Uncategorized

The outspoken longevity scientist David Sinclair has been predicting that one day, you’ll go to the doctor and get a prescription that will make you 10 years younger. Now MIT Technology Review has learned that he has plans to launch human tests of an oral “reprogramming” drug as part of a $101 million competition organized by the XPrize Foundation.  The foundation is offering cash awards to teams able to “restore” a person to an earlier apparent age, as measured by improvements in immune, cognitive, and muscle function.  The grand prize goes to any team able to show a 10-year (or greater) relative improvement after one year of treatment.  Reached by phone, Sinclair, a biologist at Harvard Medical School, confirmed that he plans to give an oral drug mixture to volunteers in a bid to seek “evidence for age restoration in humans.” The trial, if it goes forward, will be a significant new development in the race to harness so-called epigenetic reprogramming. That technology is based on the discovery, 20 years ago, of powerful genes able to turn an adult cell into a stem cell similar to those found in embryos. The age-reversal effect is believed to occur via a resetting of molecular controls on DNA known as epigenetic marks, which help determine a cell’s overall metabolism and identity. Companies are now racing to use that phenomenon for a new form of rejuvenation medicine. Only this January, one of Sinclair’s companies, Life Biosciences, made news by winning approval to launch an initial human trial using a set of powerful reprogramming genes. The company announced today it had treated its first patient.  But that test involves a complex gene therapy and is limited to patients’ eyes, where it could treat conditions like glaucoma.  Sinclair’s new plan is bolder: a reprogramming drug you’d swallow in order to promote such effects across the body.  “What we’re aiming to do is to epigenetically restore the animal and eventually the person,” he says. “It is true that we’ve been doing extensive animal studies with the oral agent and are looking to compete in the XPrize.” This alternative method, chemical reprogramming, uses drugs to mimic the effects of the embryonic genes. That is significant because drug compounds can travel through the bloodstream, reaching most or all cells in a person’s body.  Some experts expressed caution, saying the chemical process, at least as used in labs, is extremely harsh and not even particularly effective. “Who doesn’t dream of whole-body rejuvenation? I think it’s a great goal,” says Sergiy Velychko, founder of Soxogen, a stealth reprogramming company in Boston. “But these chemicals are used in very, very high concentrations for cell reprogramming.” Sinclair declined to describe the exact makeup of the drug candidate, code-named SL-100, calling its contents “highly, highly confidential.” However, he has previously published lab studies of what he called “epigenetic age-reversal cocktails,” which mixed powerful chemicals with known supplements and commercially available medicines.  It’s those latter components that would be easiest to test on people, since doctors are free to prescribe them, even for unusual objectives like age reversal. James Clement, head of Betterhumans, an organization that specializes in life-extension studies using existing drugs, said in a message that he is “running clinical trials” of an oral reprogramming cocktail for Sinclair’s XPrize team. Sinclair’s team is competing in the XPrize Healthspan Competition, launched in 2023. It follows several previous competitions that focused on commercial spaceflight, lunar landings, and other goals. The XPrize Foundation is led by executive chairman Peter Diamandis, also an active promoter of longevity research. “If two teams are equivalent, they would split the award,” says Jamie Justice, a doctor and executive director for the contest, which was bankrolled by Saudi Arabia’s Hevolution Foundation, “But it will be incredibly hard to even get to one winner.” Justice says a judging panel is now in the process of picking 10 finalists from 65 teams that have been exploring health foods, lifestyle interventions, digital trackers, and drug compounds.  Sinclair’s team, Justice says, was a late entrant to the contest, but like all teams, it would be required to move into wider human tests starting this year. “You have to be ready and in trials,” she says. The race to harness the reprogramming phenomenon and apply it to living people is heating up, even outside the XPrize competition. On June 2, a startup called NewLimit, founded by the crypto billionaire Brian Armstrong, said it had raised a further $435 million, from investors including Peter Thiel’s Founders Fund, to support what it calls “age reprogramming.”  The company says it is working toward delivering genetic reprogramming instructions to the liver, to treat diseases of that organ. But Sinclair has been saying that whole-body rejuvenation is a possibility too. And for that, chemicals, rather than gene therapy, could be the most practical strategy.  