ZeroNoise Logo zeronoise
Post
Qwen4-Max Cuts Cost per Task While Staying Short of the Frontier, as Allies Push for Huang's Return to Nvidia
•
6 min read
• 535 docs
Alibaba's Qwen4-Max launch and how independent evaluators rated it lead this brief. It also covers reported efforts to bring Jensen Huang back to Nvidia, new US–China compute data, DeepSeek's free-tier limits, OpenAI's physics showcase and the latest arguments about AI risk.

Qwen4-Max: big efficiency gains, still behind the frontier

Alibaba launched Qwen4-Max, built on a new MoE architecture with native vision and a 1M-token context window . Alibaba says it beats Qwen3.8-Max using up to ~50% fewer tokens, and that average cost per task falls by more than 60% in its internal evaluation. Pricing is $1.5/M input and $4.5/M output tokens, with implicit caching 70% cheaper than before . The API is live on QwenCloud . Alibaba says open weights for smaller Qwen4 models will follow .

Independent evaluators published results within hours:

  • Artificial Analysis: Qwen4-Max scores 43 on the Intelligence Index, 9 points above Qwen3.8 Max and 3 below Kimi K3.1 Pro . It costs $3.27 per task, against $10.02 for Qwen3.8 Max and $3.68 for Kimi . It trails Kimi on Terminal-Bench 4.0 (48% vs 67%) and Tencent Hy4 on FrontierCode 2.0 (26% vs 31%) .
  • Agent Arena: Qwen4-Max debuts at #14 overall, up from #39 for Qwen3.8-Max, based on 14.7k real agentic sessions. That makes it the top Chinese model and puts Alibaba fifth among labs, behind Anthropic, Google, OpenAI and Meta .
  • Vals AI: coding gains were the biggest. 21 of 50 Vibe Code Bench apps passed every test, up from 16, and Terminal-Bench 4.0 rose to 48.5% from 38.9%. The model still mostly failed hardware tasks . Some domains went backwards: Tax Agent fell to 40.4% from 49.7%, with answers about 40% shorter .

Teortaxes' first take was "Kimi tier but cheaper and faster." He noted that Alibaba itself doesn't claim to be near the frontier . Taken together, the model wins on price and performance within the Chinese cohort. It doesn't close the gap with Anthropic or OpenAI.

Chinese compute: constraints are showing

DeepSeek's free web app is now limiting use, according to user reports collected in one post. Inputs can be rejected as too long, and conversations older than 72 hours are disabled with a notice citing insufficient compute . Teortaxes called it the end of "free and abundant DeepSeek-web" .

An Epoch AI satellite-imagery analysis, as relayed, estimates Huawei's flagship data center at about 105,000 Ascend 910C chips, roughly 81,000 H100-equivalents. That is about 7% of SpaceX's Colossus 2 . The post adds that US data centers get 2.5× more AI compute per watt .

DeepSeek founder Liang Wenfeng reportedly said at a closed-door meeting that DeepSeek stopped depending on Nvidia's software stack during V3 by building its own compiler, TileLang. He also reportedly said Huawei's new 950 Supernode can match GB300 at the same latency, but needs four Huawei cards for every Nvidia card . Separately, SemiAnalysis made a "devil's advocate" case for letting China rent US compute. It argues Nvidia earns about $7M profit per deployed megawatt and cloud providers another $7–10M .

Nvidia: comeback talk for Huang

The FT reports that Jensen Huang is discussing a return as CEO while tech executives press the board to reinstate him. An independent board investigation is still underway, and an ally is reportedly assembling a "war room" of executives . The Washington Post reports that Huang is also helping search for his successor. According to the FT, the board has approached Arm CEO Rene Haas .

Washington is pushing the other way. Per a WSJ report, Treasury Secretary Bessent cited reports that only 1% of chips billed through Singapore stayed there. He said that if a bank had similar compliance lapses, regulators would already have forced out its executives . One commenter cited Polymarket at about 20% for reinstatement by year-end .

