Big day in open-source AI. DeepSeek dropped V4, a 1.6 trillion parameter mixture-of-experts model with a 1 million token context window, open weights under Apache 2.0, and pricing that undercuts almost every frontier lab. Bitcoin is hovering around 77,000 ahead of the Fed decision, with traders eyeing a push back toward 80,000 while Coinbase premium just flipped negative for the first time in three weeks. Block went big at Bitcoin 2026 in Las Vegas, launching a new Bitkey hardware wallet with a touchscreen, auto-converting Cash App payments to bitcoin, and publishing on-chain proof of reserves covering over 28,000 BTC. And in London, the FCA ran its first physical crackdown on illegal peer-to-peer crypto trading, hitting eight premises with cease-and-desist notices. Let's get into it.
DeepSeek just shipped V4, and the numbers are genuinely hard to ignore. The flagship, V4 Pro, is a mixture-of-experts model with 1.6 trillion total parameters and 49 billion active per token. There's also V4 Flash at 284 billion total, 13 billion active. Both ship with a 1 million token context window, both are Apache 2.0, weights are on Hugging Face, and the API is live with OpenAI and Anthropic protocol compatibility.
The efficiency story is the real headline. DeepSeek claims V4 uses about 27% of V3.2's inference FLOPs per token, and just 10% of the KV cache at full 1 million context. They got there with a hybrid attention scheme, manifold-constrained hyper-connections, and the Muon optimizer. Pre-training ran on 32 trillion tokens with FP4 and FP8 mixed precision.
On benchmarks, V4 Pro Thinking is sitting at number 3 among open models on the Arena leaderboard, beats GPT-5.4 on some coding tasks, and is closing the gap with Gemini 3.1 Pro and GPT-5.5 on reasoning. Knowledge benchmarks still trail frontier models by maybe three to six months, and it's text-only, no audio or video yet. But pricing — Flash is around 14 cents per million input tokens, 28 cents output. That's a serving cost that makes a 1 million token context actually usable in production.
Meta also showed up at LlamaCon with Llama 4 Scout and Llama 4 Maverick, both MoE, Scout designed to run on a single GPU. And Satya Nadella casually mentioned that 30% of Microsoft's code is now AI-generated. The competitive dynamic is clear: open weights are no longer the budget option. They're catching the frontier, and developers are responding by building multi-model routers instead of betting on one lab. If you're still hardcoding a single model into your stack, you're accruing technical debt by the week.
AI agents are graduating from demos into actual workflow infrastructure this week, and three releases stand out.
First, Anthropic's Claude Dispatch. It turns your phone into a remote control for your desktop — Claude can read files, pull up documents, summarize what's on screen, and prep you for a meeting while you're walking back from lunch. Early reviews call it genuinely useful for context-heavy work, less of a chatbot and more of an integrated assistant that actually touches your apps.
Second, Google reimagined Chrome at Cloud Next 2026. The new Auto Browse feature, powered by Gemini 3, completes multi-step tasks autonomously — scheduling, form filling, expense reports — with confirmation gates for sensitive actions. There's also Chrome Skills, reusable AI workflows triggered from the address bar, and a Gemini side panel scoped per tab. Enterprise pricing lands at 6 dollars per user per month with real-time data loss prevention. Google is leveraging 3.8 billion Chrome users to make the browser itself the agentic platform, and that's a serious threat to standalone agent startups.
Third, OpenAI released Symphony, an open-source spec for orchestrating coding agents around issue trackers like Linear. Instead of supervising individual coding sessions, you let agents work tickets in parallel, with a state machine handling dependencies and CI checks. OpenAI says internal teams saw a notable jump in landed pull requests because agents handle speculative work cheaply while humans focus on judgment calls.
And to ground all this, TestMu launched Kane CLI, a terminal-native browser automation tool that acts as the verification layer — making sure agent-generated flows actually run end-to-end in a real Chrome session, with human-in-the-loop for things like CAPTCHAs.