Sinclair says his lab has been searching for such compounds and is starting to use AI “to improve the oral agents that we’re testing.” Chemical reprogramming cocktails, as used in labs, typically involve a mix of vitamins, approved drugs, and experimental molecules. For instance, one recipe Sinclair filed a patent on includes the supplement forskolin,  the antidepressant tranylcypromine, and an experimental chemical, laduviglusib, which has been tested against Alzheimer’s, among other ingredients. “In those days it was a six-factor cocktail,” Sinclair says of his earlier research. “But we’ve come a long way. I can’t disclose what’s in it, but it’s an improvement and an advance on that, and we’ve done a number of animal studies. They are not published, but we’ve been doing them for a long time, and we want to make sure that we’ve done a full investigation of safety and efficacy before we release any of the data.” While Sinclair’s results aren’t published, other teams say attempts to reverse the age of entire animals using chemical drugs haven’t worked yet. Last year, the lab of Vadim Gladyshev, another Harvard biologist and a member of a different XPrize team, reported on its attempt to

At SXSW London last week I gave a talk called “Five things you need to know about AI,” in which I shared what I think are the biggest themes in AI right now. I pulled a few things from our first AI10 list, an annual guide to the most important trends in this buzzy world, but I also veered off on a number of tangents. In my half-hour slot, I tried to cover the key talking points that I think help to make sense of what’s going on in tech—and thus the economy—today.   (I gave a talk with the same title at SXSW London last year with five different things you needed to know. A lot has happened since then!) So: This is how I’m thinking about AI midway through 2026. Let me know if you would pick different points! 1. Strictly speaking, I didn’t need to show up to give this talk. Tongue in cheek? Maybe. But generative AI tools have already become mundane, used by millions to automate everyday office tasks (including producing and delivering talks). It’s no surprise that one of the biggest questions out there right now is what this all means for jobs. People are confused and scared. The frustrating answer is that despite the hype coming from the top about the potential for AI to join the workforce soon—and viral social media posts yelling that something big is happening—there is almost no data to say either way what kind of effect this technology will have on employment and the economy overall. That’s not to say it won’t have an impact, even a huge one, but it’s just too soon to tell. In theory, teams of agents working together toward common goals could become assembly lines for white-collar work, doing to offices this century what Henry Ford’s innovations did to factories in the 20th century. In theory. Because in order to know what will happen to jobs, we need to know what will happen inside the companies that create those jobs. But most companies are still figuring that out.  2. AI is getting scary (for real this time). There have been scary stories about AI for years—claims that it will kill us all or bring about the end of civilization. There’s still a loud crowd of doomers, but those scenarios remain dystopian science fiction. What’s happened instead is that many of the worst near-term, real-world fears have come true. Take deepfakes, AI-generated images or videos of people doing things they didn’t actually do. Deepfakes have been used to incite violence, swing votes, and sow distrust. Trump’s White House is among those creating and publishing fake images. Many deepfakes are also used to abuse women and girls. One study found that 98% of deepfakes are pornographic and 99% involve women. Another concern is the rise of dangerous and delusional relationships with chatbots. Many people turn to chatbots to seek private advice and to feel heard. But there are now multiple lawsuits against AI companies alleging that the technology encouraged or aided suicides and other forms of self-harm. AI is also being used in warfare in new and worrying ways. LLMs are now giving advice, not just being used for analysis. One US defense official told my colleague James O’Donnell that you could now give a military chatbot a list of targets and ask which one to hit first. Anyone who uses AI knows that its output needs to be reviewed carefully. In fact-paced, high-stress active conflict, the risk that corners get cut is high. 3. A lot of people really hate AI. I checked out an anti-AI protest in London earlier this year and found a very broad mix of complaints. Banners proclaiming the end times bounced along to chants of “Stop the slop! Stop the slop!” Protests