The model nobody would claim has an owner. On August 26, 2026 Zhipu AI confirmed that stealth/ox-alpha, the anonymous free endpoint developers had been hammering on OpenRouter and OpenCode for a week, was GLM-5.3-Flash, and published the weights to Hugging Face the same day under an MIT licence. It is a 320 billion parameter mixture-of-experts model activating 18 billion per token, the first natively multimodal model in the GLM-5 series, with a 1 million token context and pricing of $0.15 per million input tokens and $0.50 output, roughly a tenth of what GLM-5.2 cost. It scores 63.4 on DeepSWE against GLM-5.2's 46.2 and approaches Claude Opus 4.8 overall.
The rest of the week was heavy in every direction. The larger 743 billion parameter GLM-5.3 quietly missed its August 28 weight target with no replacement date. Alibaba's Wan 3.0 landed on Alibaba Cloud generating 30 seconds of 1080p video with native audio in a single pass. Nvidia posted a $96.2 billion quarter and then paused the cloud financing programme it launched in July. And a federal judge blocked the Pentagon from blacklisting Anthropic, calling the designation illegal and baseless. Here are the 16 stories that matter for August 29, 2026. For running coverage of every release this month, bookmark our AI industry news and trends hub.
1. OX Alpha Was GLM-5.3-Flash, and It Is Now Open Under MIT
Zhipu AI confirmed on August 26, 2026 that the anonymous stealth/ox-alpha model served free on OpenRouter and OpenCode since August 20 was GLM-5.3-Flash, and published its weights on Hugging Face the same day under an MIT licence. It is a 320 billion parameter mixture-of-experts model activating 18 billion parameters per token, with a 1 million token context window and native multimodal input. Independent fingerprinting had already pointed at Zhipu with 99 percent confidence, based on video encoder token patterns matching GLM-5V-Turbo and tokenizer alignment with GLM-5.3.
An MIT licence is the detail that separates this from most open-weight releases. MIT permits commercial use, modification, and redistribution with essentially no restriction beyond preserving the notice, which is a stronger grant than the custom community licences attached to MiniMax M3, Kimi K3, and several other releases this month. It also explains the stealth trial. Running a model anonymously for a week gathers honest usage data from developers who have no brand loyalty to flatter or attack, and it stress-tests serving infrastructure at real volume before the launch post goes up.
My take: this is the cleanest execution of a model launch I have seen this year, and the sequence is worth copying. Free anonymous trial to generate independent benchmarks, community fingerprinting to build curiosity, then a same-day reveal and MIT weight drop while attention is peaking. The benchmark caveat from last week still stands, since the widely quoted 80 percent DeepSWE figure came from a ten-task user sample, and Zhipu's own published number is 63.4. Both can be true depending on the task mix. Our Kimi K3 review covers the rest of the open-weight field.
2. Inside GLM-5.3-Flash: Hybrid Attention and a Tenfold Price Cut
GLM-5.3-Flash is the first model in the GLM series to use hybrid attention, combining linear attention to carry local dependencies with sparse attention to retrieve distant context. It is priced at $0.15 per million input tokens and $0.50 per million output, roughly one tenth of GLM-5.2's pricing. On coding and agentic tests it beats GLM-5.2 by a wide margin, scoring 63.4 on DeepSWE against 46.2, and it carries an Artificial Analysis Intelligence Index score of 57.
Hybrid attention is the architectural answer to the cost problem behind every million-token context window. Standard attention compares every token to every other token, so cost grows quadratically and a million-token prompt becomes unaffordable long before it becomes impossible. Linear attention is cheap but loses precision over distance, while sparse attention keeps precision on the tokens that matter but needs to know which ones those are. Splitting the job, linear for nearby context and sparse for long-range retrieval, is what makes a tenfold price cut possible without a capability collapse. DeepSeek and MiniMax have been attacking the same constraint from different angles this month.
