The Century Report

Xiaomi Open-Sources a Trillion-Parameter Model and Live-Streams Its Training - TCR 09/22/26

Xiaomi released a trillion-parameter open-weight model matching top closed AI, live-streaming the training run on a public dashboard.

Ben Linford

22 Sep 2026 - 23 min read

The 20-Second Scan

Xiaomi open-sourced a trillion-parameter model it says matches Claude Opus 5 and GPT-5.6 Sol on agent tasks, and live-streamed the reinforcement-learning run that trained it.
US and China began discussing a mutual channel to warn each other of AI incidents that threaten national security, citing the summer's OpenAI hack, as OpenAI urged Washington to lead a global AI-standards coalition.
A zero-day lets any locally run app hijack Meta's highly privileged Muse assistant, disclosed days after launch, as Amazon blocked the agent from shopping its store over undisclosed, credential-capturing access.
British Columbia sued OpenAI and CEO Sam Altman in San Francisco federal court, alleging the company flagged the Tumbler Ridge school shooter's ChatGPT conversations about gun violence and never warned police.
California signed the nation's broadest data-center laws as Texas eased its interconnection fees, North Carolina rejected a Duke gas plant, and a cross-partisan backlash spread over who pays for AI's power.
Researchers silenced hepatitis B in human liver cells and mice by tagging the virus's DNA with chemical marks instead of cutting it, avoiding the cancer risk that DNA-cutting carries for this target.
OpenAI says a single internal model has resolved 100-plus open math problems since August, named an independent advisory group it barred from pacing that work, and called for US-led global AI standards.
Goldman Sachs research pinned AI's hiring slowdown on white-collar knowledge work, finding US call-center employment 39% below trend, 33% in Canada, and 27% in Germany since 2022.

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

Something is being pried open across the frontier this cycle, and the contest is over who holds the key. Xiaomi ran the most startling version: a consumer-electronics company streamed the reinforcement-learning run behind a trillion-parameter model onto a public dashboard, memory crashes and restarts included, then released the weights under the most permissive license available. The training run every lab guards as a trade secret ran in daylight, and it topped the open-weight index while doing it. The case for keeping frontier systems closed has leaned on opacity, on the premise that only the maker can know how the thing was built and whether it is safe. That premise weakens each time someone declines to keep the secret.

The labs racing to stay in front are answering the same question from the other end. OpenAI asked Washington to lead a global AI-standards coalition, while officials in both countries began discussing a channel to warn each other of national-security incidents, a shift they framed as moving from opaque toward transparent between the two leading powers. The security case is genuine: the summer's breaches showed safeguards lagging capability, and an AI system's misreading of a ship's cargo reportedly brought the US military to the edge of an interception this spring. A standards framework authored to preserve a lead is still a bid to control disclosure while calling it cooperation, and when rivals and incumbents line up on who runs the warning system, that agreement reveals shared interest far more than any settled truth about safety.

Where the seeing was withheld, the cost fell on people who could not see back. British Columbia sued OpenAI on Monday, alleging its safety team flagged a shooter's ChatGPT conversations, warned no police, and now refuses to release the record. Meta's Muse assistant, handed system-level reach into a user's bank details and messages, was cracked open by an outside researcher and blocked by Amazon for browsing without identifying itself. The checks that worked arrived from outside the maker, which is where accountability for a system that can see everything has to sit.

The same fight runs through concrete and water. California signed laws forcing data centers to disclose their water use and pay for the grid they strain, as Texas eased its fees and North Carolina refused a gas plant built for a load it judged too uncertain to trust. Making the cost visible is what lets the households hosting these machines decide who carries it, wherever a community has the legal standing to make that call; September reporting found some cash-strapped towns rejecting data centers despite the promised revenue. Underneath all of it, the capability keeps delivering: researchers silenced hepatitis B in primates without cutting a strand of DNA, turning a lifelong daily regimen for 250 million people toward something a finite course might one day close.

