Notizie

Alibaba’s Qwen team has released Qwen3.8-Flash-Next, an open-weight multimodal Mixture-of-Experts model built for cost per token. The checkpoint pairs a 125B backbone with a 51B N-gram embedding table and a 4B multi-token prediction module. Only 6B parameters activate per token. The team positions it as an early preview of the architecture that will underpin Qwen4, the same role Qwen3-Next played for Qwen3.5. Four changes carry the release: a Gated DeltaNet and Qwen Sparse Attention hybrid, Gated Residual, N-gram Embedding, and the Muon optimizer. Qwen team reports training cost at roughly one-ninth that of Qwen3.7-Plus. Is it deployable? Yes but not on a workstation. The FP8 checkpoint is 172.78 GiB and the BF16 checkpoint is 335.28 GiB. Per vLLM recipes, TP2 is the minimum validated FP8 configuration on GB300 and TP4 is recommended. On an 8×H200 node, use TEP8; plain TP8 is incompatible with the checkpoint’s 128-wide quantization blocks. Sparse activation cuts compute, not storage. https://qwen.ai/blog?id=qwen3.8-flash-next What is actually new Qwen3.8-Flash-Next pairs a 125B main model with 51B N-gram embedding parameters and a 4B multi-token prediction module, totaling 180B on disk. Only 6B parameters activate per token. Four changes drive this: Hybrid attention (GDN + QSA): Three of every four layers use Gated DeltaNet, a linear-attention layer that compresses history into a fixed-size recurrent state. The fourth layer runs Qwen Sparse Attention (QSA), which uses a lightweight indexer to select context at micro-block granularity rather than per token. The layer layout is 12 × (3 × GDN → 1 × QSA) across 48 layers, with a QSA budget of 512 blocks or 2048 tokens. Gated Residual: The residual stream widens into 4 parallel branches, with an element-wise read gate and a per-branch scalar write gate, at bottleneck rank 320. N-gram Embedding: A 20,000,000-entry bigram/trigram table at layer 2 adds capacity through deterministic lookups. It can be offloaded to host memory with asynchronous prefetch — though offload currently runs only on NVIDIA devices. Training recipe.:The Muon optimizer is applied alongside AdamW to specific weight categories, with batch-size warmup eliminated and scaling laws refitted. The MoE layer carries 512 experts, activating 10 routed plus 1 shared, at expert intermediate dimension 640. Benchmarks Qwen reports 58.7 on DeepSWE 1.1, 62.5 on SWE-bench Pro, 81.0 on SWE-bench Multilingual, and 91.9 on LiveCodeBench v6. On agentic tasks it posts 73.9 on CoWorkBench, 55.7 on JobBench, and 73.5 on Toolathlon Verified. Multimodal results include 84.5 on AndroidWorld, 76.6 on LVBench, 88.5 on RealWorldQA, and 95.7 on MathVision with code interpreter. The model does not lead everywhere. Claude Opus 4.6 (Max) takes HLE at 40.0 against Qwen’s 35.9, and DeepSeek-V4-Flash-0731 leads NL2Repo-Bench at 54.2 versus 48.1. Frontier reasoning remains the gap. Efficiency Qwen states training cost roughly 1/9 that of Qwen3.7-Plus. On serving, the announcement cites QSA kernel speedups of up to 7.6× prefill and 4.9× decode at 1M tokens, while the SGLang cookbook and vLLM recipes cite 10.2× and 6.6×. Treat the range as vendor-reported until independently measured. Qwen also reports 8.6× the prefill throughput of Qwen3.7-Plus at a 90% prefix-cache hit rate. Context is 262,144 tokens natively, extensible to 1,000,000 with YaRN. Running it The model serves through vLLM, SGLang, TokenSpeed, transformers serve, and llama.cpp for GGUF quants. Fine-tuning is supported via Unsloth, Swift, and LLaMA-Factory. It already powers the “Standard” mode on QwenWork and works with Qwen Code. Thinking mode is on by default, with reasoning_effort at xhigh, medium, or low. Qwen recommends temperature 1.0 and top_p 0.95 for thinking mode, and temperature 0.7 with top_p 0.80 for instruct mode. Key Takeaways 125B backbone + 51B N-gram embeddings + 4B MTP, with only 6B parameters active per token. Three of four layers use Gated DeltaNet; the fourth runs Qwen Sparse Attention at micro-block granularity. Trained at roughly 1/9 the cost of Qwen3.7-Plus, with 262K native context extensible to 1M via YaRN. FP8 weights are 172.78 GiB, so self-hosting needs a multi-GPU node, not a workstation. Licensed under qwen-community-1.0, not Apache-2.0 — verify terms before commercial use. Check out the GitHub Page, HF Model Card and Technical Details. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well. The post Alibaba’s Qwen Team Releases Qwen3.8-Flash-Next: A 125B Multimodal MoE With 6B Active Parameters Previewing the Qwen4 Architecture appeared first on MarkTechPost.