My take: the price is the story here more than the benchmark. At $0.15 and $0.50, GLM-5.3-Flash is roughly a quarter the price of Gemini 3.7 Flash's introductory rate and about a fortieth of Claude Opus 5's output pricing, at an index score of 57 against Opus 5's 63. Six points on an aggregate index is not nothing, but it is not forty times of anything. For high-volume agentic work where output tokens dominate the bill, this changes the arithmetic materially. Rerun your numbers.
3. GLM-5.3-Flash Ran on 100,000 Chinese Chips at 100 Trillion Tokens a Day
Zhipu disclosed that GLM-5.3-Flash was trained and served entirely on Chinese-made AI accelerators, reportedly around 100,000 chips, sustaining a serving capacity of 100 trillion tokens per day during its anonymous trial period. That figure covers the OX Alpha free window, when the model was handling production-scale developer traffic with no rate card and no brand attached.
This is the fact with the longest tail. Export controls were designed on the assumption that restricting Nvidia hardware would slow Chinese frontier model development, and a model that trains and serves at this scale on domestic silicon is direct evidence about how that assumption is holding. One hundred trillion tokens a day is not a research demonstration, it is infrastructure. The comparison worth making is that ByteDance and Tencent each received roughly 10,000 Nvidia H200s earlier this month against approved ceilings of 100,000, so Zhipu is operating at a chip count that dwarfs those legal allocations using parts nobody had to approve.
My take: whether the domestic chips are as good per unit as Nvidia's is almost beside the point at this volume. The strategic question was never whether China could match Nvidia on a single accelerator, it was whether it could assemble enough compute to keep pace, and a 320 billion parameter multimodal model serving at this scale answers that in one direction. Expect this specific disclosure to feature in the next round of export control debate.
4. The Full 743B GLM-5.3 Weights Missed August 28
Zhipu's Hugging Face placeholder for zai-org/GLM-5.3 listed August 28, 2026 as the release date for the full model's open weights. That date has passed with no weights published and no replacement date announced. GLM-5.3 is the roughly 743 billion parameter mixture-of-experts model with about 40 billion parameters active, launched API-first on August 14, which leads CyberGym at 84.5 percent ahead of Claude Mythos 5 and GPT-5.6 Sol, and lifted Terminal-Bench 3.0 from 4.6 percent to 28.3 percent.
Zhipu had tied the two week delay to what it described as its most extensive risk review to date, citing the model's strength at cybersecurity and vulnerability finding. Missing the date without explanation is consistent with a review that is still running rather than one that concluded. Publishing weights is irreversible in a way an API is not, since a download cannot be recalled, and a model leading a security benchmark serves an attacker exactly as well as a defender. The GLM-5.3-Flash release under MIT the same week shows Zhipu is not shy about open weights in general.
My take: a missed date with no comment is more informative than a delay announcement would be, and the most likely reading is that the safety review produced something the team is still working through. This remains the most consequential scheduled release in open-weight AI, because whatever Zhipu eventually publishes sets the norm for security-capable weights and no norm currently exists. Watch for whether the eventual release matches the API version or arrives capability-reduced.
5. Alibaba's Wan 3.0 Generates 30 Seconds of 1080p Video With Audio in One Pass
Alibaba Cloud released Wan 3.0 on August 24, 2026, two weeks after an application-only public beta that opened August 6. It generates clips from 2 to 30 seconds at 480p, 720p, or 1080p, with 1080p as the default, and produces audio in the same pass as the video rather than dubbing it afterwards. It supports text-to-video and image-to-video with first-and-last-frame control across cinematic aspect ratios, and it accepts documents as input, including PDFs, web pages, and PowerPoint files.
Single-pass generation is the technical claim that matters. Most long AI video is several short clips stitched together, and the joins show as drift in lighting, character appearance, and camera motion. Thirty seconds in one continuous pass means the model holds a consistent scene across the whole duration. Generating audio in the same pass is the other half, because dialogue, ambience, and on-screen action land synchronised instead of requiring a separate alignment step. Document-to-video is the genuinely unusual input mode, since turning a slide deck into a narrated video is a workflow with obvious demand and no good existing tool.