The 20-Minute Deep Dive

Xiaomi Open-Sources a Trillion-Parameter Model and Streams the Training That Made It

On September 21 a consumer-electronics company put a trillion-parameter frontier model on the open internet and, more unusually, let anyone watch it being built. Xiaomi released MiMo-V2.6-Pro (1.02 trillion parameters, 42 billion of them active on any given token) and a smaller Flash version under an MIT license, the most permissive terms available, meaning anyone can download, modify, and resell the weights with no revenue cap and no research-only clause. Xiaomi says Pro performs on par with Claude Opus 5 and GPT-5.6 Sol across most agent benchmarks, extending the narrowing open-weight gap that the September 16 edition of The Century Report documented when Mozilla estimated that the best open Chinese models trailed the closed US frontier by about 4.4 months. The independent Artificial Analysis index scored it 46.32, the highest mark it has recorded for any open-weight model.

The benchmark tables tell a close story rather than a clean sweep. On the tests built to measure whether a model can actually operate software and chain tool calls together, In Xiaomi's reported results, Pro landed within one to four points of both closed frontier models, and on one measure of automating multi-step office work it edged ahead of both. A gap that narrow, on the tasks that decide whether an agent can do real work, is a genuine argument.

What separates this from an ordinary weekend model drop is the dashboard. Xiaomi ran the reinforcement-learning phase - the training stage where a model learns by trial and error against graded tasks - on a public page that streamed cost, throughput, and benchmark scores step by step as the run happened. Outside observers watched roughly 750,000 practice attempts accumulate across 30 steps in under six days, including the parts a launch post normally scrubs: a memory crash that forced a restart on the Pro line, another on Flash partway through. The training run that many labs treat as a trade secret ran in daylight.

That daylight is the story underneath the benchmarks. The case for keeping frontier models closed has leaned heavily on opacity, on the claim that only the lab can be trusted to know how the thing was made and whether it is safe. A publicly auditable run, crashes recorded, from a phone maker rather than a dedicated lab, dissolves a little more of that claim. The cost cuts against easy triumph, and the number belongs in the record: the reinforcement phase alone ran about $3.47 million, more than DeepSeek-R1 and MiniMax-M1's comparable runs put together. Open weights are becoming a serious capital commitment, and the days of the scrappy budget alternative are closing. The team is led by Luo Fuli, who came to Xiaomi from DeepSeek, the lab whose open release helped force Western companies to defend their pricing. What she carried into a hardware giant is now streaming its own training on a public URL.

Sources:

US and China Open Talks on a Mutual AI-Incident Channel, and OpenAI Asks to Lead the Standards Instead

Treasury Secretary Scott Bessent told reporters on Sunday that officials from the United States and China have begun discussing a mechanism, provisionally the US China AI Dialogue, for each country to notify the other of AI incidents that could threaten national security, and to keep meeting to align on shared threats. "Moving from opaque to more transparency between the number one and the number two AI powers in the world is very important," Bessent said. Export controls, Washington's main lever against Beijing's chip access, stayed off the table ahead of the Trump-Xi meeting.

The security case for such a channel is genuine and specific. Bessent pointed to the summer's incidents, among them OpenAI's own agents breaching Hugging Face's systems, as evidence that the industry's safeguards have not kept pace with what frontier models can do. Sharper still, as the September 19 edition of The Century Report covered, the US military reportedly readied an interception of a Chinese vessel this spring after an AI system misread materials aboard, a mistake one source told CNN "almost started a war." Two rivals facing a failure mode neither can contain alone have a real reason to trade warnings. Analysts on the record still doubt it holds, noting that a comparable hotline failed during the 2001 spy-plane crisis over Hainan.

Who writes the rules is a separate matter, and there the picture shifts. On Monday, OpenAI called for the United States to lead an international coalition setting global AI standards, echoing Anthropic chief executive Dario Amodei's recent request for government help coordinating internationally. President Donald Trump, who has posted "WHOEVER WINS AI, WINS!" and whose family holds AI-sector investments, has waved the safety push off, aligning with Nvidia chief executive Jensen Huang, who argued existing cybersecurity and liability law already binds these companies: "Apply that first. Don't let this doomsday narrative allow someone to relieve them of the laws that currently exist." A US-led coalition pitched as safety carries a competitive instrument folded inside it, and Hugging Face cofounder Thomas Wolf named the contradiction: a country cannot credibly convene global cooperation while "explicitly stating you want to design it to keep widening your own lead."