Even as a teenager, Laurie Stach ’06 says, she had a “crazy ambition to take on the world and solve problems.” At MIT, she realized she wasn’t the only one. “Adults always see these youth who are good at math and science and say, ‘You’re going to do great things—someday,’” says Stach. “Meanwhile, traditional education isn’t preparing them.” That’s why in 2012 Stach founded LaunchX, a for-profit entrepreneurship program that helps high school students build practical skills through an intensive summer program in which they start real companies. Stach says that students form teams, identify a problem, and build an actual product or service that they then take to market—often generating preorders or even revenue before the four-week program ends. Raised in Texas and the Carolinas, Stach majored in mechanical engineering at MIT and earned her MBA from Harvard Business School in 2011. She ran her first summer program on MIT’s campus with 30 students in 2013 and says “it kind of snowballed.” LaunchX now serves about 500 students per year across both in-person and online programs. Read more at www.technologyreview.com/alumni-profiles.

For Rupert Young ’95, SM ’95, his career in data science and cybersecurity began when his grandfather gifted him thousands of stamps: He built intricate databases to catalogue them, displaying the “precise eye” for detail and nuance that his MIT application essay said would make him a good engineer. Young is now chief product officer of ­MaxMind, whose work with IP-address location data and fraud prevention has made it a go-to resource. Streaming companies, security vendors, merchants, and ad-network providers use its signature tool, GeoIP, to determine how and where users are accessing services. This information makes it possible to ensure retail sites are using the correct currency, flag suspicious bank logins, and more. Young’s passion for engineering was shaped by internships, and he has paid it forward by volunteering at his children’s California high school and trying to keep up with what “the next generation of engineers are tinkering with.” The next generation of product innovation at MaxMind excites him too. “Working with my team to try to find patterns in data and solve challenging problems—to me, there’s no greater joy,” he says. Read more at www.technologyreview.com/alumni-profiles.

When Don Yansen ’63 arrived at the MIT AgeLab for a study on technology in care­giving for older adults, he didn’t plan to launch another company. But when he heard participants talk about how hard modern devices can be to use, he decided to develop an alternative. Yansen, a serial entrepreneur who retired to care for his wife, earned an MIT degree in electrical engineering and a master’s in physics from Northeastern. “What I took from MIT was a good education and the feeling I could solve difficult problems,” he says. One such problem: Many older adults “can’t even use a modern cell phone,” he says, “and very few people can use something like an Apple Watch.” So Kiwi Health, which he founded in October 2024, sells a wristband controlled primarily by voice. Wearers can set reminders, make calls, send texts, and share health data, all without navigating screens. AI helps interpret these commands, accounting for the way voices may change with age. The device also measures health metrics and automatically alerts caregivers to falls.  Yansen’s hope, he says, is that it can “help seniors live the best life that they can.” Read more at www.technologyreview.com/alumni-profiles.