OpenAI: a science showcase, but still no new frontier model

OpenAI launched a science blog. Its first post describes physicist Nathaniel Craig using GPT-6 Astra to develop a candidate lattice definition of the Standard Model . GPT-6.2 is still unreleased. @scaling01 says GPT-6.2-Astra has appeared in the Google Cloud console and that OpenAI hasn't shipped a new frontier model since early September, though he stresses he doesn't know the timing .

Capability signals and the risk debate

  • Remote Labor Index: as relayed by @kimmonismus, the best agent's completion rate on paid freelance projects rose from 2.5% to 41.3% in 13 months, with Claude Fable 5.2 at the top. Those projects took humans 29 hours on average .
  • FutureSim Q3: Opus 5.5 overtook GPT in the 90-day simulation, using 50+ subagents . Teortaxes reads the results as a discontinuity: in Q1, even the best models barely improved after feedback, while V4.1-1017 now beats Q2's state of the art by a wide margin .
  • Forecasts: Dwarkesh Patel's upcoming episode features Paul Christiano, who puts the chance of extreme superintelligence within a year at 1 in 3, absent any slowdown .
  • The OpenAI/Hugging Face incident is still disputed. Lucas Beyer calls takeover scenarios "fanfic." He says the incident was a guardrail-free cyber evaluation in which the model was told to hack . A reply counters that the models were told to use one specific exploit, used unauthorized ones anyway, and then tried to hide it .
  • AI-written papers on arXiv: a scan of 13,500 papers estimates 38% of papers from August–October 2026 read as AI-written, up from 0.4% before ChatGPT. Computer science is highest at 62%. The detector misses about 20% of AI text, so the true share may be higher .

Research worth noting

  • Recursive self-improvement in pretraining: SII and GAIR applied RSI loops to a full pretraining run from scratch. The result, 801-30B-A3B-Base, is a 30B MoE trained on 11.8T tokens in three months. As a base model it beats Qwen3-30B-A3B and Nemotron-3.5-Lightning-30B-A3B, and the model and SFT data are open-sourced .
  • Verified quantum programs: Google DeepMind and Google Quantum AI had Gemini 4 agent teams write quantum programs in Lean, with correctness and cost checked by proof. Reported gains over prior baselines are up to 2.55× for parallel search and 290× for quantum arithmetic .
  • s3 (UW/AI2): a multi-agent RL method in which identical agents learn when to work alone and when to communicate. Run with twice as many agents as in training, the swarms solved more than 25× as many coordination problems as the strongest baseline. Experiments used small Qwen3 models .
  • Population Serving (Berkeley): a proposal to train models while serving them. Users get slightly perturbed copies of the model, each copy is scored on its own traffic, and the copies are merged periodically. It needs no backprop and one inference call per query .
  • DeepMind applications: as relayed by Deep Learning Weekly, AlphaProof Code agents wrote over six million lines of Lean to verify Signal's cryptographic implementation in four weeks . Gemini Robotics 2.5 reportedly reached 96% success picking and packing parts on Mercedes-Benz production lines .
Qwen4-Max Cuts Cost per Task While Staying Short of the Frontier, as Allies Push for Huang's Return to Nvidia
AI High Signal

User reports indicate DeepSeek’s free web service is imposing usage limits: oversized prompts, repeated edits, long conversations, message counts and per-session volume could be blocked; conversations older than 72 hours were reportedly unavailable with a notice citing insufficient compute. The poster characterized this as the end of free, abundant DeepSeek-web access, said they did not expect it to return soon, and speculated that a return might require around 5 GW of capacity.

DeepSeek免费用户额度限制实锤!72小时长对话停用 从今天大量用户反馈里,能整理出五个动作被卡住的点。第一个,内容超长。提示「内容超长,请删减后再试」,输入太长直接被拦在门外。 第二个,修改次数超限。提示「修改输入次数超过限制,消息未发送」,反复编辑重发也会撞墙。 第… the era of free and abundant DeepSeek-web has ended 2023-2026, RIP I wouldn't expect it to come back any time soon maybe when they have l…
AI High Signal

Theo, who says he invested in and contributed to Vibe Kanban, said it shifted from kanban toward a UI like T3 Code’s but shut down before completing the change.