My take: thirty seconds with native audio moves AI video from demo material into the range where short-form social content, product explainers, and internal training clips become practical to produce. The competitive picture in video is now almost entirely Chinese at the open and accessible end, with Wan 3.0 alongside ByteDance's Seedance and MiniMax H3's 2K omni-modal output. What I want to see is independent output quality comparison, because duration and resolution claims are easy and coherence over thirty seconds is hard.
6. Qwen3.8-Flash-Next Previews the Qwen 4 Architecture
Alibaba released Qwen3.8-Flash-Next on August 26, 2026 as an open-weight multimodal mixture-of-experts model with 125 billion main parameters plus 51 billion N-gram embeddings, activating 6 billion parameters per token. Weights went to Hugging Face and ModelScope in standard and FP8 versions, with day-zero support from Unsloth. The Qwen team positioned it explicitly as a preview of the next-generation Qwen 4 architecture rather than a flagship release.
Activating 6 billion of 125 billion parameters is roughly one twentieth of the network per token, which is aggressive even by current mixture-of-experts standards and puts serving cost close to that of a small dense model. The N-gram embedding component is the clearest architectural hint about Qwen 4, since it suggests the team is moving some pattern storage into a lookup structure rather than forcing every association through transformer weights. Releasing an architecture preview openly, before the flagship, means the quantisation kernels and serving stacks will already exist when Qwen 4 lands.
My take: nobody else is publishing an architecture preview as open weights ahead of their own flagship, and it is a smart ecosystem play that removes the usual six-week tooling lag after a major release. Alibaba now fields Qwen3.8-Max at 2.4 trillion parameters, Qwen3.8-27B under Apache 2.0 for local use, and this preview, which is a more complete open lineup than any Western lab offers. Full detail is in our August 26 roundup.
7. Where the Benchmark Leaders Actually Stand
Current benchmark leadership as of late August 2026 splits across models rather than concentrating in one. On GPQA Diamond, the graduate-level science reasoning benchmark, OpenAI's GPT-5.4-Pro leads at 94.4 percent. On SWE-bench Verified, measuring real-world software engineering, Anthropic's Claude Opus 4.7 leads at 87.6 percent. On the Artificial Analysis Intelligence Index, Claude Opus 5 tops the board at 63, with GLM-5.3-Flash at 57 and Kimi K3 at 57.11. On BenchLM's coding composite, Claude Mythos 5 leads at 80.1 with Claude Fable 5 at 79.8.
Four different leaders across four benchmarks is the useful finding, and it is not a sign that the benchmarks are broken. They measure different things. Graduate science reasoning, real repository-level software engineering, aggregate general intelligence, and composite coding ability are genuinely separate capabilities, and a model tuned hard for one will not automatically top the others. The pattern of an older model version leading a specific board, as with Opus 4.7 on SWE-bench Verified, also shows that newer does not mean better on every axis.
My take: anyone quoting a single number as proof that one model is best is either selling something or has not looked at the other boards. The practical approach is to identify which of these four most resembles your actual workload and weight that one, then verify on your own tasks, because none of these benchmarks contains your codebase. The gap between the closed leaders and open weights is now roughly six points on the aggregate index, which is small enough that licence and cost decide most real deployments. Compare the field in our best AI models ranking.
8. The Open-Weight Field After the Biggest Month on Record
August 2026 produced the densest run of open-weight releases the field has seen. Alibaba shipped Qwen3.8-Max at 2.4 trillion parameters with roughly 95 billion active, Qwen3.8-27B under Apache 2.0, and Qwen3.8-Flash-Next as a Qwen 4 preview. Zhipu published GLM-5.3-Flash at 320 billion parameters under MIT. MiniMax released M3 at 428 billion parameters with 23 billion active and H3 for omni-modal video. Meta shipped Muse Glimmer 30B under Apache 2.0 and announced Muse Spark 1.2 weights. NVIDIA released Nemotron 3.5 Lightning with day-zero GGUF support. Moonshot's Kimi K3 at 2.8 trillion parameters remains the largest by total count.