Oral and forehead thermometers may not accurately capture a person’s core body temperature, and the few ingestible temperature sensors on the market are so big they are hard to swallow and risk obstructing the GI tract. But MIT engineers created one that can send continuous temperature updates at a size of just six by four millimeters. To create it, the researchers built a circuit that can fit on a one-square-­millimeter silicon chip and designed an oscillator based on leakage current—the small current that flows through a circuit when it’s off. Its frequency varies depending on the temperature of the chip’s surroundings. The circuit detects temperature within 0.01 °C and can be powered with a 1.55-volt coin cell battery. A strategy known as backscattering further cuts energy consumption, allowing most of the power requirements to be outsourced to an antenna outside the body. It emits an ultra-high-frequency radio wave, which is modulated by an antenna within the sensor and sent back to the external one. By interpreting the changes in the radio wave, the external antenna can calculate the temperature value. The result is “the smallest ingestible capsule that we have seen for temperature-­sensing paradigms,” says Saransh Sharma, a former MIT postdoc now at the University of Cambridge, who is lead author of a paper on the work.  The researchers envision using it to monitor infection, observe patients during and after anesthesia, track fevers in children, help mark ovulation, and more. It could also be used to monitor athletes, soldiers, or anyone else exposed to extreme temperatures.  “A sensor like this gives us the ability to monitor infections and identify them early,” says Giovanni Traverso, an associate professor of mechanical engineering and one of the paper’s senior authors, along with MIT provost Anantha Chandrakasan. “That’s very relevant, particularly for at-risk populations like people who are immunosuppressed.” Ultimately, he hopes, it could replace other types of thermometers for everyone.

Model cards report quality under server-class, full-precision conditions. Those numbers rarely predict how the same model behaves on a phone. This week, Liquid AI released Pipette. It is an open-source platform for benchmarking foundation models on edge devices, built in partnership with Artificial Analysis as an independent methodology validator. Pipette treats on-device behavior as a property of the deployed system, not the model in isolation. Its unit of measurement is a full configuration: model + quantization + runtime + device. The launch dataset covers five on-device performance metrics across more than 1,000 model × quantization × runtime × device × context configurations, spanning 30+ models, llama.cpp builds for macOS, iOS, Windows and Android, and context lengths from 256 to 8,192 tokens. Initial verified results come from a MacBook Pro with M5 Max, an iPhone 17 Pro and a Galaxy S26 Ultra. The practical claim is testable: two 350M models at the same quantization on the same phone retain 78.4% and 33.8% of decode throughput at 4,096 tokens. Is it deployable? Yes, Pipette ships as Apache 2.0 infrastructure (pipette-mgmt, pipette-clients, pipette-scores), a public results dataset, a hosted dashboard, and native iOS and Android benchmark apps. Nothing is waitlisted. Publication of community-submitted results is still in beta. Which companies: Any team shipping a model onto hardware it does not own. Solo developers and seed-stage startups can use the dashboard and apps without infrastructure. Mid-market product teams can run the clients across an internal device fleet. Large OEMs, chip vendors and enterprises can operate the whole pipeline behind their own firewall. Industries: Consumer electronics and smartphone OEMs, automotive, industrial and robotics, healthcare devices, financial services, defense — anywhere latency, privacy or connectivity forces inference onto the device. Applications: Model and quantization selection before