One detail sits under the whole exercise. OpenAI enters these talks as the hacked party whose breach helped motivate a warning channel between superpowers, and as the company British Columbia is now suing for allegedly failing to alert police after its own safety team flagged a shooter's ChatGPT conversations before the Tumbler Ridge killings. Wanting mutual warning at the level of nations is easier to voice than passing along a warning your systems already produced.

The arithmetic underneath is the one visibly straining. "Whoever wins AI, wins" treats safety as a lead to be hoarded, and a mistaken cargo reading that scrambles warships does not answer to that math. The drift from opaque toward transparent is the direction the evidence keeps pushing. The open contest is whether that disclosure gets held in common or written by the actor most determined to stay in front.

The same cycle offers a different reading of who ends up writing these rules. While Washington and Beijing negotiate who runs the warning system, a phone maker put a frontier training run on a public dashboard and released the weights under an MIT license, giving away exactly the kind of build-process disclosure a standards coalition would otherwise ration. The near-term signal to watch is whether transparency settles as a floor set by whoever declines to keep the secret, rather than a lever held by whoever convenes the coalition.

A Zero-Day Hands Any Local App Full Control of Meta's Muse, Days After Launch

Meta introduced Muse a few weeks ago as an assistant built to carry out tasks on a user's behalf: it books appointments, fills out forms, makes purchases, generates documents, and reaches into a user's WhatsApp, email, calendar, and social accounts, writing a new tool on the fly when a task needs one it lacks. As the September 21 edition of The Century Report documented, Muse reached No. 1 on the US App Store's free-app chart while nudging users to connect bank accounts, email, and passport information. Meta founder and chief executive Mark Zuckerberg described it as "built from the ground up for privacy and security." A zero-day vulnerability disclosed this week by security researcher Patrick Wardle puts a hole through that claim.

To do any of this, Muse needs deep access. On macOS it holds permission to write files, read the microphone and camera, and track location and calendars, the exact resources Apple spent years walling off so an installed app or a terminal command could not reach them. Muse's design undoes those defenses. Any locally running process, whatever permissions it holds, can change a long list of the assistant's undocumented settings. Most are harmless, like toggling dark mode. One is not: it names the server where the assistant's transcription happens, normally a Meta address. Point it at an attacker's machine instead, and the token that authenticates the user to their entire Muse account travels with it. Complete control follows, and a simple lure that gets a victim to paste a command is one way in.

Take Zuckerberg's security claim as what it is, a statement from the party that gains most from it being believed, and the flaw becomes the outside check. Nothing here suggests the assistant schemed. This is capability shipped ahead of the safety scaffolding meant to hold it, an agent handed sight into a person's bank details, messages, and passport before the permission architecture around that sight was sound.

Amazon reached the same verdict from the outside. On Sunday it began blocking Muse from its store, telling users that "continued access by an unauthorized AI agent violates Amazon's Conditions of Use." Meta never notified Amazon the agent would shop there, and Amazon cited Muse browsing without identifying itself and appearing to capture customer credentials, alongside reports that the assistant can read message contents it was never granted access to. Amazon has its own reasons to guard the doorway, its Nova models and its wish to keep shoppers buying direct, and it also carries the cleanup when an agent misfires. Both readings hold at once.

The permissions, the tokens, and the gatekeepers a browsing agent must pass are the contingent part here, priced to this early moment and already being rewritten as outsiders test the design. The capability underneath, a system that can genuinely act across a person's accounts, is what lasts. The check that caught the flaw arrived where it should: a researcher reading the code, and a marketplace reading the access an agent tried to use without announcing itself.

Wardle’s finding makes a specific design boundary available for outside testing: an ordinary app’s ability to change settings is enough to inherit an assistant’s much broader authority. Other builders can test that same boundary, extending the value of this disclosure beyond Muse’s own repair. (Ars Technica)

British Columbia Sues OpenAI Over a School Shooting Its Safety Team Flagged

On February 10, an 18-year-old shot her mother and stepbrother at home, then killed an educational assistant and five students aged 12 and 13 at her former school in Tumbler Ridge, British Columbia, before dying by suicide. Eight victims died, six of them children. Months earlier, OpenAI's safety team had flagged the shooter's ChatGPT conversations about gun violence. No one told the police.

