Tencent's 770B Hy4 preview, Alibaba's Qwen4 teaser and GLM-5.3-Flash at $0.075 per million input tokens all landed as open weights, while OpenAI says unshipped Astra already clears its research-intern bar.
Tencent put a preview of Hy4 on Hugging Face on August 28, a mixture-of-experts model with 770 billion total parameters and roughly 49 billion active per text request, aimed at software work. Alibaba's Qwen3.8-Flash-Next had landed two days earlier as an explicit Qwen4 architecture preview.
GLM-5.3-Flash, previously previewed as Ox Alpha, lists at $0.075 per million input and $0.25 per million output under a 50% discount through September 9. GPT-5.6 Sol sits at $4 input and $20 output.
The glacier collapse and flood wave across the Tibet-Nepal border left more than 2,400 missing as of Aug 28, with 586 bodies recovered. Rescue access remains the constraint.
A weekly operator deep dive, the full Dojo Pro prompt packs, and the complete prompt database. Founding members lock the launch price forever.
Join the founding list →Tencent posted a preview of Hy4 to Hugging Face on August 28, an open-source mixture-of-experts model with 770 billion total parameters and roughly 49 billion active per text request, aimed at software engineering, research and financial analysis. Tencent says it will integrate the model into CodeBuddy and...
Continue reading arrow_outwardTencent posted an open-source preview of Hy4 to Hugging Face on August 28, a 770-billion-parameter mixture-of-experts model with roughly 49 billion active parameters per text request, targeted at software engineering, research and financial analysis.
The missing count from the Himalayan glacier collapse and flood wave on the Tibet-Nepal border passed 2,400 on August 28, with 586 bodies recovered and 933 hydropower workers added to Nepal's list.
Mood is Greed even as traders raised US rate-hike odds on hawkish comments from Fed Chair Kevin Warsh, with core PCE at 3.3% and Michigan consumer sentiment down to 51.7 in August from 55.2 in July.
Washington's two levers on Friday were both financial: a US-Venezuela oil agreement and a FinCEN rule to cut Banque Misr UAE off from US correspondent banking; neither instrument touches a 770-billion-parameter file uploaded to Hugging Face.
Tencent's Hy4 preview and Alibaba's Qwen3.8-Flash-Next, explicitly framed as a Qwen4 architecture preview rather than a flagship, both went out as open weights before they were finished. Publishing architectural experiments buys deployment feedback that closed labs have to simulate internally, the same sequencing Z.ai used with the model it previewed as Ox Alpha and then released as GLM-5.3-Flash.
GLM-5.3-Flash lists at $0.075 per million input tokens and $0.25 per million output under a 50% promotional discount running through September 9, while the larger GLM-5.3, whose open weights are still pending, carries a cited Artificial Analysis Agentic Index score of 59. Discount windows with hard expiry dates are a customer-acquisition tactic, not a cost structure, so the question is what September 10 pricing looks like.
OpenAI says the unshipped Astra family already clears an internal bar for an automated AI research intern: implement an idea in code, run the experiment, return results. Forbes reported the claim on August 28 alongside leadership statements that AGI arrives by year end, which puts a self-defined internal metric against Google's shipped and specific releases, Gemini 3.5 Transcribe with sub-second streaming across more than 85 languages and Gemini Omni 1.1 Flash.
Fourteen Days to Cutoff
Read full analysis arrow_outward
Deep Dive
7.1/ 10
Every social listening tool reads captions; Oriane actually watches the video, and that changes what you can ask.
Oriane is for marketers, agency strategists, and creator-marketing leads who keep reverse-engineering hooks by hand and want a queryable index of what is inside social video instead of just around it. The concept is genuinely differentiated and the free tier is enough to validate whether your niche is covered, but the paid jump is steep and early reviewers say the interface still thinks in videos when teams think in accounts. Worth a one-hour test if you brief video content weekly; skip it if you post monthly.
“Finally something taking on the old and clunky brand tracking and social listening tools.”LinkedIn
Sumly.AI condenses audio and video into short written summaries, and the listing says it is currently free and ad free. Use it on the long-form interviews and recorded webinars in your queue, then keep only the ones where the summary raises a question you need the full source to answer. It is a triage tool, not a substitute for listening to the two things that matter.
Try it arrow_outwardY Combinator's big-data directory now lists 28 funded companies, including Lyon, a two-person San Francisco team building private foundation models that run inside a bank's own cloud; it trained on 28 billion transactions for one fintech and identified premium-card converters four times more precisely than the existing rules. If you are buying prediction rather than chat, these are the teams competing with your internal feature store. Read the descriptions for the deployment model, since on-premises versus API is the real decision.
Try it arrow_outwardBrex's summer 2026 ranking of the 25 fastest-growing software startups is built from corporate card spend, which is a harder adoption signal than download counts or Product Hunt upvotes. Cross-check it against the tools your own finance team is already expensing before you approve another pilot. Growth in spend tells you a category is real; it does not tell you the winner.
Try it arrow_outwardThe list was updated on 28 August 2026 and narrows 11 evaluated AI tools into ranked picks with a stated weakness for each, covering meeting notes, dictation, email, and video. The useful part is the side-by-side of where each tool falls short, which is what most roundups omit. Note the page is FTC-disclosure compliant affiliate content, so use it to build a bake-off list rather than to pick a winner.
Try it arrow_outwardBenchmark the cheap tier honestly. Route a real workload through GLM-5.3-Flash at $0.075 input and $0.25 output, then compare quality and latency against your current GPT-5.6 Sol path at $4 input and $20 output. Log the delta before the 50% promo ends on September 9.
Treat previews as previews. Pull Hy4 and Qwen3.8-Flash-Next weights into a sandbox for coding and multimodal evals, but keep them out of production paths. Both were shipped as architecture signals, not finished flagships.
Instrument your media and research intake. Point Oriane at your social video to test hook performance rather than guessing, run Sumly.AI over the podcasts and calls stacking up unheard, and scan the YC Summer 2026 big-data batch before your vendor shortlist hardens.
Three Chinese labs used this week to ship incomplete work into the open rather than wait for a polished flagship, and it worked as positioning. Tencent and Zhipu are pushing open weights precisely because Western toolchains keep gating them, while OpenAI cut Cursor's model access and countered with an internal claim about Astra clearing a research-intern bar nobody outside the lab can test. For builders, the practical read is that capability is arriving cheaper and faster than your procurement cycle can absorb it. Price the alternatives now, keep previews in the sandbox, and treat unverified internal benchmarks as marketing until weights or an API show up.
AI analysis, world news, markets, and tools. One briefing, delivered free.