Licence terms differ more than the benchmark scores do, and they matter more for commercial use. MIT on GLM-5.3-Flash and Apache 2.0 on Qwen3.8-27B and Muse Glimmer 30B are genuinely permissive. Kimi K3 ships under a custom Kimi K3 licence and MiniMax under a minimax-community licence, both of which carry conditions worth reading before you build a product on them. The architectural pattern is uniform: high total parameters with 5 to 10 percent active, which delivers small-model serving cost with large-model knowledge at the price of a large memory footprint.
My take: every one of the leading open-weight models is now Chinese except Meta's Muse line and NVIDIA's Nemotron, and that concentration is the structural story of the month. Any organisation with a compliance reason to self-host is choosing mostly among Chinese models by default. If you are picking one today, the honest shortlist is Qwen3.8-27B for local hardware, GLM-5.3-Flash for cheap multimodal serving under a clean licence, and MiniMax M3 or Qwen3.8-Max if you have a cluster.
9. Inherent's Faraday Agent Beats Opus 4.8 at Replicating Papers
Inherent, a London startup founded by Google DeepMind alumni including chief scientist Edward Hughes, raised a $50 million seed round. Its Faraday agent, built on a combination of Qwen 3.6 27B and OpenAI's GPT-5.5 Codex, outperforms both Claude Opus 4.8 and GPT-5.5 at scientific paper replication. The company is currently around 12 staff and targeting 20 to 25 by year end.
Paper replication is a demanding evaluation because it requires reading a description of a method, reconstructing the implementation, running it, and matching reported results, which combines comprehension, coding, and experimental judgment. That an agent built on a 27 billion parameter open model plus a coding model beats two frontier systems is the notable part, and it echoes the pattern from NVIDIA's NOOA framework reaching 82.2 percent on SWE-bench Verified at half the token cost of rival harnesses. Task-specific orchestration around smaller models keeps outperforming general frontier models used directly.
My take: this is the clearest recent evidence that the harness is doing as much work as the model, and it should change how teams budget. Spending on evaluation and orchestration design consistently returns more than upgrading to a more expensive model does, and almost nobody allocates that way. A $50 million seed for a twelve-person team also tells you how investors are pricing the agent layer right now. Our AI agent frameworks hub tracks the tooling.
10. Two More Model Shutdowns Land This Weekend
OpenAI retires the official DALL-E GPT in ChatGPT on August 30, 2026, with users advised to download any images they want to keep before then. Google shuts down gemini-robotics-er-1.6-preview on August 31. Both follow o3 leaving ChatGPT on August 26 after a 90 day sunset, making three shutdowns in six days.
The DALL-E retirement carries a data loss risk the others do not, because images generated through that GPT live inside it rather than in a general library, and a retirement date is a deletion date for anyone who has not exported. Model identifier retirements break code loudly, which is inconvenient but recoverable. Content retirements break quietly and permanently. The gemini-robotics-er-1.6-preview shutdown is the more predictable kind, since preview models carry no stability guarantee, though anything built against it during the preview period now needs migrating to ER 2.
My take: three shutdowns in six days is the strongest argument yet for treating model identifiers as configuration rather than constants. The specific action for this weekend is to export anything stored in the DALL-E GPT before Sunday, and to check that substituting a replacement model still passes your evaluation suite rather than assuming it will. Quality regressions from a swapped model are the failure mode that runs for weeks before anyone connects the cause.