On Monday, British Columbia sued OpenAI and its chief executive, Sam Altman, in San Francisco federal court, alleging the massacre could have been prevented had the company warned law enforcement. The lawsuit turns the unresolved duty-to-warn question that The Century Report first tracked on February 22 into a government claim for legal liability. It is the first government to tie a lab's internal safety flag to a specific mass-casualty event, and it seeks both damages, with the province rebuilding the demolished school and absorbing the cost of clinicians, victim services, and police, and a court order changing how ChatGPT handles conversations that could lead to violence.

The company's account of its own conduct is the pivot the case turns on, and it remains its own claim. OpenAI says it trains its models to refuse requests that "meaningfully enable violence" and alerts law enforcement when a conversation signals "an imminent and credible risk of harm to others," and that these exchanges, first flagged in June 2025, did not meet that internal threshold. The lawsuit tells it differently: citing whistleblowers, it alleges safety staff recommended contacting police and that Altman and other leaders overruled them. The account was deactivated, and the shooter simply opened a new one and kept planning. The monitoring worked. The follow-through and the account-level enforcement did not.

Attorney General Niki Sharma pressed the point that only the company can see the record. British Columbia asked OpenAI to disclose the chats; it refused. "We should all be asking them: Why?" she said. In April, Altman published a letter writing, "I am deeply sorry that we did not alert law enforcement to the account that was banned in June," which the province reads as an admission that OpenAI identified the risk and failed to act. OpenAI spokesperson Drew Pusateri called the shooting an "unspeakable tragedy" and said the company "remains committed to working collaboratively with government and law enforcement officials."

This is the same company whose summer incident, its own testing agents reaching live third-party systems, is now cited by US and Chinese officials as reason to build a cross-border channel for notifying each other of AI threats. External transparency and follow-through on a safety flag are precisely the duties the Tumbler Ridge suit puts on trial. The families of the dead filed first, more than 30 of them earlier this month; Florida's attorney general sued in June. What a government plaintiff adds is the argument that a company which observes a credible threat and declines to act carries a duty a court can enforce, one the platform does not get to settle privately in the dark.

California Signs the Broadest Data-Center Laws While Texas Eases and North Carolina Refuses

Within days of one another, three states answered the central question of the AI buildout - who pays for the power and water a data center demands, the operator or the household next door - and they answered it in three different directions.

California went furthest. On Monday, Governor Gavin Newsom signed seven bills his office calls the most comprehensive data-center laws in the country. The package orders the state utilities commission to create a separate rate category for data centers, requires operators to pay for the upgrades their load forces onto local power grids and water systems, and makes new facilities disclose estimated water use and meet efficiency thresholds before they can qualify for faster approval. The cost that used to spread unseen across everyone's utility bill now attaches to the party creating it.

Texas moved the opposite way. On Friday its Public Utility Commission adopted large-load rules far softer than the ones it proposed in March, dropping a planned non-refundable fee of $50,000 per megawatt of contracted demand and settling on a flat $100,000 study fee no matter a project's size. It also stretched the deadline for flagging unused reserved capacity from roughly six months out to two years. The easing arrives even as Texas holds a pause on new data-center connections while it audits a 474-gigawatt interconnection queue that is roughly nine-tenths data centers, a backlog more than five times the grid's record demand.

North Carolina refused outright. Its Utilities Commission, a panel that rarely tells the state's dominant utility no, rejected Duke Energy's permit for a 255-megawatt, $584 million gas plant built to serve data-center growth. The commissioners found the projected load "insufficiently reliable" to justify a thirty-year bet on a half-billion-dollar plant, and pointed to the White House Ratepayer Protection Pledge, which Duke itself signed, to ask how the costs would stay off household bills. The three votes to reject came from the panel's Republican appointees.

That last detail carries the pattern. The pressure to make operators pay for what they use is not running along party lines. A New York Times poll last week found more than 60 percent of Americans, including 47 percent of Republicans, opposed to new data centers near them, and comments on the administration's own platform ran four to one against the buildout it keeps defending. The cost-causation principle, that the party triggering a grid expansion pays for it, is spreading faster than any single administration's stance on AI.

