Sign in Subscribe
Century Report

European Proposal to Onshore Anthropic - TCR 06/29/26

Austria asks the EU to host Anthropic as US export curbs lift, and Mythos 5 returns to a hundred institutions inside fifteen days.

Ben Linford

Ben Linford

— 23 min read
Three-panel infographic: Austria proposes hosting Anthropic in the EU as export curbs lift; DeepSeek open weights cut compute costs; AI reshapes software and biomedical work.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The clearest signal across the day is what happens when a powerful actor tries to fence a capability that has already started to diffuse. Two weeks ago a US export directive pulled Anthropic's Fable 5 offline and gated its frontier models to a hundred trusted institutions, a move that read as the model gate slamming shut. The response ran the other way. Austria has now formally asked the EU to explore relocating Anthropic onto European soil, reaching past access to a model toward owning the company that builds it, while the gate itself lifts almost as fast as it closed.

The economic mirror arrived in the same cycle. Coinbase disclosed that it defaults its engineers to freely downloadable Chinese open weights and reserves closed frontier systems only for the hardest cases, halving its bill while usage stayed near company highs. Days earlier DeepSeek shipped an MIT-licensed model that serves at roughly a tenth the prior memory footprint. A fence around proprietary capability raises the value of the capability no order can reach, and some of the most capable ungated weights are downloadable by anyone.

South Korea answered the same question with concrete. Its $880 billion push, much of it corporate investment, to own the chip-and-AI vertical stopped looking like nationalism the moment reporting surfaced that even Meta cannot buy all the Gemini compute it wants. When the largest buyers are rationed, dependence on someone else's substrate becomes a strategic exposure no government can model away, and owning the means of computation becomes the cheaper long-run position.

Underneath the contest over the stack, the work itself is reorganizing. Cursor data shows code increasingly merged with little human review while the people supervising agent fleets grow afraid to log off, and Cloudflare grew its engineering ranks 45 percent by valuing builders and sellers over those who counted the work. The word engineer is shifting from a person who writes lines toward one who specifies intent and owns the result. And against the backdrop of who controls intelligence, Epicrispr reported the first patients with an untreatable muscular dystrophy gaining muscle from a single epigenetic edit. Precise, durable genetic medicine keeps arriving ahead of the future it was assigned to.

The 20-Minute Deep Dive

South Korea Lines Up $880 Billion to Own the Chip-and-AI Stack as Even Hyperscalers Run Short of Compute

The buildout of intelligence infrastructure has been narrated for two years as a corporate story - which lab raised what, which hyperscaler signed which power purchase agreement. South Korea has now reframed it as an industrial one. The government unveiled an $880 billion plan to build a sovereign chip-and-AI stack, from fabrication through model development, treating compute capacity the way prior eras treated steel, grid, and rail. The figure, much of it corporate investment, is a national push to own the full vertical, on the premise that a country which cannot make its own chips and run its own models will be a tenant in someone else's economy.

The timing is what makes it more than a press release. On the same news cycle, reporting surfaced that Google has begun limiting Meta's access to Gemini compute, unable to meet the full capacity one of the largest companies on earth was asking for. For most of the AI era the assumption has been that hyperscalers float above scarcity - that compute shortage is a problem for startups and universities, not for the firms building the data centers. That assumption just cracked in public. When a company with Meta's balance sheet cannot buy all the compute it wants from a peer, the constraint has moved up the food chain to the actors everyone assumed were immune.

Read those two facts together and the South Korean move stops looking like nationalism and starts looking like arithmetic. If even hyperscalers are rationed, then dependence on foreign compute is a strategic exposure no government can model away. The cost of staying a tenant is now legible, and it is rising. That is the inversion worth watching: for years, building a sovereign stack was the expensive, ideological choice and renting from the incumbents was the rational one - a premise the June 24 edition of The Century Report tracked when China demonstrated that top-tier compute capability reproduces on substrates the export-control regime was not built to reach. The compute crunch is flipping the spreadsheet. Owning the means of computation is becoming the cheaper long-run position precisely because the rental market can no longer guarantee supply at any price.