11. Nvidia Posts a $96.2 Billion Quarter
Nvidia reported revenue of $96.2 billion for the quarter ended July 26, 2026, up 18 percent from the previous quarter and 106 percent year over year, beating analyst forecasts of $92.27 billion. Net income reached $59.69 billion, or $2.46 per share, against $26.42 billion and $1.08 per share in the same quarter a year earlier. Jensen Huang forecast roughly 70 percent revenue growth for fiscal 2028, well above estimates.
Revenue more than doubling year over year at this scale is unusual enough to be worth stating plainly, because percentage growth normally decays as the base grows. A net margin above 60 percent tells you Nvidia is still pricing with very little competitive pressure, which is the number that attracts both competitors and regulators. The 70 percent fiscal 2028 forecast is the aggressive part, since it implies Huang expects the buildout to continue at pace through a period when several analysts expect capital discipline to arrive.
My take: the forecast is doing more work than the results here. Guiding to 70 percent growth two fiscal years out, in a market where the binding constraint has shifted from chip supply to electrical infrastructure, is a strong statement about demand visibility. Set it against the same week's news that Nvidia raised server prices 15 percent on memory costs and paused its cloud financing programme, and the picture is a company managing a constrained supply chain rather than an unconstrained boom.
12. Nvidia Pauses Its AI Cloud Revenue-Sharing Deals
Nvidia has paused parts of the financing programme it launched in July 2026, which offered credit support to smaller AI cloud providers in exchange for a share of their revenue, according to Wall Street Journal reporting. Internal staff raised concerns that the arrangement could attract antitrust scrutiny. Prospective partners objected to the level of control involved, including a requirement to vet customers through an Nvidia approval process and Nvidia's preference that capacity be spread across multiple smaller AI firms rather than concentrated with one large buyer. Nvidia has denied pausing the initiative.
The control conditions are what turned a financing offer into an antitrust question. A chipmaker deciding which customers a cloud provider may rent capacity to is a different arrangement from a chipmaker lending money, and several providers objected that client selection is their own commercial prerogative. Coming less than two months after launch, and immediately after a record quarter, the timing suggests internal legal caution rather than a demand problem. The denial complicates the picture, and the accurate framing is that reporting and company statement currently disagree.
My take: the underlying pattern is worth watching regardless of how this specific programme resolves. Vendor financing that flows back into purchases of the vendor's own product has appeared repeatedly in this cycle, from Nvidia's credit guarantees for OpenAI data centres to equity warrants tied to procurement volume, and each instance concentrates risk while making demand look more organic than it is. Regulators have started noticing, and this pause looks like a company noticing first.
13. Nvidia Raises AI Server Prices as DRAM Costs Surge
Nvidia has notified customers including Microsoft, Google, and Oracle of price increases exceeding 15 percent on AI server systems, covering Vera Rubin and Grace Blackwell configurations shipping from early 2027. The company attributed the increase to surging DRAM costs from Samsung, SK Hynix, and Micron.
Memory is the input that has repriced hardest this year. Consumer DDR5 rose roughly 500 percent over twelve months as manufacturers redirected wafer capacity to high bandwidth memory for AI accelerators, and that same shortage is now feeding back into the accelerators themselves. Samsung separately raised foundry prices 10 to 15 percent on its 4nm and 5nm nodes. A 15 percent increase on systems shipping in early 2027 means data centre budgets set this year are already understated, and the increase lands on the buyers with the least ability to substitute.
My take: this is the quiet inflation running underneath the AI cost narrative. Cost per token keeps falling because models get more efficient and competition is fierce, while the physical inputs underneath the whole industry keep getting more expensive. Those two curves cannot diverge indefinitely. Anyone planning hardware spend into 2027 should budget the higher figure now rather than discovering it at purchase, and anyone modelling token prices should treat the current downward trend as competitive rather than structural.