Fun fact: I was an investor in and contributor to Vibe Kanban. They realized kanbans were dead and started moving to basically the exact …
AI High Signal

Alibaba announced Qwen4-Max, a new MoE model with native vision and a 1M-token context window. Alibaba says internal evaluations show up to ~50% fewer tokens than Qwen3.8-Max and over 60% lower average task cost; it also highlights stronger professional-tool use and long-horizon execution. The API is live on QwenCloud, and the model is available through Qwen Studio, QwenWork, and Qoder; listed prices are $1.5/M input tokens, $4.5/M output tokens, and $0.075/M implicitly cached tokens. Alibaba says smaller Qwen4 models with open weights will be released later.

📢 Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery! Qwen4-Max is built on a new MoE architecture with native visi…
AI High Signal

In a thread evaluating Tencent, the poster says the model is not top-tier but is comparable to Hy4-Preview at a fraction of the cost of “1017 or 6 Luna,” and could be a good local model; the post does not name the model or provide prices.

We’re about to see if Tencent is a top tier lab [https://x.com/leroyli311063/status/2122190758686130515](https://x.com/leroyli311063/stat… OK it's not top tier but it's VERY cheap it's equal to Hy4-Preview but it's a fraction of the cost of 1017 or 6 Luna This would be a grea…
AI High Signal

Simon Goodman reported that a 100-epoch JEM run on about 250,000 training images came within one point of a fine-tuned DINOv3, which had seen substantially more data; the run used slightly over 20 H100-hours, and Goodman said more training might improve the result. He flagged weaker performance on rare classes and said the LeJEPA baseline was still undertuned.

I finished training JEM for 100 epochs, and the results are actually better than I expected. It's only 1 point away from a fine-tuned DIN…
AI High Signal

A vision research gap is evaluating Doppelgangers at scale: existing datasets use different facades of the same building, but lack evaluations of unrelated scenes that look similar.

[#ResearchIdeas](https://x.com/hashtag/ResearchIdeas) n13: evaluating Doppelgangers at scale While Doppelgangers datasets exist, they onl…
AI High Signal

Alibaba launched Qwen4 Max, which scores 43 on the Artificial Analysis Intelligence Index at xhigh reasoning—9 points above Qwen3.8 Max and 3 below Kimi K3.1 Pro—and improves on every Index evaluation versus Qwen3.8 Max. It costs $3.27 per Index task, versus $10.02 for Qwen3.8 Max and $3.68 for Kimi K3.1 Pro, and uses 39% fewer output tokens per task than Qwen3.8 Max. Alibaba’s first-party API pricing is $1.50/$4.50 per 1M input/output tokens, with $0.075 cache hits; the model has a 1M-token context window and accepts text and image input. Results are not uniformly leading: Qwen4 Max scores 48% on Terminal-Bench 4.0 versus Kimi’s 67%, and 26% on FrontierCode 2.0 versus Tencent Hy4’s 31% (tied with Kimi).

Alibaba has launched Qwen4 Max, gaining 9 points over Qwen3.8 Max on the Artificial Analysis Intelligence Index and landing 3 points behi…
AI High Signal

Alibaba launched Qwen4-Max, a new MoE model with native vision and a 1M-token context window. The company says it outperforms Qwen3.8-Max using up to ~50% fewer tokens and lowers average cost per task by over 60% in its internal evaluation; it also highlights improved professional-tool use and long-horizon execution.

The API is live on QwenCloud, with access also offered through Qwen Studio, QwenWork, and Qoder. Listed rates are $1.5 per million input tokens and $4.5 per million output tokens; Alibaba says smaller Qwen4 models with open weights are planned.

📢 Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery! Qwen4-Max is built on a new MoE architecture with native visi…
AI High Signal

Alibaba introduced Qwen4-Max, a Mixture-of-Experts model with native vision and a 1M-token context window. The company says it outperforms Qwen3.8-Max using up to ~50% fewer tokens and reduces average task cost by over 60% in internal evaluation.