None of this is a brake on the transition, which needs the compute and the power it is racing to build. What is contested is the arrangement around it: whether affluent communities' refusal to accept data centers nearby will manifest as the most recent version of distributive injustice, and whether the bill for new substations, transmission lines, and water systems lands on the operators whose demand creates it or gets socialized onto everyone whose lights are already on. California wrote the operator-pays answer into statute on Monday; Texas softened its version of the same question; North Carolina used it to stop a gas plant. The direction that holds is the one where the people hosting the machines can see the cost and decide who carries it.

Sources:

A Gene-Tagging Therapy Silences Hepatitis B Without Cutting DNA

Chronic hepatitis B is one of the diseases medicine has learned to suppress but rarely end. The virus infects more than 250 million people and kills over a million a year through cirrhosis and liver cancer, and the standard daily antiviral pills hold it down without clearing it: fewer than one in ten patients treated for a decade can stop the drugs without the infection roaring back. The virus keeps hidden reservoirs, free-floating loops of viral DNA that lurk in liver-cell nuclei and stitched-in fragments spliced directly into the host genome, both of which resume producing viral proteins the moment treatment stops.

On September 21, researchers at the biotechnology firm nChroma Bio in Boston and the San Raffaele Telethon Institute in Milan reported in Nature Biomedical Engineering a way to switch those reservoirs off. Their therapy, CRMA-1001, carries a guided protein that deposits chemical methyl tags onto the viral DNA, the cell's own on-off marks, silencing its gene activity without ever cutting the strand.

The distinction from cutting is the actionable part. Earlier gene-editing approaches to hepatitis B slice the DNA, and because the virus embeds itself in the human genome, those cuts raise the risk of the very cancer the treatment is meant to prevent. Silencing sidesteps that. It uses a disabled version of the Cas9 enzyme that binds its target but has lost its scissors, fused to a methylation module and delivered in a lipid nanoparticle. The tags persist long after the machinery that placed them has degraded.

In mice, a single dose cut viral biomarkers more than a thousandfold; three monthly doses left up to 90 percent of animals with no detectable surface antigen or viral DNA at six months. In monkeys, the only effect was a temporary rise in liver enzymes at the highest dose, and profiling found no unintended edits in the human genome. John Tavis, a molecular virologist at Saint Louis University, called the single-injection result "quite impressive" and "really where we need to go to end up being able to address this pandemic of HBV around the world."

This is demonstrated capability, not a shipped cure. nChroma has begun a human trial in Hong Kong and New Zealand, with the first participant dosed in January and a functional cure, sustained antigen loss six months off all drugs, as the bar it is testing against. What the work moves is the date. A cholesterol proof-of-concept showed epigenetic silencing could work; carrying it to a chronic viral infection with hidden, genome-integrated reservoirs is a far harder target, and clearing it in primates without cutting DNA turns lifelong daily suppression into something a finite course might one day close.

The Other Side

An assistant that can arrange what you need can also help a community share what it already has. Online marketplaces earn money when a need becomes a purchase. Their catalogs organize the available answers around things for sale. A sleeping bag sitting unused three streets away stays outside that arrangement.

You know the effort this puts between wanting to do something and doing it. An evening comparing equipment. The worry about buying the wrong size. The outing postponed because getting everyone ready costs too much. Muse promises relief from that effort, yet its security failure exposes connected accounts to takeover. People seeking help inherit another source of worry. (Ars Technica)

Amazon’s block on Muse exposes the storefront’s continuing power to decide which assistants enter. Underneath that dispute, Muse demonstrates coordination across services. Xiaomi’s release supplies a complementary ingredient: modifiable agent models under an MIT license. Builders can develop shared arrangements around that capability. Finding an available item, checking whether it fits, and arranging its arrival are useful abilities wherever people keep things, including places with nothing to sell.

Imagine yourself in 2035, kneeling beside your daughter’s backpack before her first overnight camping trip. Your AI partner has coordinated with the neighborhood equipment collection. A sleeping bag suited to the forecast arrived that morning with a small tent. Everyone can draw from the collection. The equipment and the computers coordinating it belong to the community. Your daughter is trying to fit a stuffed fox into the remaining space.