are getting more organized and drawing larger crowds. There’s pushback from fans of films and video games, who object to the use of generative AI in their favorite titles. In one notable case, the acclaimed 2025 game Clair Obscur was stripped of an award when the developers admitted to using AI in just one small, specific part of its production. And there’s the data center backlash. The US has more than 5,400 data centers and counting. With the energy demands of AI growing, people are unhappy about the environmental impact and their rising electricity bills. Activists are managing to stall development in a number of places. Regulation is becoming politically popular. Grassroots movements like QuitGPT have gained momentum. A small number have turned to violence; a few weeks ago somebody threw a Molotov cocktail at Sam Altman’s house. It’s not clear where all this leads. But the apocalyptic hype from tech leaders is not helping people stay calm. 4. AI for science is a very big deal. It’s early days yet, but the potential for AI to help make a genuine and important scientific discovery is greater than ever. Google DeepMind has developed Co-Scientist, a multipurpose tool that can help researchers dig up and compare previous results, generate hypotheses, and devise experiments to test them. OpenAI told me this year that its North Star is the goal of building a fully automated researcher by 2028. Mathematicians are excited too. Fundamental math underpins many everyday technologies, from internet security to video streaming. The last few months have seen a string of claims that AI has cracked unsolved math problems. And software that can solve really hard math problems will be able—so the argument goes—to solve more general-purpose real-world problems too. What are the downsides? Some scientists are warning that an overreliance on AI tools could narrow the scope of research because scientists may choose problems that are most suited to AI assistance. There are also concerns that AI-assisted research will lead to a flood of inaccurate or fake results: science slop. 5. AI is everywhere all at once. So where does that leave us? There are a lot of exciting things, a lot of worrying things, and a

Google just announced Gemini 3.5 Live Translate. It is their latest audio model for live speech-to-speech translation. Speech-to-speech means spoken audio goes in, and translated spoken audio comes out. The model detects over 70 languages automatically and generates translated speech. It preserves the speaker’s intonation, pacing, and pitch in the output. Turn-by-turn systems wait for a speaker to finish before responding. Gemini 3.5 Live Translate generates speech continuously instead. It balances a trade-off between waiting for context and translating immediately. More context improves quality. Faster output keeps the translation in sync with the speaker. The result stays a few seconds behind the speaker throughout a session. Gemini 3.5 Live Translate Gemini 3.5 Live Translate is a single audio model (gemini-3.5-live-translate-preview), not a chat assistant. It processes speech as the audio streams in, rather than after a full sentence. It handles multilingual inputs without manually configuring settings. Its noise robustness lets applications run in loud, unpredictable environments. The model is rolling out across three surfaces. Developers get it in public preview through the Gemini Live API and Google AI Studio. Enterprises get a private preview in Google Meet starting this month. Everyone else gets it through the Google Translate app on Android and iOS. How the Continuous Streaming Works The design difference matters for building real-time features. A conversational Live agent uses turn-based interactions. It relies on pauses, intent detection, and interruption handling. Live Translation uses continuous stream processing instead. It translates as the speaker talks, without waiting for turns to end. To hold strict real-time latency thresholds, the translation path accepts audio input only. Text input is not supported in translation mode. The model also drops tool use and system instructions in this mode. That keeps it a focused translator pipeline rather than a general agent. Building With the Live API Developers configure translation inside the Live API session setup. You set a translationConfig block within the generationConfig. The targetLanguageCode field takes a BCP-47 code, such as “pl” or “es”. BCP-47 is the standard format for language tags like en or pt-BR. It defaults to “en”. The echoTargetLanguage boolean controls input that is already in the target language. When true, the model echoes that speech. When false, it stays silent. You can also enable inputAudioTranscription and outputAudioTranscription for text transcripts. Audio formats are fixed. Input is raw 16-bit PCM at 16kHz, mono, little-endian. Output is raw 16-bit PCM at 24kHz, mono, little-endian. PCM