a sprint commits; SoC and hardware procurement validation; regression testing when a runtime, OS or driver updates; context-length capacity planning; independent verification of vendor performance claims. What Liquid AI shipped Liquid AI released Pipette in partnership with Artificial Analysis, an independent validator that reviewed and verified the methodology. The premise is narrow and useful: on-device behavior is a property of the deployed system, not of the model in isolation. The launch dataset covers five on-device performance metrics across more than 1,000 model × quantization × runtime × device × context configurations. It spans 30+ models, multiple quantization formats, llama.cpp builds for macOS, iOS, Windows and Android, and context lengths from 256 to 8,192 tokens. Initial published results come from a MacBook Pro with M5 Max, an iPhone 17 Pro and a Galaxy S26 Ultra, with AMD Ryzen AI Max+ 395 and Radeon 8060S results listed as coming soon. In Pipette, the unit of measurement is a deployment configuration: model + quantization + runtime + device. A benchmark then defines the metric and token shape, producing a latency, throughput or memory result. Quality is tracked separately on IFBench, GPQA Diamond and MATH-500. Those quality scores currently come from llama.cpp evaluation runs on NVIDIA H100 80GB reference systems, then get matched to on-device runs sharing the same model and quantization — a quality number shown next to phone throughput was not produced on the phone. Why the deployment context changes the answer Four published comparisons show how far a configuration can move a decision: Context scaling can diverge at identical parameter counts. At Q4_K_M on Galaxy S26 Ultra, Granite-4.0-H-350M retains 78.4% of its decode throughput from 256 to 4,096 input tokens, while Granite-4.0-350M retains only 33.8%. Sparse activation buys speed, not memory. At 2,048 input tokens on the same phone, LFM2.5-8B-A1B decodes 2.4x faster than Qwen3.5-4B and 2.6x faster than Ministral-3-3B-Instruct-2512. It activates 1.5B of 8.5B parameters per token, yet still peaks at 5.29 GiB because all expert weights occupy memory. Speed and quality do not co-locate. On iPhone 17 Pro at Q4_K_M, MiniCPM5-1B completes a 2,048-in / 256-out workload in 3.47 seconds versus 4.12 seconds for LFM2.5-1.2B-Instruct, a 15.8% reduction in elapsed time. On the same artifacts, LFM scores 9.0 points higher on MATH-500. Near-identical system profiles can hide task-level reversals. At Q4_K_M and 2,048 input tokens on M5 Max, Granite-4.1-8B and Ministral-3-8B-Instruct-2512 differ by 2.4% in decode throughput and 1.2% in peak RAM. Granite leads IFBench by 7.3 points; Ministral leads GPQA Diamond by 14.0 points. How the measurements are produced Performance runs follow a published methodology: fixed token shapes, greedy decoding, a discarded warm-up, five measured repetitions and readiness gating. Before each timed repetition, a platform-specific check verifies thermal and load conditions; failing runs are not published. Evaluations use a separate protocol with deterministic, model-blind scoring, and pipette-scores never sees generation provenance. Every submission records benchmark version, token shape, model artifact, quantization, runtime version and settings, and device hardware and OS. Interactive explainer Key Takeaways Pipette benchmarks configurations, not models: model + quantization + runtime + device. Apache 2.0 stack, 1,000+ configurations, 30+ models, three verified devices at launch. Quality evals run on H100 references and are matched to on-device performance, not measured on-device. Identical parameter counts can differ 78.4% vs 33.8% in context-scaling retention. Check out the Technical Details and Leaderboard. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well. The post Liquid AI Open-Sources Pipette: A Reproducible Benchmarking Suite That Measures On-Device Models, Quantization, Runtime and Hardware Together appeared first on MarkTechPost.

