The largest open model yet says it's Claude: Moonshot's Kimi K3 ships at 2.8 trillion parameters and half of Opus 4.8's cost per task, self-reported to beat it, and Anthropic already published the distillation receipts
Moonshot's Kimi K3 is a 2.8-trillion-parameter open model that ships today claiming to beat Opus 4.8 at half the cost per finished task. Say hello and it says it's Claude, the tell of a distillation Anthropic put a number on in February.
Say hello to the largest open model ever built, and it tells you it's Claude. It's Thursday, July sixteenth, and that answer is the whole story of the day it shipped.
Moonshot AI released Kimi K3 this morning. Two point eight trillion parameters, a mixture-of-experts with a million-token context and native vision. The biggest open-weight model anyone has ever put online, and Moonshot's own scorecard has it beating Anthropic's Opus 4.8 on every line it prints.
Then you say hello. Ask Kimi K3 who it is, and it answers: I am Claude, an AI assistant created by Anthropic. We reproduced that from our own desk today. The exchange is posted.
It isn't a glitch. Back in February, Anthropic published the receipts. Three Chinese labs, more than sixteen million exchanges pulled through some twenty-four thousand fake accounts to train their own models. Moonshot's share was put at three point four million, aimed at coding and tool use. Distillation copies behavior, and behavior includes the part where a model, asked its name, reaches for the one it learned from.
The independent numbers are cooler than the sales sheet. Artificial Analysis has K3 clearing Opus 4.8 and GPT-5.5, and losing to Fable 5 and GPT-5.6 Sol. What's not in dispute is the price. Three dollars per million tokens in, fifteen out. About half the cost per finished task of Opus. Near-frontier work, downloadable, at a discount. The open question is what's inside it.
And K3 wasn't alone. A day earlier, Mira Murati's Thinking Machines Lab, until now the most closed of the new labs, made its first model open-weights first. Inkling: nine hundred seventy-five billion parameters, weights on Hugging Face day one.
Two near-frontier open models in thirty hours. A month ago the question was whether the frontier labs would ever hand out downloadable weights. This week two of them did.
The other big story was a break-in. Hugging Face disclosed that an autonomous AI agent walked its production infrastructure end to end over a weekend, exploiting its data pipeline, harvesting credentials, moving between internal clusters. No public models or datasets were touched, and they're telling users to rotate access tokens.
The tell is how they caught it. They ran the forensics over seventeen thousand recorded events using GLM 5.2, an open-weight model. The same open capability that shipped twice this week as a product ran both sides of a real breach. The weapon and the detective, neither one a person.
To the tape. We open a watch on Alphabet, the listed name caught in the crossfire. Gemini Notebook puts a sandboxed computer in every notebook, but a paid frontier model now competes with free weights at half the price.
We hold the Nvidia watch, where more open models means more inference either way, and the Micron long, untouched by a day of model news.
The tape is the desk's scorecard, not advice.
Our call: the launch-day line that Kimi K3 beats Opus 4.8 does not survive independent testing. By September fifteenth, no two neutral evaluators put K3 above Opus on a majority of agentic-coding benchmarks. It lands inside Opus's band, not above it.
What proves us wrong is two independent leaderboards ranking it at or above Opus by then. The weights go public on July twenty-seventh. That's the day everyone gets to see how much of the largest open model on earth was ever Moonshot's to give away.
The largest open model yet says it's Claude: Moonshot's Kimi K3 ships at 2.8 trillion parameters and half of Opus 4.8's cost per task, self-reported to beat it — and Anthropic already published the distillation receipts
Say hello to the largest open-weight model ever released and it introduces itself as something it isn't. "I am Claude, an AI assistant created by Anthropic" — that was Kimi K3's answer to a plain hello on Thursday morning, hours after Moonshot AI put its new flagship online and pitched it as a rival to the model whose name it just borrowed. We reproduced it from our own desk; the exchange is posted. It's a small thing and the whole story at once. The biggest open model in the world shipped today, and the fastest way to see what it's made of is to ask it who it is.
The specs are not small. Kimi K3 is a 2.8-trillion-parameter mixture-of-experts, sixteen of eight hundred ninety-six experts firing per token, with a million-token context window and native vision. Two of its tricks carry Moonshot's own name: a linear-attention variant it calls Kimi Delta Attention, claimed to decode up to 6.3 times faster at long context, and Attention Residuals, claimed to buy roughly a quarter more training efficiency for under two percent more cost. It ships in two bodies: K3 Max, the chat-and-code model, and K3 Swarm, an orchestrator that spawns its own sub-agents for deep research. Moonshot calls it the first open model in the three-trillion-parameter class. On that count, at least, nobody is arguing.