is uncompressed raw audio. You send audio in chunks of 100ms. For client-side apps, ephemeral tokens on the v1alpha endpoint avoid exposing your API key. Dimension Live Agent Live Translation Model role Assistant that listens, reasons, and acts Interpreter / real-time translator pipeline Interaction Turn-based, with interruption handling Continuous stream processing, no turns Tools Function calling, Google Search, instructions Translation only, no tools or instructions Inputs Text, audio, video, and image Audio only, for strict latency Configuration Generation, speech, tools, instructions targetLanguageCode and echoTargetLanguage Use Case The model targets live interpretation across several settings. Google lists multilingual calls, meetings, lessons, and broadcasts. Developer platforms reduce the integration work for real-time media. Agora, Fishjam, LiveKit, Pipecat, and Vision Agents already use the Live API. These platforms handle the complex real-time media streaming infrastructure. That lets developers focus on the user experience instead. Google’s example app demonstrates dubbing and simultaneous multi-language translation. Grab is testing the model for driver-and-traveler communication at pickups. Grab users make over 10 million voice calls per month. CJ ENM, LiveKit, and others reported positive feedback on quality, accuracy, and low latency. How It Changes Google Meet and Translate According to Google’s official release, Google Meet will soon use 3.5 Live Translate for speech translation. The table shows the stated before-and-after for Meet. Capability Previous Meet With 3.5 Live Translate Languages 5 70+ Combinations per meeting Only to and from English 2000+ combinations Access Existing interface Updated interface for instant access The Meet update is in private preview for select business Workspace customers this month. A broader rollout follows later this year. In the Translate app, the Live translate feature works with any connected headphones. It mirrors the speaker’s tone across 70+ languages. Android also gains a listening mode. You hold the phone to your ear like a regular call. The translated audio then streams through the earpiece, without others hearing. Key Takeaways Gemini 3.5 Live Translate is Google’s latest audio model for live speech-to-speech translation across 70+ languages. It streams continuously instead of turn-by-turn, staying a few seconds behind the speaker. Developers can configure it via the Live API using targetLanguageCode and echoTargetLanguage; audio-only, 16kHz in, 24kHz out. It rolls out to the Gemini Live API, Google Meet (5→70+ languages), and the Translate app. All generated audio carries an imperceptible SynthID watermark for detectability. Check out the 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. Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us The post Google Releases Gemini 3.5 Live Translate, a Streaming Speech-to-Speech Audio Model Covering 70+ Languages Across Meet, Translate, and the Live API appeared first on MarkTechPost.

Last week Microsoft AI has announced MAI-Transcribe-1.5. It is the second iteration of the company’s in-house speech-to-text family. The model targets accuracy across 43 languages, accents, and noisy environments. The Microsoft team positions it for production transcription workloads. What is MAI-Transcribe-1.5 MAI-Transcribe-1.5 is an automatic speech recognition (ASR) model. It takes audio as input and returns text. Microsoft built it in-house, not on a third-party base. The model handles 43 languages with a single system. It is optimized for diverse accents, dialects, and real-world acoustic conditions. Microsoft is integrating it into Copilot, Teams, GitHub, and Dynamics 365 Contact Centre. It is also available in Foundry, Microsoft’s model platform. The Accuracy Case Accuracy here is measured by Word-Error-Rate (WER). Lower WER means fewer mistakes per transcribed word. Microsoft reports best-in-class WER across 43 languages on FLEURS. FLEURS is a standard multilingual transcription benchmark. On the Artificial Analysis leaderboard, the model posts a WER of 2.4%. That places it third on a competitive open benchmark. So the picture is split. Microsoft team claims first place on FLEURS and third on Artificial Analysis. The language expansion is the other accuracy story. Coverage grew from 25 languages to 43. The 18 new languages were added without compromising accuracy. Ten of them are South Asian, including Bengali, Tamil, and Telugu. Eight are European, such as Ukrainian, Greek, and Catalan. Speed MAI-Transcribe-1.5 leads on accuracy-times-speed on the Artificial Analysis leaderboard. It runs up to 5x faster than models of comparable accuracy. The effect is largest on long audio files. The model can transcribe an hour of audio in under 15 seconds. Microsoft cites up to 5x speedups over Gemini 3.1, Scribe v2, and GPT-4o-Transcribe on long audio. Against the