as open weights. The post commits without a date. The useful check is whether they carry the same Gated DeltaNet and sparse attention stack or a cheaper variant, because the architecture is the interesting part and a distilled dense model would not test it.
- Whether anyone publishes a latency measurement for Haru-4B. Apache-2.0 weights, a 12 GB footprint, PyTorch and MLX implementations and 2,098 downloads mean this costs an afternoon. The figure worth having is end-to-end latency against the 133 ms frame budget, since the card offers nothing.
- Whether a vendor publishes a threshold-conditioned fail-open rate, day 38. Qwen4-Max ships reasoning effort levels and no abstention semantics, adding a fourth endpoint that cannot answer the question.
- Still open: Voxtral 3 weights, flagged daily since 21 October, day 27 with nothing under
mistralaisince 15 October. No independent number for Tencent's DSpine drafter or Nanbeige4.3-33B; no Edge0 ablation separating quantisation from post-training, asked on the 13th; no assessment from the three complexity theorists who agreed on 26 October to examine the Exponential Time Hypothesis manuscript; and no response from any company named in the Blackwell forensics of the 13th.