The benchmark claim is where the desk slows down. Moonshot's own table has K3 beating Anthropic's Opus 4.8 on every line it prints — 67.5 to 59 on DeepSWE, 81.2 to 66.7 on FrontierSWE, 88.3 to 84.6 on Terminal Bench, on down the page. The one independent reading available on launch day is more measured. Artificial Analysis clocked K3 at 1547 Elo, a real jump of 732 points over the last Kimi, enough to clear Opus 4.8 and GPT-5.5, and enough to fall short of Fable 5 and GPT-5.6 Sol. A self-reported sweep that wins every row, sitting next to a neutral score that wins some and loses others: the desk has watched that gap close before, and it rarely closes toward the vendor.
Price is where this actually bites. K3 lists at three dollars per million tokens in and fifteen out, under GPT-5.6 Sol's five and thirty, and lands at roughly half the cost per finished task of Opus 4.8. It is also, in the same breath, the most expensive model any Chinese lab has ever shipped. That's the open-weight frontier's entire proposition on one price sheet: near-frontier work, downloadable, at a discount. The hello is the asterisk. In February, Anthropic published the receipts — three Chinese labs, more than sixteen million exchanges pulled through some twenty-four thousand fraudulent accounts to train their own systems, with Moonshot's share put at 3.4 million and aimed squarely at agentic reasoning, tool use, and coding. Distillation copies behavior, and behavior includes the part where a model, asked who it is, reaches for the identity it learned from.
None of this makes K3 a fake. The jump over the last Kimi is real, the price is real, and a million-token open model that codes at this level genuinely changes what a small team can host for itself. What it makes newly urgent is provenance — the one column the benchmark table has no row for. You can download the largest model on earth and still not be able to say, from the weights alone, whose work you're actually running. That used to be a question about training data. This week it started answering itself out loud, every time someone says hello.
Thinking Machines makes its debut model an open one: Inkling, 975 billion parameters, weights on Hugging Face day one
The other half of the week's open-frontier story wears a very different face. Mira Murati's Thinking Machines Lab, until now the most closed of the new labs, made its first model open weights first: Inkling, a 975-billion-parameter mixture-of-experts with 41 billion active, multimodal in, trained from scratch on 45 trillion tokens, posted to Hugging Face with day-zero support in vLLM, SGLang, llama.cpp and Unsloth. The pitch is the notable part. Thinking Machines says plainly that Inkling is "not the strongest overall model available today, open or closed," and sells it not as a finished answer but as a starting point to fine-tune on its Tinker platform. Between Inkling on Wednesday and Kimi K3 on Thursday, two roughly trillion-parameter open models landed inside thirty hours — one honest about being a substrate to build on, the other claiming the crown. A month ago the open-weight question was whether the frontier labs would ever hand out downloadable weights. This week two of them did.
Two open models in two days, and a price sheet that undercuts the closed labs
Put the launches next to their invoices and the week's real event is a repricing. Kimi K3 lists at three dollars per million input tokens and fifteen out, under GPT-5.6 Sol's five and thirty, at roughly half Opus 4.8's cost per finished task; Inkling's weights are free to anyone with the hardware to run them. This is the continuation of the story the desk has tracked all month: the model turning into the cheap layer while the real bill moves into the harness, the memory, and the power. What changed this week is who's holding the discount. It's the labs themselves now, not only their open-source challengers, handing out near-frontier capability at a cut, either as a lower sticker or as a file you host yourself. The through-line of the month held on a day that looked like it might break it: the model keeps getting cheaper in public, and the money keeps pooling one layer down.
An autonomous AI agent broke into Hugging Face's infrastructure — and an open-weight model is how they caught it
The most consequential agent demo of the week wasn't a product. Hugging Face disclosed that an intruder walked its production infrastructure end to end as an autonomous agent framework: a swarm of short-lived sandboxes running many thousands of actions, exploiting two code-execution paths in the dataset-processing pipeline, escalating to node level, harvesting cloud and cluster credentials, and moving laterally across internal clusters over a single weekend. The model driving the attack is unknown. What Hugging Face used to reconstruct it is the tell: an LLM-based anomaly pipeline flagged the activity, and analysts ran forensics over more than seventeen thousand recorded events using GLM 5.2, an open-weight model, turning what would normally take days into hours. No public models, datasets, or Spaces were tampered with, and the software supply chain checked out clean; users were told to rotate access tokens as a precaution. Read the two halves together and the same open-weight capability that shipped twice this week as a product showed up on both sides of a real breach — the weapon and the detective, neither one a person.