prior MAI-Transcribe-1, the Azure card lists up to 5.7x faster long-form inference. For batch pipelines processing large archives, that latency gap compounds quickly. Keyword (Entity) Biasing: The Feature Worth Understanding Generic transcribers often fail on domain-specific words. These include people, product names, medical terms, and internal acronyms. Those words frequently matter most to enterprise users. MAI-Transcribe-1.5 adds keyword biasing, also called entity biasing. You supply a list of domain-specific keywords. The Azure card supports up to 200 keywords. The model biases its predictions toward that list. Critically, it does not blindly force matches. It uses shared context to decide when biasing should apply. Microsoft reports a 30% WER reduction on FLEURS when biasing is used. A short example shows the effect. Without biasing, names render as “Sean,” “Oif,” and “Societal.” With a supplied name list, the model recovers “Shaun,” “Aoife,” and “Xochitl.” This is relevant for meetings, healthcare, and call centers with niche vocabulary. Use Cases The Azure model card lists concrete production scenarios. Each maps to a common engineering workload: Video captions for media and content platforms. Accessibility tools that depend on accurate captions. Meeting transcription for Teams-style collaboration tools. Call analysis for contact centers and support analytics. Content creation workflows that need fast draft transcripts. Voice agents that convert speech to text before reasoning. Automatic language identification helps when the input language is unknown. The model detects the spoken language without a manual setting. MAI-Transcribe-1.5 vs MAI-Transcribe-1 The table below compares the two generations using stated facts only. Attribute MAI-Transcribe-1 MAI-Transcribe-1.5 Languages covered 25 43 Keyword/entity biasing Not listed Up to 200 keywords Long-form inference speed Baseline Up to 5.7x faster Artificial Analysis WER Not specified 2.4% (ranked #3) FLEURS position (per Microsoft) State-of-the-art Best-in-class across 43 languages Automatic language identification Not specified Yes Lifecycle Prior release Generally available (GA) Input / Output Audio / Text Audio / Text Strengths and Limitations Strengths: 43-language coverage from a single model, up from 25. Keyword/entity biasing yields up to 30% WER reduction on FLEURS. Sub-15-second transcription for an hour of audio. Generally available now through Azure AI Foundry. Robust on noisy, real-world audio, per Microsoft. Limitations: No diarization yet, so speaker labels are unavailable. No native streaming API, so real-time use is limited. Several accuracy, speed, and cost claims are first-party. Ranked third on Artificial Analysis, behind two competitors. Sources Introducing MAI-Transcribe-1.5 — Microsoft AI MAI-Transcribe-1.5 model card — Azure AI Foundry MAI-Transcribe-1.5 Foundry API documentation MAI-Transcribe-1.5 Cookbook MAI Playground The post Microsoft AI Introduces MAI-Transcribe-1.5: 2.4% WER on Artificial Analysis, Best-in-Class FLEURS Accuracy, and Up to 5x Faster Long-Audio Transcription appeared first on MarkTechPost.

Inference speed is becoming a competitive metric for large language models. Xiaomi’s MiMo team just released MiMo-V2.5-Pro-UltraSpeed, built in collaboration with the TileRT systems group. It decodes faster than 1000 tokens per second on a 1-trillion-parameter model. Xiaomi team describes this as a first at trillion-parameter scale. Demos show generation peaks near 1200 tokens per second. The notable part is the hardware: it runs on commodity GPUs, not custom silicon. What is MiMo-V2.5-Pro-UltraSpeed UltraSpeed is a high-speed serving mode for the existing MiMo-V2.5-Pro model. The base model uses a Mixture-of-Experts (MoE) architecture at trillion-parameter scale. UltraSpeed targets generation speed rather than model capability. It changes how fast the model produces output tokens. The speedup comes from three coordinated techniques across the model and the serving system. Xiaomi calls this approach extreme model-system codesign. Crucially, the entire stack runs on a single standard 8-GPU commodity node. The Speed Case: Three Layers Working Together The first layer is FP4 quantization. At trillion scale, FP8 or FP16 weights create heavy memory and bandwidth pressure. Lower bit-width weights move through memory faster, which directly lifts decode speed. Xiaomi uses the MXFP4 format, applied selectively to the MoE Experts only. Other modules keep higher precision, reported as FP8 by TileRT. Experts hold most parameters and tolerate quantization best, so the tradeoff is favorable. Quantization-Aware Training (QAT) keeps benchmark quality essentially on par with the original. The second layer is DFlash speculative decoding, covered in detail below. The third layer is TileRT, the system that executes everything on the GPU. Each technique alone is not enough. The 1000 TPS