The pattern across all four: agents are no longer just generating output. They're executing, verifying, and coordinating across systems. The bottleneck is shifting from model capability to orchestration and trust.
Block came to Bitcoin 2026 in Las Vegas with a complete overhaul of its bitcoin stack, and it's worth taking seriously because Block touches more bitcoin users than almost anyone outside the exchanges.
The centerpiece is the new Bitkey hardware wallet. It's got a built-in touchscreen for on-device verification, keeps the 2-of-3 multisig model, no seed phrases, and adds inheritance features. Preorders are open. The pitch is that self-custody shouldn't require memorizing 24 words on a metal plate, and the multisig design means losing one key isn't catastrophic.
On Cash App, eligible users can now auto-convert peer-to-peer payments into bitcoin, get 5% Bitcoin Back at Square merchants, and withdraw up to 10,000 dollars per day or 25,000 per week. Fees and spreads disappear on purchases over 2,000 dollars. Block also demoed NFC tap-to-pay using the Lightning Network at over 800,000 Square merchants — a new merchant gets auto-enrolled roughly every eight seconds — with zero processing fees through 2026. That's a direct shot at Apple Pay's economics.
The transparency play is proof of reserves. Block's Q1 2026 disclosure shows 28,355 BTC, roughly 2.2 billion dollars, with about 19,357 BTC attributable to customers and 8,998 BTC in the corporate treasury, all verifiable on-chain via cryptographic signatures. Michael Saylor has publicly argued that proof of reserves creates security risks by exposing wallet structure. Block clearly disagrees.
On the policy side, Representative Nick Begich used the same stage to push self-custody as a constitutional and property-rights issue, and rebranded his bitcoin reserve bill as the American Reserves Modernization Act, still aiming for up to 1 million BTC over five years. Meanwhile, the Czech central bank, which actually bought 1 million dollars in bitcoin last October to test it, just declared the asset too risky for reserves. Two very different conclusions from people closer to the question than most.
The UK is moving fast on crypto regulation, and the implications go well beyond British borders.
First, enforcement. The FCA ran its first coordinated crackdown on illegal peer-to-peer crypto trading, with HMRC and the South West Regional Organised Crime Unit hitting eight London premises with cease-and-desist notices. The FCA is blunt: there are zero registered peer-to-peer crypto traders in the UK, so anyone running that business is operating illegally under the 2017 money laundering regulations.
More consequential is Consultation Paper CP26/13, published April 15. It clarifies the perimeter of the new Cryptoassets Regulations 2026. Seven activities will fall under the FCA's remit: issuing qualifying stablecoins, custody, operating trading platforms, dealing as principal, dealing as agent, arranging deals, and arranging staking. The application window for licenses opens September 30, 2026 and runs to February 28, 2027. The full regime takes effect October 25, 2027.
The DeFi implications are sharp. The FCA reads "arranging" broadly. Any front-end, web3 interface, or non-custodial wallet that helps a UK user place an order, discover prices, or source liquidity could be considered to be arranging deals — and could need authorization. Licensed firms would have to subsidiarise in the UK, hold regulatory capital, and may be limited to tokens trading on UK venues. The realistic alternative for many DeFi projects is geofencing UK users out entirely.
Stablecoin issuance is captured regardless of where the issuer is domiciled, which means anyone serving UK users with a stablecoin needs to pay attention now, not in 2027. Consultation feedback closes June 3.
Elsewhere, Canada is proposing an outright ban on crypto ATMs, citing scam losses. MoonPay bought Israeli security firm Sodot for 100 million dollars in stock to build out an institutional custody arm under former CFTC Acting Chair Caroline Pham. And in the US, the CLARITY Act has stalled in Senate Banking, with Galaxy Research putting the odds of enactment this year well below 50%. The regulatory map is hardening in some places and fragmenting in others, and where you're domiciled is starting to matter as much as what you're building.
One thing to watch: when the cheapest open-source model can read your entire codebase in a single prompt, the moat isn't the model anymore — it's whatever you build on top of it that nobody else can copy.