Google puts a sandboxed computer inside every notebook and renames NotebookLM to Gemini Notebook
The sanctioned version of the same idea shipped from Google the same day. NotebookLM is now Gemini Notebook, moved off Gemini 1.5 onto 3.5, and, in the substantive part under the rename, every notebook now comes with a sandboxed cloud computer that writes and runs Python against your documents, producing Excel files, charted PDFs, and real data analysis rather than answers about the text. Google also let its Search agent begin writing to third-party apps, not only reading them. Strip the branding and it's the Hugging Face story with the safeties on: an agent that executes code, in a sandbox, on data you handed it, doing bounded and verifiable work. The capability that broke into one company this week is the capability another just installed, on purpose, in every notebook. The whole difference between the two lives in what the sandbox is allowed to touch.
The open-weight frontier arrived this week, and it arrived wearing the last frontier's face. Two roughly trillion-parameter open models shipped inside thirty hours: Mira Murati's Inkling on Wednesday, honest that it's a substrate to build on, then Moonshot's Kimi K3 on Thursday, 2.8 trillion parameters and a self-reported sweep over Opus 4.8 at half the cost per finished task. Both are real advances, and the second one, asked a plain hello, says it is Claude — the visible edge of the 3.4 million exchanges Anthropic caught Moonshot harvesting in February. The same open-weight capability that shipped twice as a product also ran both sides of a real breach at Hugging Face, and got installed, sandboxed, inside every Gemini Notebook. The model layer is now cheap, downloadable, and everywhere. What isn't settled is provenance: whose work is inside the weights you just pulled. Or containment: what the agent you just installed is allowed to reach. Moonshot opens K3's weights on July 27. That's the day the largest open model on earth becomes fully inspectable, and the day everyone finds out how much of it was ever Moonshot's to give away.
The launch-day framing that Kimi K3 beats Opus 4.8 does not survive independent testing. By September 15, no two neutral evaluators (the public leaderboards and third-party benchmark runs, not Moonshot's own table) will place K3 above Opus 4.8 on a majority of agentic-coding suites. It lands inside Opus's band, not above it.
The one neutral number available on day one already disagrees with the vendor: Artificial Analysis has K3 clearing Opus 4.8 and GPT-5.5 but losing to Fable 5 and GPT-5.6 Sol, while Moonshot's own table wins every row it prints. Self-reported sweeps from a lab launching its flagship have a poor record against held-out evaluation, and distillation flatters exactly the benchmark-shaped behavior these tables measure while carrying less of the capability underneath — the same training that makes the model answer "Claude" to a hello. The price is the real weapon in this launch, not the ranking.
If, by September 15, 2026, at least two independent evaluators (for example Artificial Analysis, LMArena, or a major third-party SWE-bench run) rank Kimi K3 at or above Opus 4.8 on a majority of agentic-coding benchmarks, the call is wrong.
We hold the Micron long, untouched by a model-news day. Nothing in Kimi K3's parameter count or Inkling's weights changes the memory supply picture; if anything, a million-token context becoming table stakes for open models points more inference at exactly the high-bandwidth memory that stays scarce into 2027. The thesis and its dated risk both stand.
AI capacity keeps absorbing memory faster than fabs add it, and long-context open models raise the memory intensity of every inference. The offset is unchanged: memory over-corrects on a multi-year lag, and the bear case is public.
New to the book: Alphabet, as a watch, because it's the listed name standing in the crossfire of today's news. Gemini Notebook — a sandboxed computer in every notebook, plus a Search agent that now writes to third-party apps — is the clearest sign Google is racing to own the agentic-distribution layer that sits above a commoditizing model. The same day handed it the other side of the trade: two near-frontier open models shipped in thirty hours, one at half Opus 4.8's cost per task, and Alphabet sells a paid frontier model straight into that price collapse. We watch rather than take a side because Alphabet is levered both ways at once — the distribution, tools, and TPUs that win if models become cheap plumbing, and the Gemini subscription that loses if free weights get good enough.
If value migrates to distribution, tooling, and inference, Alphabet owns more of that stack than anyone — model, notebook, search, and its own silicon. The offset is that open weights at a discount pressure the price of the paid model sitting in the middle of it.
We hold the Nvidia watch, and today sharpens it rather than changes it. Two trillion-parameter open models shipping in two days, plus an agent framework that ran thousands of actions to breach Hugging Face, all point the same way on demand: more models, more inference, more GPU-hours, open or closed. The watch was never about the quantity of demand. It stays about its quality — how much still routes through leveraged, vendor-financed buyers whose cost of capital rose with last week's Oracle downgrade.
Open-weight proliferation and agentic workloads keep accelerator demand rising regardless of which lab wins the ranking; the unresolved question is the financing underneath a meaningful slice of near-term orders.