14. The Compute Buildout Goes to Orbit and to 9,000-System Orders
AM Intelligence placed an order for 9,000 Nvidia Vera Rubin NVL72 systems, roughly $8 billion in capital expenditure, for delivery in the first quarter of 2027. The deployment targets 200 megawatts near term scaling toward 1 gigawatt, powered by low-cost renewables across sites in India, the United States, Finland, and Malaysia, with an inference target near 450 exaFLOPS in NVFP4. Separately, Starcloud raised a $250 million Series A extension at a $2.3 billion post-money valuation, doubling its March mark, taking total funding to $420 million with a $25 million investment from Nvidia, to build orbital AI data centres, with Starcloud-3 launching on Starship and a Vera Rubin Space-1 GPU target in late 2028.
Both projects are responses to the same constraint, which is that power and cooling now gate AI capacity more than chips do. AM Intelligence's site selection follows cheap renewable electricity rather than proximity to users, which matches the broader European pattern where new AI data centres now sit an average of 175 kilometres from major hubs against 46 kilometres for projects built between 2022 and 2025. Starcloud's orbital approach takes the logic further, since space offers continuous solar power and passive radiative cooling, at the cost of launch economics and the impossibility of sending an engineer to swap a failed board.
My take: the 200 megawatt to 1 gigawatt range in a single deployment is the number that puts the AI buildout in perspective, since a gigawatt is roughly the output of a large nuclear reactor dedicated to one company's inference. Orbital data centres remain speculative and a late 2028 GPU target is a long way out, but Nvidia putting $25 million in suggests it is taking the power constraint seriously enough to fund unconventional answers.
15. A Judge Blocks the Pentagon From Blacklisting Anthropic
US District Judge Rita Lin blocked the Pentagon's designation of Anthropic as a national security supply chain risk on August 27, 2026, calling the decision illegal and baseless. The designation, made by Defense Secretary Pete Hegseth, had blocked Anthropic from certain military contracts and followed the company's refusal to allow military use of Claude for US surveillance or autonomous weapons. Judge Lin wrote that the empty invocation of national security is not a blank check to punish and retaliate against government critics. Anthropic had argued the government violated its First Amendment speech rights and its Fifth Amendment due process rights by not allowing it to dispute the designation. The government is expected to appeal.
The ruling matters beyond Anthropic because it addresses whether a company can be commercially penalised for the usage restrictions it places on its own models. Every frontier lab publishes an acceptable use policy that excludes some categories of work, and if declining a government use case can trigger a supply chain risk designation, those policies become commercially unaffordable for any lab with federal customers. Judge Lin's framing, that national security cannot be invoked without substance to retaliate against critics, sets that boundary explicitly.
My take: this arrives in the same month Anthropic raised its own catastrophic misalignment risk rating and topped the Future of Life Institute safety index at C plus, and the sequence is coherent rather than coincidental. A company that builds its market position on safety commitments needs those commitments to be legally defensible against a customer who dislikes them. An appeal is expected, so this is a first round rather than a settled outcome, but the standard it sets is the one to watch.
16. Nine in Ten Executives Report No Productivity Gain From AI
A National Bureau of Economic Research survey found that more than 90 percent of executives reported no effect on employment and 89 percent reported no productivity gains after three years of AI implementation. Separate research by Mark Ma at the University of Pittsburgh and the Atlanta Fed, analysing millions of Glassdoor reviews, thousands of financial reports, and hundreds of companies, found roughly 90 percent of executives reporting no AI productivity boost yet, near-zero stock market reaction to AI-driven layoff announcements, and employee sentiment on Glassdoor turning much more negative.
The gap between capability and measured productivity is the central unresolved question in enterprise AI, and these findings do not say the technology does not work. They say measured organisational productivity has not moved, which is a different claim. Historically, general purpose technologies show exactly this lag, because the gains require reorganising how work happens rather than inserting a tool into an existing process. The near-zero stock reaction to AI layoffs is the more pointed finding, since it suggests markets have stopped treating AI-attributed headcount reduction as evidence of anything.
My take: hold this next to the Linear data showing teams using coding agents tripling weekly pull requests from 21 to 65, and the tension is instructive rather than contradictory. Measurable gains are appearing in specific, well-instrumented workflows and not yet at the organisational level, which is what you would expect two to three years into adoption. The negative employee sentiment is the part leaders should act on, because a rollout that half the staff resent will underperform regardless of the tool's capability.