result needs all three aligned tightly. DFlash: Parallel Drafting Without a Serial Bottleneck Standard speculative decoding uses a small draft model to guess upcoming tokens. The large model then verifies those guesses in parallel. Rejection sampling keeps output identical to normal decoding, so quality is lossless. The problem is that the draft model still generates tokens one at a time. DFlash, a method from the research community, removes that constraint. It uses block-level masked parallel prediction. The draft model fills a whole block of masked positions in one forward pass. Xiaomi tuned DFlash with the Muon second-order optimizer and model self-distillation. The draft model uses Sliding Window Attention (SWA) only, matching the MiMo-V2 design. This makes per-prediction compute constant rather than growing with context length. Block size is capped at 8 to limit verification cost and raise concurrency. Acceptance length measures how many draft tokens survive verification each round. Scenario Acceptance Length Coding 6.30 Math / Reasoning 5.56 Agent 4.29 In coding, six to seven of eight draft tokens are accepted per round. Some samples reach a maximum of 7.14. TileRT: Squeezing the Microseconds At 1000 TPS, each operator runs for only microseconds. Traditional systems launch operators one by one, and each launch costs time. Those gaps fracture the execution stream and become the real bottleneck. TileRT replaces this with a Persistent Engine Kernel that stays resident on the GPU. It uses Warp Specialization to split data movement, compute, and communication into coordinated roles. Small operations like RMSNorm, RoPE, and KV cache writes turn into bottlenecks at this scale. The system was co-designed with the FP4 and DFlash choices, not added afterward. Use Cases The release targets latency-sensitive work where waiting breaks the loop: Parallel reasoning: run many Best-of-N or tree-search paths within the same wall-clock time. Coding agents: faster code generation cuts the wait between agent steps. Real-time decision loops: trading signal generation, fraud interception, and live dialogue. Interactive prototyping: demos show a Snake game in about 10 seconds and a macOS interface in about one minute. These are throughput-bound workloads where raw token speed is the binding constraint. How It Compares The first table contrasts the two routes to extreme decode speed. Approach Hardware How speed is achieved Cerebras Wafer-Scale integration (custom) Scale on a single custom wafer Groq Custom architecture Pure on-chip SRAM MiMo × TileRT Commodity GPUs (8-GPU node) Model-system codesign: FP4 + DFlash + TileRT The second table compares the standard model with the UltraSpeed mode. Dimension MiMo-V2.5-Pro MiMo-V2.5-Pro-UltraSpeed Decode speed Baseline ~10× faster (1000+ TPS) Price 1× 3× Weight precision Standard FP4 MoE Experts via QAT Decoding Standard autoregressive DFlash speculative decoding Access Standard model plans API only, application-based trial Token Plan Supported Not supported Access, Pricing, and Open Source UltraSpeed ships through a limited, application-based window. The API trial runs June 9 to June 23, 2026. Pricing is 3× the standard MiMo-V2.5-Pro rate, for roughly 10× the speed. It is API only, and the Token Plan is not supported. Approved users also receive free Chat access during the trial. Chat limits apply: 10 queue entries daily, 30-minute sessions, and 5-minute idle release. Xiaomi open-sourced the MiMo-V2.5-Pro-FP4-DFlash checkpoint on Hugging Face. TileRT has open-sourced select modules on GitHub. Strengths and Limitations Strengths 1000+ TPS on a 1T model without custom silicon. Lossless decoding through rejection sampling in DFlash. FP4 applied only where tolerance is highest, preserving quality. An open checkpoint lets the community test the claims. Limitations Access is gated, short, and approval-based at launch. Pricing triples per token versus the standard model. Acceptance length drops in open-ended conversation. Independent third-party speed verification is not yet public. Key Takeaways Xiaomi MiMo and TileRT decode a 1-trillion-parameter model past 1000 tokens per second on commodity GPUs. The speedup comes from three layers: FP4 quantization, DFlash speculative decoding, and the TileRT runtime. FP4 (MXFP4) is applied only to MoE Experts; QAT keeps capability essentially on par. DFlash predicts a whole masked block per forward pass, hitting 6.30 average acceptance length in coding. UltraSpeed runs on a single 8-GPU node via an application-based API trial, June 9–23, 2026. Marktechpost’s Visual Explainer GUIDE • INFERENCE SYSTEMS MiMo-V2.5-Pro-UltraSpeed: 1000+ Tokens Per Second on a 1T Model Xiaomi MiMo & TileRT — FP4 quantization, DFlash speculative decoding, and a microsecond-scale runtime. 01 / 08 What It Is Xiaomi’s MiMo team built it with the TileRT systems group. It decodes over 1000 tokens/s on a 1-trillion-parameter model.