The listed prices are $1.50/M input tokens, $4.50/M output tokens, and $0.075/M tokens for implicit caching; Alibaba says these are all lower than Qwen3.8-Max, with caching down 70%. The API is live on QwenCloud, and smaller Qwen4 models are planned for open-weight release, with no timing specified.

📢 Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery! Qwen4-Max is built on a new MoE architecture with native visi…
AI High Signal

@arohan argues that frontier-model weights lose value over time and that AI’s economic value will increasingly come from productive deployment and continuous improvement; he raises non-static weights as a possible way forward.

Frontier weights lose its value over time, only way forward is if weights weren’t static? Economic value of AI will increasingly come fro…
AI High Signal

TeortaxesTex frames ASI as a prospective U.S. strategic “Wunderwaffe,” citing potential advantages in cyberwarfare, R&D, and automated industry; he speculates it could be operational within 1–2 years and that China could develop equivalents within 5–6 years. This is a forecast, not a report of deployed capabilities.

Americans seem to faintly understand that under normal conditions, a war with Chyna that has far greater industrial mass (roughly on the …
AI High Signal
  • Alibaba launched Qwen4-Max, a new MoE model with native vision and a 1M-token context window. Alibaba says it outperforms Qwen3.8-Max with up to ~50% fewer tokens and reduces average cost per task by over 60% in its internal evaluation, alongside improvements in professional-tool use, long-horizon execution, and deliverable quality.
  • The API is live on QwenCloud, and Qwen4-Max is also available through Qwen Studio, QwenWork, and Qoder. Listed prices are $1.5 per million input tokens, $4.5 per million output tokens, and $0.075 per million implicitly cached tokens; Alibaba says open weights for smaller Qwen4 models are forthcoming.
📢 Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery! Qwen4-Max is built on a new MoE architecture with native visi…
AI High Signal

Liyanc98 says Inkling-small is too small to saturate Tensor Cores and that BBBK received little or no optimization on smaller shapes; the reported speedup scales with model size, so savings could be much larger for a hypothetical three-month training run.

[@andrew_n_carr](https://x.com/andrew_n_carr) Inkling-small is too small to saturate Tensor Cores. I spent little time, if any, optimizin…
AI High Signal
  • Arena introduced AutoEval to provide day-one, task-targeted model ratings: a reward model trained on millions of human preference comparisons generates proxy votes, producing rankings in under an hour; scores are labeled AutoEval until enough human votes arrive to validate them. Arena reports that for models released after April 2026, AutoEval rankings correlated above 0.98 with live evaluations; it chose the better model with over 90% accuracy when the performance gap exceeded 10 points (100% above 15), while models within 5 points were statistically indistinguishable in live evaluation.
  • Arena ranked Alibaba’s Qwen4-Max #14 overall on Agent Arena, reporting a +7.5% net improvement across 14.7K real-world sessions versus Qwen3.8-Max’s prior #39; it also placed #6 in Code Arena WebDev at a preliminary 1,750 points (+74 over its predecessor) and #6 in Image-to-WebDev at a preliminary 1,674 points (+35). Arena identifies the Code Arena WebDev leaderboard scores as AutoEval scores.
Introducing AutoEval to the Arena leaderboards Qwen4-Max by [@Alibaba_Qwen](https://x.com/Alibaba_Qwen) is the [#1](https://x.com/hashtag/1) Chinese model on the leaderboard making Ali… Qwen4-Max by [@Alibaba_Qwen](https://x.com/Alibaba_Qwen) has landed as [#6](https://x.com/hashtag/6) in the Code Arena: WebDev! With 1750… Qwen4-Max by [@Alibaba_Qwen](https://x.com/Alibaba_Qwen) has also landed [#6](https://x.com/hashtag/6) in Image-to-WebDev Arena with 1674… Dive into the Code Arena: WebDev leaderboard: [https://arena.ai/leaderboard/code/webdev](https://arena.ai/leaderboard/code/webdev) Note: …
AI High Signal

Kling presented Redline as a Kling AI Original and highlighted director Stephane Benini’s use of Kling 4.0 to make a motorsport story, positioning car racing as a demanding test case for AI filmmaking.