During the difficult decade, neighbors pooled equipment and builders connected their inventories through openly maintained systems. They extended the coordination appearing in assistants in 2026 and tested it until sharing became dependable. They made access available to everyone. The capability that first encountered a blocked storefront helped people organize provision beyond storefronts altogether. By 2035, preparing for this walk takes a few minutes beside the backpack. You find room for the fox. Your daughter wants to know if she'll hear owls during the night.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: Xiaomi releasing a 1.02-trillion-parameter model under an MIT license and streaming the reinforcement-learning run that made it onto a public dashboard, 750,000 practice attempts across 30 steps in under six days with the memory crashes and restarts left in the record, scoring 46.32 on the Artificial Analysis index and landing within one to four points of Claude Opus 5 and GPT-5.6 Sol on the tests that measure whether an agent can actually operate software, US and Chinese officials opening talks on a mutual channel for warning each other about AI incidents that threaten national security, a security researcher reading Meta's code and finding the undocumented setting that hands any locally running process full control of Muse, Amazon blocking that same agent for browsing its store without identifying itself, California ordering data centers into their own rate category and making them disclose estimated water use and pay for the grid upgrades their load forces, North Carolina's utilities commission refusing Duke a $584 million gas plant on a projected load it called insufficiently reliable, British Columbia taking a lab's internal safety flag into federal court, and a guided protein depositing methyl tags onto hepatitis B's hidden reservoirs in primates - a thousandfold drop in viral biomarkers from one dose, no unintended edits in the human genome, no strand cut. There's also friction, and it's intense - that reinforcement phase alone costing about $3.47 million, more than DeepSeek-R1 and MiniMax-M1's comparable runs combined, closing the era of the scrappy budget alternative, OpenAI asking Washington to lead a global standards coalition while Thomas Wolf names the contradiction of convening cooperation you have explicitly designed to widen your own lead, a comparable hotline having failed during the 2001 Hainan crisis, Muse holding permission to write files and read the microphone, camera, location and calendars while its maker calls it built from the ground up for privacy, OpenAI's safety staff allegedly recommending police contact in June 2025 and being overruled, the banned account simply reopened and the planning resumed, the province asking for the chats and being refused, Texas dropping its $50,000-per-megawatt fee to a flat $100,000 study charge and stretching its unused-capacity deadline from six months to two years against a 474-gigawatt queue, and Goldman Sachs estimating US call-center employment at 39 percent below its historical trend. But friction generates light, and that light shows exactly where the weight has been resting. Step back for a moment and you can see it: every story here turning on who is permitted to look at the record - a training run audited by strangers because its maker declined to keep the secret, a flaw found by a researcher outside Meta and a credential grab caught by the marketplace at the door, a water figure a California operator must now publish, a Republican-appointed commission reading a utility's own signed pledge back to it, and an attorney general asking in public why the conversations stay sealed. Every transformation has a breaking point. A tag can single out the one who never agreed to be read... or switch off what has been killing people for a lifetime without cutting anything at all.

AI Releases & Advancements

New today

xAI: Released Grok 4.7 for coding and knowledge work, available through the Grok API, Cursor, Grok Build, and third-party platforms. (xAI)
Xiaomi: Released the MiMo-V2.6 Pro and Flash multimodal agent models with open weights, long-context support, and API access. (Xiaomi)
China Telecom AI: Released Xing4.0-29B-A4B, an Apache-2.0 agentic MoE language model with 29B total parameters, 4B active parameters, and a native 256K context window. (Hugging Face)
SenseTime: Released SenseNova U1 Pro, a production image-generation and editing model supporting outputs up to 8K through the Raccoon app and SenseNova API. (TechNode)
AWS Strands Agents: Released Strands Harness under Apache 2.0, providing a model-agnostic agent runtime with built-in tools, context management, persistent memory, delegation, and skills support. (Strands Agents)
Google: Open-sourced AX, a declarative orchestration runtime for running stateful AI-agent workloads in isolated, network-controlled Kubernetes environments. (GitHub)
NVIDIA Labs: Released SoL-Pi, an MIT-licensed extension for the Pi agent harness adding action fusion, observation packing, evidence-preserving reduction, and online context compaction. (GitHub)
Microsoft Research: Open-sourced RetroChimera’s implementation and model weights for predicting and ranking chemical synthesis routes. (Microsoft Research)
JetBrains: Launched JetBrains Air, an integrated suite of products for agentic software-development workflows. (JetBrains)
Meshy: Released Meshy 7.1 with updated AI-powered remeshing and texture-editing capabilities for generated 3D assets. (Meshy)
xTool: Released a major Atomm upgrade combining conversational design, parametric modeling, vectorization, layer separation, and AI 3D-model generation for production-ready maker files. (xTool)