This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. Kids outlearn AI—and we still don’t know why Teaching a computer to use human language requires an inhuman amount of data. An LLM can easily churn through a hundred thousand times more words than a person will experience in the process of mastering their mother tongue. This yawning divide between children and machines is called the data efficiency gap. It raises a tantalizing question for cognitive scientists and a challenge for the architects of AI models: how can children still outperform the most linguistically sophisticated machines ever built? Kids show that it could be possible to learn more with less. Far less. By reverse-engineering the way they learn, scientists hope to create more data-efficient AI models—and perhaps settle enduring questions about language and children’s developing minds. Here’s how researchers are trying to close the data gap. —Elise Cutts Job titles of the future: space travel agent As co-owner of a luxury lifestyle firm, Roman Chiporukha has long turned wild vacation dreams into reality. In 2018, he got a phone call that would open up a new frontier: Axiom Space wanted to find citizen explorers willing to pay $50 million each to join the first fully private mission to the International Space Station (ISS), slated for April 2022. This showed Chiporukha that the sky was no longer the limit; it was the market. He successfully signed up a private astronaut and then launched SpaceVIP in 2021 to offer celestial experiences that mix culture, science, and purpose. Find out what it takes to become the Expedia of the cosmos. —Linda Childers These stories are from the next issue of our magazine, which is all about kids. Subscribe now to get your copy as soon as it lands.  The must-reads I’ve combed the internet to find you today’s most fun/important/scary/fascinating stories about technology. 1 Both parties are turning against AI data centers ahead of the midtermsThe backlash is spilling onto the campaign trail. (NYT $)+ Republicans have sharply turned against them. (WP $)+ Texas’ governor says data centers “dug their own grave.” (Axios)+ Here’s why data center opposition crosses party lines. (MIT Technology Review) 2 Chinese humanoids have smashed Usain Bolt’s 100-metre world recordOne ran it in 9.39 seconds at the World Humanoid Robot Games. (Verge)+ Another, also built by X-Humanoid, beat a high jump record. (ESPN)+ But intricate real-world tasks pose greater challenges. (Reuters $)+ Gig workers are training humanoids at home. (MIT Technology Review) 3 Uber has been fined nearly $1 billion over automated suspensionsDutch regulators say drivers were deactivated without human review. (Quartz)+ It’s the second-largest fine issued yet under the EU’s GDPR. (Reuters $) 4 New Zealand plans to ban under-16s from social mediaIt’s joined a throng of countries trying to restrict access. (Bloomberg $)+ But opposition parties could torpedo the bill. (Reuters $) 5 TikTok will pay $400 million to settle a US child privacy caseIt was sued for allegedly violating ‌children’s online privacy. (BBC)+ The US DOJ said TikTok illegally collected kids’ information. (Reuters $) 6 China’s largest-ever car recall covers nearly 3 million TeslasThey’re being recalled over door safety concerns. (Quartz)+ The action covers roughly 4.3 million vehicles in total. (NYT $) 7 Dr. Dre and Jimmy Iovine think AI is good for musicThey compared fear of AI to resistance to earlier technologies. (NYT $)+ AI let a musician with MLS sing again. (MIT Technology Review) 8 Taiwan has indicted nine people over alleged illegal AI exports to ChinaThey include employees of ​Nvidia and Super Micro (Reuters $)  9 NASA’s new telescope could discover 200,000 new planetsRoman could also reveal new details about dark matter. (Wired $)+ And detect killer asteroids. (MIT Technology Review) 10 Plastic bottles can be turned into edible, vanilla-flavour cookiesThanks to genetically-engineered yeast. (New Scientist $) Quote of the day “Credit where credit is due: the tech sector has managed to unite a deeply divided country at a time of maximal partisanship.” —Max Steele, Senior Director of Communications at Everytown, reacts on X to a new poll showing that support for data centers in the US has plummeted. One More Thing How next-generation nuclear reactors break out of the 20th-century blueprint As worries about climate change and energy independence drown out concerns about meltdowns and radioactive waste, demand for nuclear power has surged. The problem is, building nuclear power plants is expensive and slow. Now, a new generation of nuclear power technology could reinvent what a reactor looks like—and how it works. Advocates hope it can refresh the industry and help replace fossil fuels without emitting greenhouse gases. Small modular reactors could bring the assembly line to nuclear development, while new designs are experimenting with different fuels and coolants, from TRISO to molten salt. Here’s what the next generation of nuclear reactors could look like.  — Casey Crownhart We can still have nice things A place for comfort, fun, and distraction to brighten up your day. (Got any ideas? Drop me a line.) + Welcome to WikiCity, where every building is a Wikipedia article.+ Animal lovers have built a tiny village for dogs that don’t have homes.+ Here are 18 ridiculous yet cathartic new words to describe modern tech irritants.+ Feast your eyes on these Porsche 912s customized for the Japanese police in the 1960s.