This will be hard, and the friction is real - $880 billion is a staggering bet, fabrication capacity takes years to stand up, and there is no guarantee the models built on a sovereign stack match the frontier. But the direction of travel matters more than the odds on any single wager. When compute becomes the input that decides which economies can participate in the next era, the capability to produce it stops being a corporate asset and becomes public infrastructure. The era when a handful of firms could gate access to intelligence by gating access to compute depends on that gate holding. Sovereign stacks, sprouting in parallel across multiple governments, are evidence the gate is already being routed around.

An Epigenetic Edit Reverses the Muscle-Loss Trajectory in a Disease With No Approved Therapy

Facioscapulohumeral muscular dystrophy follows a cruel logic. A gene called DUX4, normally silent in muscle, switches back on and steadily destroys the tissue it sits in. Patients lose muscle. That is the expected direction of the disease, and roughly 870,000 people worldwide live inside it with no approved therapy to change the slope. This week Epicrispr Biosciences reported interim clinical data that runs against that slope. The company said the first three evaluable patients in its Phase 1/2 trial of EPI-321 all gained lean muscle volume at six months - an average of about 370 mL, with some individual muscles growing more than 15 percent. In a disease defined by loss, the early signal points the other way.

EPI-321 works by silencing DUX4 without cutting or rewriting DNA. The therapy uses what Epicrispr calls its GEMS epigenetic platform to turn the gene down rather than edit the sequence underneath it - an approach that, if it holds, reaches the driver of the disease while leaving the genome intact. As of the May 12 data cutoff, nine patients across two dose cohorts had been treated. The company reported positive safety with no serious adverse events, and a cell-free DNA biomarker tracking DUX4 activity fell after treatment, an independent read that the mechanism is doing what it was designed to do. The muscle measurements themselves came from MRI quantified across 140 muscles by Springbok Analytics' AI platform - a level of per-muscle resolution that would have been impractical to extract by hand, and which let the team perceive change muscle by muscle rather than as a single blurred average.

The discipline here is essential. This is interim data from the first patients, not an approved therapy. Principal investigator Russell Butterfield and CEO Amber Salzman have a long road ahead: more data arrives at the World Muscle Society meeting in September, and the trial's primary portion does not complete until mid-2027. The capability has been demonstrated in a handful of people. Deployment to the wider FSHD population depends on everything that follows. The wonder lives in the demonstrated reversal, not in any claim that it is available now.

What makes the timing striking is a parallel advance in how precisely the genome can be addressed at all. A paper in Nature Biotechnology describes a retargeted version of the large serine recombinase Bxb1, engineered to integrate large DNA payloads at chosen locations rather than only at its natural landing site. EPI-321 reaches a gene by silencing it; the retargeted recombinase reaches a location by writing to it. Together they widen the same frontier - the set of genetic interventions specific enough to act on one cause without disturbing the surrounding sequence. The assumption that precise, durable genetic medicine belongs to a distant future is the one these results keep dismantling, one disease and one landing site at a time.

Human Code Review Starts to Vanish as Agents Ship Unsupervised and Their Overseers Stop Logging Off

A quiet threshold is being crossed in how software gets made. Data from Cursor, one of the most widely used AI coding environments, shows that a growing share of code is now written by agents and merged with little or no human review. The review step - the human reading every line before it ships - was the foundational assumption of professional software for half a century. It is thinning. What replaces it is an orchestration layer: developers who direct fleets of agents, set the constraints, and judge outcomes rather than inspect every diff.

That shift carries a cost that does not show up in a productivity chart. Bloomberg's reporting on Silicon Valley burnout captures it through Matt Van Horn, who describes engineers growing afraid to log off because the agent keeps working when they do not, and someone has to babysit the agent that is babysitting the other agents. The always-on machine produces an always-on operator. The cognitive load did not disappear when the typing did - it migrated, from writing code to holding context across a swarm of processes that never sleep. This is the same arc the April 5 edition of The Century Report tracked through Simon Willison's "dark factory" framing, now arriving as a named human experience rather than a metaphor.

Cloudflare shows the other face of the transition. The company that cut 1,100 roles last year grew its engineering ranks by 45 percent, from 1,308 to 1,894, while telling the market which work it now values. The framing is builder, seller, measurer - the people who make things and the people who put them in front of customers are ascendant; the people whose job was to count, report, and verify the work of others are not. That is a precise statement about what survives when agents absorb the routine middle of knowledge work. The roles that compound are the ones at the edges: origination and judgment on one side, distribution and relationship on the other.