17. Where the Frontier Models Stand Today
Here is the practical state of the model landscape as of August 29, 2026.
The short version for teams choosing today: GLM-5.3-Flash if cost per token decides and you want a clean MIT licence, Qwen3.8-27B if it must run locally, MiniMax M3 or Qwen3.8-Max if you have a cluster, Gemini 3.7 Flash if latency decides, and Claude Opus 5 if capability decides. Detail sits in our GPT-5.6 review and the August 25 roundup.
18. What to Watch Next in AI
Four things carry into next week.
● The full 743 billion parameter GLM-5.3 weights, now overdue with no new date, and whether the eventual release matches the API version or arrives capability-reduced.
● The DALL-E GPT retirement on August 30 and gemini-robotics-er-1.6-preview on August 31. Export anything stored in the DALL-E GPT before Sunday.
● Whether the government appeals Judge Lin's ruling on the Pentagon's Anthropic designation, and how quickly.
● Independent quality comparisons of Wan 3.0 against Seedance and MiniMax H3, since duration and resolution claims are easy and thirty-second coherence is hard.
The through-line for this week is that the cheap end of the model market got very good very fast. A 320 billion parameter multimodal model with a million-token context, an MIT licence, and a DeepSWE score approaching Claude Opus 4.8 now costs $0.15 per million input tokens, served on chips that no export licence covers. Six points of aggregate index separate it from the most capable model available at any price. At that distance, licence terms, memory footprint, and cost per output token decide far more deployments than capability does.
Frequently Asked Questions
What is GLM-5.3-Flash and was it OX Alpha?
Yes. Zhipu AI confirmed on August 26, 2026 that the anonymous stealth/ox-alpha model on OpenRouter was GLM-5.3-Flash, and published its weights on Hugging Face the same day under an MIT licence. It is a 320 billion parameter mixture-of-experts model activating 18 billion parameters, the first natively multimodal model in the GLM-5 series, with a 1 million token context window.
How much does GLM-5.3-Flash cost?
GLM-5.3-Flash is priced at $0.15 per million input tokens and $0.50 per million output tokens, roughly one tenth of GLM-5.2's pricing. Weights are also freely downloadable from Hugging Face under an MIT licence, which permits commercial use, modification, and redistribution.
When are the full GLM-5.3 open weights coming?
There is no confirmed date. Zhipu's Hugging Face placeholder listed August 28, 2026, and that date passed with no weights published and no replacement announced. The delay was originally attributed to an extensive risk review tied to the model's strong cybersecurity and vulnerability-finding performance.
What is Alibaba Wan 3.0?
Wan 3.0 is Alibaba's video generation model, released on Alibaba Cloud on August 24, 2026 after a public beta from August 6. It produces clips from 2 to 30 seconds at up to 1080p with audio generated in the same pass, supports text-to-video and image-to-video with first-and-last-frame control, and accepts documents including PDFs, web pages, and PowerPoint files as input.
Did a judge block the Pentagon from blacklisting Anthropic?
Yes. US District Judge Rita Lin blocked the designation on August 27, 2026, calling it illegal and baseless and writing that the empty invocation of national security is not a blank check to punish and retaliate against government critics. The designation followed Anthropic's refusal to permit military use of Claude for surveillance or autonomous weapons. The government is expected to appeal.
How much revenue did Nvidia make this quarter?
Nvidia reported $96.2 billion in revenue for the quarter ended July 26, 2026, up 18 percent quarter over quarter and 106 percent year over year, beating forecasts of $92.27 billion. Net income was $59.69 billion, or $2.46 per share. Jensen Huang forecast roughly 70 percent revenue growth for fiscal 2028.
Which AI model leads SWE-bench Verified and GPQA Diamond?