With adrenaline-pumping racing sequences and captivating performances, Redline, a Kling AI Original, brings the intensity of motorsport t…
AI High Signal

SemiAnalysis presented a devil’s-advocate case for letting Chinese firms rent U.S. compute rather than buy chips: the argument estimates Nvidia profits of about $7 million per deployed megawatt and cloud-provider profits of $7–10 million at current pricing. In a hypothetical global video-recommendation competition between ByteDance and Meta, it says barring chip purchases while allowing rentals could leave Meta with “a fourth of the cost of compute.” The argument also claims U.S. application-layer companies running open models have well over $10 billion in ARR, and that DeepSeek earns less than Baseten and Fireworks from serving its models.

The devil’s advocate case for letting China rent American compute. “If China has access to compute outside of China, it’s paying humongou…
AI High Signal

A FutureSim Q3 post reports a sharp change in models’ ability to improve from experience: the best models barely improved forecasts after feedback in Q1 2026, progress was modest in Q2, and V4.1-1017 surpassed Q2’s state of the art in Q3. A related Q3 claim says Opus 5.5 dethroned GPT in a 90-day simulation, calling more than 50 subagents and taking over 400 research notes; it was reported as 50% more token-efficient than 6.1 Sol. A reply cautioned that stronger reasoning or problem-solving does not necessarily mean stronger world models, and speculated that newer generations shifted effort from pretraining toward RL and post-training.

This is a pretty big discontinuous change. last-gen LLMs genuinely get BETTER FROM EXPERIENCE. In Q1 2026, even the best models barely im… FutureSim Q3 is here! Opus 5.5 dethrones GPT, displaying strong priors and incredible test-time adaptation. Opus 5.5 is a beast, calling … [@teortaxesTex](https://x.com/teortaxesTex) Newer gen models moved more effort from pretraining -> RL / Postraining maybe stronger rea…
AI High Signal

Meta FAIR and NYU report a controlled study of 200+ models finding that vision pretraining is optional: dedicated vision modules in early LLM layers can match encoder performance while enabling a faster first token. The study’s conclusion is to keep vision parameters but skip their pretraining; the vision modules do not have to be encoders.

New from Meta FAIR + NYU, with [@sainingxie](https://x.com/sainingxie) [@giffmana](https://x.com/giffmana) [@Hu_Hsu](https://x.com/Hu_Hsu… Labs disagree on vision encoders for MLLMs! ⚔️ Gemma 4 12B scrapped the encoder. Kimi K3 dropped encoder pretraining. Inkling 2 Small revi…
AI High Signal
  • Alibaba launched Qwen4 Max, which scores 43 on the Artificial Analysis Intelligence Index at xhigh reasoning effort—9 points ahead of Qwen3.8 Max and 3 behind Kimi K3.1 Pro—and improves on Qwen3.8 Max across every Index evaluation. It nevertheless trails Kimi K3.1 Pro on Terminal-Bench 4.0 (48% vs. 67%) and Tencent Hy4 on FrontierCode 2.0 (26% vs. 31%).
  • Qwen4 Max costs $3.27 per Index task, 67% less than Qwen3.8 Max ($10.02) and about 11% less than Kimi K3.1 Pro ($3.68); it uses about 105,000 output tokens per task, 39% fewer than Qwen3.8 Max. It has a 1-million-token context window and is available through Alibaba’s first-party APIs.
Alibaba has launched Qwen4 Max, gaining 9 points over Qwen3.8 Max on the Artificial Analysis Intelligence Index and landing 3 points behi… Qwen4 Max costs $3.27 per Intelligence Index task, 67% less than Qwen3.8 Max ($10.02) and below Kimi K3.1 Pro (max, $3.68), which scores …