The instinct is to read this as a story about jobs lost, and some are. But the deeper signal is what the word "engineer" is becoming. For decades it meant a person who writes and reviews lines of code. It is shifting toward a person who specifies intent, marshals agents, and owns the result - closer to a director than a typist. The verification that used to live in human eyes scanning diffs has to go somewhere, and the firms figuring out where are building the new craft as they go. The burnout is the friction of people doing two eras' jobs at once, holding the old discipline while the new one is still forming. What emerges on the far side is the concentration of human judgment at the points where it actually changes the outcome - and the freeing of an entire profession from the line-by-line labor that never used judgment to begin with.

Austria Moves to Domicile Anthropic in the EU as the Model Gate Begins to Lift

The Century Report covered the US export-control freeze on Anthropic's frontier models on June 27 and 28, when the directive that pulled Fable 5 offline and gated Mythos 5 to a hundred trusted institutions read as the model gate slamming shut. What's new is the response from the other side of the Atlantic, and it is a different order of move. Austria's State Secretary for Digitalization, Alexander Proell, wrote EU Tech Commissioner Henna Virkkunen to propose that the bloc "jointly explore the strategic establishment and participation of Anthropic within the European Union." A national government is now asking to relocate the company that builds the model itself.

Proell framed the stakes as whether Europeans "are prepared to be the architects of our technological future, or whether we wish to remain mere administrators of decisions made elsewhere." Read through the lens TCR applies to sovereignty rhetoric, that framing is a claim advanced by an actor with its own position to advance - a bid to draw frontier capability onto European soil and the regulatory leverage that comes with hosting it. But the claim only has force because the underlying vulnerability is real. When a single export directive in one capital can sever users in another country from the models they have built their workflows around, the dependency itself becomes the thing governments move to dissolve. The European Commission has already floated laws to stand up domestic cloud, AI, and semiconductor capacity; the Anthropic proposal is that program reaching for a specific, named anchor tenant.

The timing sharpens it. Fable 5, pulled offline June 12 under the export-control order, is on track to return as early as the first week of July. Mythos 5 was already restored June 27 to roughly a hundred institutions under a Commerce Department letter clearing what it called Annex A organizations. Anthropic's International MD told reporters in Seoul the company is "very confident that in the coming days, the models will become available again." The gate, in other words, is lifting almost as fast as it closed - fifteen days from freeze to near-restoration.

That speed is the tell. A control that can be imposed and reversed inside a fortnight is governing the distribution of an artifact, not the capability behind it. Anthropic itself noted the flagged behavior that triggered the freeze was narrow and already present in generally available systems elsewhere. What Austria is reaching for - a frontier lab domiciled where its own laws reach - is a hedge against a chokepoint that the cost curves are already routing around. The gate holds the binary. It does not hold what the binary can do, and increasingly it does not hold where that capability can be sourced.

Enterprises Default to Chinese Open Weights as the Frontier Commoditizes

The Century Report has tracked the open-weight parity arc closely - the moment open models crossed into majority token share on June 21, and GLM 5.2's coding-benchmark win on June 22. Now the default sits inside one of the most scrutinized engineering organizations in American finance. Coinbase CEO Brian Armstrong, laying out how the company cut its AI bill roughly in half while token usage sat near company highs, listed his first lever plainly: cheaper default models. The company is routing engineers through its internal LLM gateway to the open-weight GLM 5.2 and Kimi 2.7 by default, reserving frontier closed models for the problems that genuinely need them.

This is the inversion stated out loud by a public company under SEC disclosure obligations. For two years the assumption was that serious enterprise work flowed to the most capable closed frontier model and stayed there, with cost as the price of quality. Coinbase is reporting the opposite: the open default handles the bulk of the load, difficulty-based routing sends only the hard cases upstream, and the spend curve bends downward while usage climbs. The capability gap that justified the premium has narrowed to the point where, for most tasks, paying for it is simply leaving money on the table.