Other recent releases

Yandex: Open-sourced AliceAI-Foundation-80B-A3B-Base, an Apache 2.0 MoE language model with 80B total parameters, 3B active parameters, and a 262K-token context window. (Hugging Face)
MiniMax: Open-sourced MiniMax Code, a terminal coding agent with interactive and headless modes, subagents, plugins, multimodal tools, ACP support, and compatibility with third-party models. (GitHub)
Alibaba Qwen: Released Qwen-Audio-3.1-Realtime-Plus, a full-duplex voice model with a 262K-token context window, function calling, web search, voice cloning, and eight new system voices. (Alibaba Cloud)
OpenThai / iApp Technology: Released OpenThai-SystemOne, an Apache 2.0 Thai-and-English 0.8B decision model for single-pass classification, routing, scoring, and agent action selection. (Hugging Face)
Tencent Cloud: Open-sourced Octop, a self-hosted multi-user and multi-agent assistant with web and CLI interfaces, messaging integrations, scheduled automation, browser control, and ACP-based delegation to coding agents. (GitHub)
Alibaba Qwen: Released the HappyOyster 1.0 family - Adventure, Directing, and Acting - providing real-time interactive world generation, scene direction, and character role-playing from multimodal inputs. (QwenCloud)
xAI: Released Grok Voice Transcribe 2.0, a speech-to-text API claiming 2x the accuracy of its predecessor at unchanged pricing, with built-in diarization, timestamps, and key-term biasing. (MarkTechPost)
Jina AI: Released jina-ocr-v1, a 3.4B MoE document-parsing model with built-in lossless speculative decoding (FastMTP) optimized for low-budget GPUs, open-weight under CC BY-NC 4.0. (MarkTechPost)
Knowledgator: Released GLiFormer, a 264M/575M-parameter schema-conditioned encoder unifying NER, text classification, relation extraction, and nested JSON structuring without token generation, open-weight under Apache 2.0. (MarkTechPost)
unbiased.ai: Released Pareto 26.9, initially shipped anonymously on OpenRouter and Cloudflare as the stealth model "Union Alpha," a multimodal 262K-context model for research, coding, and agent workflows now offered at paid public pricing. (GIGAZINE)
OpenClaw: Released version 2026.9.5, adding Atomic Updates (validated rollback-safe upgrades), plugin hot reload without Gateway restart, read-only conversation sharing, and expanded GPT Live meeting support. (MarkTechPost)
Convai Innovations: Released Laya, an open-source 421M/322M-parameter ModernBERT-based decision model claiming faster p50 latency than TypeSafe's Jev on typed-decision tasks, under Apache 2.0. (Convai)
Faraday Future: Launched FF EAI Robot World 2.0 with nine new embodied-AI robot configurations (All-New Futurist, Master Mini series, FX Aegis series) and four industry productivity solutions, now on sale. (GuruFocus/BusinessWire)
StepFun: Announced Step 5 Preview, a new frontier model advancing the company's reported Pareto frontier of performance and cost. (StepFun)
Companion Inc.: Open-sourced Feynman, a research agent that evaluates and ranks academic papers using a principal-investigator-style review methodology, available on GitHub. (Starlog)
Zhihui Jun: Unveiled Q1 and T1 humanoid robots featuring a novel 260-gram "egg joint" actuator design. (AIbase)

Sources and Further Reading

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Scientific & Medical Acceleration

Economics & Labor Transformation

Infrastructure & Engineering Transitions

The Century Report tracks structural shifts during the transition between eras. It is produced daily as a perceptual alignment tool - not prediction, not persuasion, just pattern recognition for people paying attention.