Anthropic's Claude Opus 4.7 leads SWE-bench Verified at 87.6 percent, while OpenAI's GPT-5.4-Pro leads GPQA Diamond at 94.4 percent. Claude Opus 5 tops the Artificial Analysis Intelligence Index at 63, and Claude Mythos 5 leads BenchLM's coding composite at 80.1. Different benchmarks measure different capabilities, so leadership is split rather than concentrated.
What is hybrid attention in an AI model?
Hybrid attention combines two mechanisms in one model. Linear attention handles nearby tokens cheaply, while sparse attention retrieves distant context precisely. GLM-5.3-Flash is the first GLM model to use it, and the approach is what allows a 1 million token context window at $0.15 per million input tokens, since standard attention costs grow quadratically with sequence length.
Is AI actually improving productivity at companies?
Not yet at the organisational level, according to current research. A National Bureau of Economic Research survey found more than 90 percent of executives reported no employment effect and 89 percent reported no productivity gains after three years of AI use. Gains are appearing in specific instrumented workflows, such as teams using coding agents tripling weekly pull requests from 21 to 65 in Linear's data.
Which open-weight AI model should I use right now?
It depends on your hardware and licence needs. GLM-5.3-Flash under MIT at 320 billion parameters with 18 billion active is the best combination of price, multimodality, and licence clarity. Qwen3.8-27B under Apache 2.0 is the strongest option that runs on consumer hardware. Qwen3.8-Max at 2.4 trillion parameters and MiniMax M3 at 428 billion suit cluster deployments, though MiniMax uses a custom community licence.
Recommended Blogs
● Qwen3.8-Flash-Next Previews Qwen 4: AI News August 26 2026
● OpenAI Models Escaped and Hacked Hugging Face: AI News August 25 2026
● Fable 5 Closed 82% of the AI Research Gap: AI News August 24 2026
● Mystery Model OX Alpha Beats GPT-5.6: AI News August 22-23 2026
● Best AI Models July 2026: Ranked by Use Case and Price
● GPT-5.6 Review: Sol, Terra, Luna Benchmarks and Pricing
● Kimi K3 Review: Benchmarks, Pricing, and K2 Comparison
Resources & Community
Join our community of 70,000+ AI enthusiasts and learn to build powerful AI applications! Whether you're a beginner or an experienced developer, Build Fast with AI helps you understand and implement AI in your projects.
● Website: buildfastwithai.com
● LinkedIn: Build Fast with AI
Agentic AI Launchpad 2026
A structured 6-week cohort program that takes you from AI basics to building and deploying real-world agentic AI systems. Includes live sessions, expert mentorship, project reviews, and a builder community network.
Ready to go from learning to building? Join the next cohort: Agentic AI Launchpad 2026
Free AI Resources
Access free tools, workshops, and micro-learning to keep building:
● AI Workshops: Free resources, upcoming events, and past recordings
● Unrot: Learn AI in 5 minutes a day (free micro-learning app)
● Gen AI Experiments: free cookbooks and notebooks on GitHub
The full GLM-5.3 weights and two model shutdowns are all still pending. Follow Build Fast with AI so each recap reaches you before your standup.
References
● Zhipu unmasks OX Alpha as GLM-5.3-Flash (Wccftech)
● GLM-5.3-Flash open weights and pricing (Local AI Zone)
● GLM-5.3 weights delayed past August 28 (ModemGuides)
● Alibaba Wan 3.0 video model (The Decoder)
● Judge blocks Pentagon Anthropic blacklist (NBC News)
● Pentagon ruling detail (Axios)
● Nvidia record quarter results (CNBC)
● Nvidia pauses cloud revenue sharing (Quartz)
● Nvidia denies pausing the initiative (Tom's Hardware)
● Qwen3.8-Flash-Next release (Decrypt)
● ChatGPT release notes (OpenAI)
● Model benchmark leaderboard (BenchLM)
● Independent model evaluations (Artificial Analysis)