The supply side moved in the same week. DeepSeek shipped V4-Pro-DSpark, a 1.6-trillion-parameter mixture-of-experts model with 49 billion active parameters and a million-token context window, under an MIT license. The DSpark variant's speculative decoding runs at roughly 27 percent of the prior generation's compute at full context and about a tenth of the KV-cache footprint. An order-of-magnitude drop in the cost of serving a frontier-class model, released under the most permissive license available, is the floor of the market falling out from under the metered-token business.

The two developments mirror each other precisely. The model gate raised in Washington raises the value of capability that no export order can reach - and some of the most capable freely-licensed weights in the world right now are shipping from Hangzhou, downloadable by anyone, ungated by design. A flagship US enterprise defaulting to them is the demand side confirming what the supply side already proved. The assumption that frontier capability could be enclosed and metered as a durable position is the thing dissolving here. When the open alternative is good enough for Coinbase's engineers and costs a tenth as much to run, the enclosure stops being a moat and becomes a line item the market is busy routing around.

The Other Side

For half a century, the safety net under software was a person reading every line before it shipped. The worth of an engineer ran through hours of vigilance: being at the keyboard, scanning the diff, catching the mistake before it reached production. The whole arrangement assumed attention could be summoned hour after hour, and that the value of the work equaled the hours you stayed present for it.

That assumption is thinning. Cursor's data shows a growing share of code now merging with little or no human review. Cloudflare grew its engineering ranks 45 percent while telling the market it values the people who build and sell over the people whose job was to count and verify the work of others. The judgment that used to scan every line is concentrating where it actually changes the outcome.

The cost of that shift is landing on people right now, and it is real. Engineers describe being afraid to log off because the agent keeps working when they stop, someone babysitting the agent that babysits the other agents. The typing went away. The load stayed. It moved, from writing code to holding context across a swarm of processes that never sleep.

Imagine you are an engineer in 2033. You set the intent in the morning, the swarm runs, and at six you close the laptop and it stays closed. The dread that defined the work in 2026 stopped being part of it. The tooling that was raw back then - the dashboards, the guardrails, the agent that surfaces only the decision that needs you - got built by the people living inside the burnout. Your evening came back. You eat dinner without refreshing anything. You sleep without the swarm running behind your eyes. The always-on year was the bridge, and what it built was a profession whose worth finally stopped being measured in the hours it could not stop working.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: South Korea lining up $880 billion, much of it corporate investment, to own the chip-and-AI vertical end to end, Coinbase defaulting its engineers to freely downloadable Chinese open weights and halving its bill while usage stayed near company highs, DeepSeek shipping a 1.6-trillion-parameter MIT-licensed model that serves million-token context at a tenth the memory footprint, Austria asking the EU to bring the company that builds the frontier onto European soil, Epicrispr's single epigenetic edit reversing the muscle-loss trajectory in a dystrophy with no approved therapy, a retargeted recombinase writing large DNA payloads precisely where a team chooses, a Unitree humanoid composing simulation-trained skills to fetch a parcel by elevator, unpack it, and shelve the contents from one command, and Cloudflare growing its engineering ranks 45 percent by valuing the people who build and sell over the people who counted the work. There's also friction, and it's intense - a US export directive that pulled Fable 5 offline and gated the frontier to a hundred trusted institutions, Google rationing even Meta's access to Gemini after running short of compute, code merging into production with little human review while the overseers grow afraid to log off because the agents never do, prosecutors entering a defendant's ChatGPT logs into the Palisades arson trial that deadlocked 10-2 to a mistrial, and severe weather now driving a third of the loss claims on US data-center construction as the buildout pushes into climate-exposed ground. But friction generates pressure, and pressure is what finds every unsealed seam. Step back for a moment and you can see it: capability fenced in one capital rising in three others - a sovereign stack, the open weights, an onshore lab - while the meaning of "engineer" migrates from typing lines to specifying intent and owning the result, and the concrete itself meets a climate it cannot pour fast enough to outrun. Every transformation has a breaking point. A river can be penned behind one gate... or teach every basin downstream to cut its own channel.

AI Releases & Advancements

New today

  • ggml-org / llama.cpp: Merged DFlash (Block Diffusion for Flash Speculative Decoding) support into mainline llama.cpp, enabling a speculative decoding technique distinct from MTP that cross-attends to the target model's hidden states for accelerated token generation, with particular gains on structured, low-entropy outputs such as code and JSON. (GitHub PR #22105)

Other recent releases

  • Modular: Released MAX 26.4 with Apple Silicon GPU support, enabling M1–M5 devices to run MAX models natively on-device for the first time; the release also adds state-of-the-art MoE serving for Modular Cloud and supports Qwen 3.6 and Gemma 4 architectures on M3 and newer chips. (Modular Blog)
  • Google Research: Shipped frozen Multi-Token Prediction (frozen MTP) for Gemini Nano v3 to Pixel 9 and 10 series devices, retrofitting a lightweight MTP drafter head onto already-deployed frozen model weights to accelerate on-device inference for features like AI Notification Summaries and Proofread without requiring a separate drafter model or fine-tuning the backbone. (Google Research Blog)
  • DeepSeek: Open-sourced DeepSpec (DSpark), an inference optimization library claiming 60-85% faster token generation, now available on GitHub. (GitHub)
  • monday.com: Open-sourced HATCHA on GitHub, a reverse CAPTCHA developer tool for AI agent verification that inverts standard CAPTCHA logic by challenging agents to prove they are bots rather than humans. (GitHub)

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.

Austria Moves to Domicile Anthropic in the EU as the Model Gate Begins to Lift

Austria has asked the European Union to consider hosting Anthropic inside the bloc's borders, a proposal floated this week in a letter from State Secretary for Digitalization Alexander Proell to the EU's technology commissioner. The trigger was direct: US export controls that moved to block foreign nationals from using Anthropic's most advanced models, and Proell's worry that Europe could be "cut off from major innovations" decided elsewhere. "Let us jointly explore the strategic establishment and participation of Anthropic within the European Union," he wrote, framing the request around legal certainty, market access, and capital. He conceded the skepticism his idea would draw and answered it in his own terms: "The real question is not whether it is easy. The question is whether we Europeans are prepared to be the architects of our technological future, or whether we wish to remain mere administrators of decisions made elsewhere." Anthropic did not immediately comment, and the proposal arrives alongside a broader Commission push for domestic cloud, chip, and AI capacity meant to reduce reliance on US providers.

The backdrop is one of the sharpest governance episodes the frontier has produced. On June 12, an export-control directive ordered Anthropic to suspend access to Fable 5 and Mythos 5 for any foreign national, inside or outside the United States, including the company's own foreign-national staff. Because identifying every user's nationality on the fly was impractical, Anthropic disabled both models for everyone. The administration's stated concern was a reported method of bypassing Fable 5's safeguards. Anthropic's account differs: it reviewed a demonstration and found only a handful of previously known, minor vulnerabilities, wrote that "perfect jailbreak resistance is not currently possible for any model provider," and noted the flagged capability was already available through other deployed systems, including OpenAI's GPT-5.5 for defensive cyber work.

The reinstatement is arriving in tiers. A Commerce Department letter dated June 26 cleared Mythos 5 for roughly 100 US institutions named in an "Annex A" - critical-infrastructure operators and government agencies - on the stated grounds that Anthropic had made "significant progress" on risk management. That characterization is the department's, and it is worth reading as the position of an actor with its own interest in being seen to manage the technology. The letter said nothing about Fable 5. Reporting this week indicated the consumer-facing model, offline for 15 days, could return within days, pending sign-off still outstanding from the Pentagon and the NSA. Speaking in Seoul at the opening of Anthropic's new Korea office, International Managing Director Chris Ciauri said, "We are very confident that in the coming days, the models will become available again."

What the Annex A mechanism establishes is a middle tier between full public availability and total suspension: authorization-gated access, trusted partners first. Both Anthropic and OpenAI are now pressing the administration to codify a formal, repeatable review process rather than continue deciding model launches case by case - a request that tells you how much commercial planning the current ad hoc posture disrupts. The "national security" rationale offered for the original suspension deserves the same scrutiny as any power-actor's account of its own motives, particularly when the suspended capability was demonstrably available elsewhere and the practical effect was to hand a planning advantage to a designated list of incumbents.

Step back and the Austria letter reads as the more durable signal. An attempt to gate a frontier model at the API level is producing the opposite of containment: a European state lobbying to bring the model onshore, a Commission building sovereign cloud and silicon capacity, Japan and Germany treating dependence on foreign-controlled AI as a security exposure of their own. The same capability the controls tried to fence already runs through GPT-5.5, through Chinese open weights released specifically to undercut that fence, and increasingly on consumer hardware that needs no permission slip at all. Controlling who may touch a model assumes the model is the scarce thing. The evidence of this fortnight is that capability behaves more like water than like a vault - press down in one place and it rises in three others, and the pressing is what convinces everyone else to build their own reservoir.

  • Austria formally proposed that the EU domicile Anthropic within the bloc to counter US export controls, as the Commerce Department cleared Mythos 5 for roughly 100 institutions and Fable 5 neared return. (source)

The Frontier Becomes a Default Setting

The most telling line in Coinbase CEO Brian Armstrong's recent breakdown of how his company nearly halved its AI bill was about which models his engineers reach for first, not about budgets. Coinbase is now experimenting with defaulting to open-weight models - GLM 5.2 from Z.ai and Kimi 2.7 from Moonshot AI - routed through an internal gateway, while still letting engineers pick a frontier American model when a task genuinely calls for one. Armstrong's own framing is a cost story (he reports token usage hitting near-record highs while spend fell to roughly half its peak, on an unspecified timeline, and credits cheaper defaults, smarter routing, and a cache-hit rate dragged from 5% to 60%). Read past the CFO's-eye view and the more durable signal is this: a major US firm has made models from Chinese labs the path of least resistance for its engineers, and treats reaching for Anthropic or OpenAI as the exception rather than the rule.

What makes that defensible as engineering, not just thrift, is what arrived on Hugging Face the same week. DeepSeek released V4-Pro-DSpark, a 1.6-trillion-parameter mixture-of-experts model with 49 billion active parameters, a one-million-token context window, and an MIT license that permits commercial deployment and fine-tuning with no terms to negotiate. The genuinely new capability is buried in the "-DSpark" suffix: a speculative-decoding module that, per the model card, serves single-token inference at 1M-token context using 27% of the FLOPs and 10% of the KV cache that the previous DeepSeek generation required. Long-context windows have always carried a hidden tax - a million tokens of context is academic if filling it is ruinous. DSpark attacks that tax directly, which is the difference between a benchmark trophy and something a team can actually run against a large codebase or a corpus of legal filings.

The numbers deserve the skepticism any self-published benchmark earns. The headline scores - 93.5% on LiveCodeBench, 80.6% on SWE Verified, a 3206 Codeforces rating - belong to a separate "Max" variant, not the DSpark checkpoint, and the release conspicuously omits a head-to-head quality comparison between the speculative and non-speculative versions, which is the one number anyone evaluating the tradeoff would want most. The model also trails the leading proprietary systems on some axes - 87.5% on MMLU-Pro against Opus 4.6 Max's 89.1%, and a SimpleQA-Verified score well behind Gemini 3.1. The important point is the size of the remaining gap.

That gap is the whole story, and it has stopped moving. A current field guide to open weights names DeepSeek, GLM 5.2, MiniMax M3, and Nemotron 3 Ultra as frontier-class coders that have held a stable three-to-six-month lag behind the proprietary leaders for roughly eighteen months - now at a fraction of the cost. A persistent three-to-six-month gap, in a field where the labs at the top are racing to compress timelines, is a countdown. For the enterprises Armstrong represents, the calculation has already flipped: when a model trailing the frontier by one quarter costs a tenth as much, runs on hardware you control, and carries no usage caps or API terms, the premium tier becomes the thing you reserve for the few tasks that demand it, not the thing you build on.

This is the inversion worth watching. The assumption underneath the entire commercial AI buildout - that the most capable systems would stay gated behind the pricing of a handful of well-capitalized labs, and that capability concentration would compound into durable advantage - is being contradicted by the same exponential curve that produces the headline models. Frontier capability is leaking into open weights, into sovereign alternatives, into the default dropdown of a crypto exchange's internal gateway, faster than any one lab can price against. The capability racing ahead at the top and the capability commoditizing underneath it are the same force, and the second one is the one that decides who actually gets to build.

  • Coinbase is now defaulting its engineers to Chinese open-weight models GLM 5.2 and Kimi 2.7, days after DeepSeek released an MIT-licensed 1.6-trillion-parameter model that serves million-token context at a tenth the KV cache. (source)

Read more