Two Western open models arrive with their weights still to come

AI² — AI Investor Sydney · Friday 11 December, The Mint. One widely cited forecast puts cumulative AI and data-centre capital spending at about US$11 trillion by 2029. One day, six evidence-led briefings on where that capital is creating and destroying value — no panels, no pitches. Every ticket includes a year of NOOPS Pro. A$995 until 30 October → Book your seat

Two Western labs put large open-weight models into preview on consecutive days, both with the weights still to come, while DeepSeek, the lab one of them was reported to be answering, weighs doubling its round. Underneath, a supplier argues the buildout is limited by supply rather than demand, and agents get a commerce standard and a named attack class in the same week.

Western open weights, two days running

Reflection has released Beam, its first open-weight model, a mixture-of-experts with 501 billion parameters and 23 billion active, in early access while red-teaming finishes; the weights, technical report and model card are due this month under Apache 2.0. Reflection concedes that Kimi K3 is ahead on raw capability and pitches efficiency instead, claiming reasoning comparable to GLM-5.2 on three to four times less inference compute, by its own approximation that leaves out prefill, attention and serving overhead; its reinforcement-learning run used 10,500 Nvidia GB300s for four weeks. A day later Mistral opened a public preview of Mistral Large 4, 1 trillion parameters with 49 billion active, which Artificial Analysis scores at 38 on its Intelligence Index against a median of 26 for comparable models; the weights are due by the end of October, under a licence Mistral has not stated. The pitch is sovereignty and cyber defence: Mistral reports 82 per cent on a cyber test on which, it says, Claude Opus 5.5 and GPT-6 Astra score near zero because they refuse, which is the vendor's framing, and the model used 200 million tokens to run the index against a median of 81 million, so cost per task rather than list price is the number to compare. Further down the stack, openTPU, a GitHub project of about 160 stars, puts an accelerator design, instruction set, bit-exact simulator, compiler and host software in one repository and runs ten small open models with real weights on an old Kintex-7 FPGA card, Qwen3.5-2B at 12.1 tokens a second; it describes itself as developed by AI, which is the author's claim with the repository offered as evidence, and its interest is less the speed than how far agents can go in designing the chip that runs their own inference.

The money is in the register and the meter

DeepSeek is considering expanding its funding round to as much as 100 billion yuan, about US$14.9bn and double its original target, two people familiar with the talks told CNBC; state-backed funds, the investment arms of listed companies and venture firms are competing for allocation while funds raised from individual investors are turned away, the company is said to be seeking a valuation of about 500 billion yuan (US$75bn), and Bloomberg separately reports it is close to securing at least 80 billion yuan with an IPO in view for early 2027. A register of state funds and strategic industrials reads more like national-champion financing than venture capital, and the figures remain fluid: CNBC's sources say the amount could still change. In the West the pricing question sits inside the plans. SemiAnalysis's limit tests estimate that Claude Opus 5.5 on an Anthropic plan delivers roughly five times the API-equivalent value of GPT-6.1 Sol on OpenAI's at the mid tier, with the two close at the top tier; it estimates subscriptions are about 10 per cent of Anthropic's revenue but can take more than 40 per cent of its inference compute, and at 20 per cent utilisation models gross margins of about 6 per cent for heavy Opus use. One provider confirmed a small A/B test on limits during the testing, so limits can move silently, which leaves them a disclosure gap for anyone modelling either lab ahead of a listing.

A supplier says the buildout is supply-bound

Nokia chief executive Justin Hotard told CNBC that customers would build AI datacentres twice as fast if they could, citing shortages of memory chips and energy, and that even with no new frontier release for three years the industry could make large progress deploying what already exists; he sells the networking inside and between datacentres, so his view is not neutral, but the claim is a specific one, that capacity is limited by supply rather than demand. Networking names rose in Tuesday's session, Marvell 5.8 per cent, Arista 4.1 and Broadcom 3.7, and we do not attribute those moves to the interview. Where the capacity goes is shifting as well. From 1 January, datacentre projects in rural census tracts become eligible for opportunity-zone tax benefits under the One Big Beautiful Bill Act, Wired reports, with Searchlight Institute research finding more than 100 rural projects in development that could qualify and the government estimating the rural expansion's cost at US$40.9bn over a decade. Microsoft, Meta and Amazon told Wired they do not use the program, and Google did not respond. The only requirement is capital investment, with no test for jobs, claimants need not disclose, and Senator Josh Hawley has introduced a bill to strip opportunity-zone funding from datacentres, so the likeliest beneficiaries are lower-profile developers and private-capital vehicles, taking on legislative risk with the saving.

Agents get a handshake and a named attack

Meta, Walmart, Stripe and a handful of other companies, in an effort led by Sierra co-founder and OpenAI chairman Bret Taylor, are publishing a personal agent protocol, an open standard for how agents acting for consumers authenticate to businesses and exchange card and personal details, CNBC reports; Amazon is not among the members CNBC names, and neither OpenAI nor Anthropic is on board yet, though Taylor says he expects them to participate. In the same week, Ars Technica reports that Google and four other organisations have acknowledged flaws in which one compromised agent passes harmful instructions to other internal agents, the most severe, rated 8 out of 10 and since fixed, in Google's MCP toolbox for databases. The independent researcher who found them, Syed Anas Mohiuddin, calls the class protocol pivoting, while X41 D-Sec's Markus Vervier calls it a subclass of indirect prompt injection; both readings agree that agents built to trust one another will act on an instruction delivered as a delegated task that a model might otherwise reject. Maintainers face the volume side of the same shift: GNOME issued 13 CVEs in 2023 and 141 in 2026 to 30 September, a rise developer Michael Catanzaro attributes mostly to AI-generated reports, and he argues quality software can no longer be maintained without AI vulnerability scanning, while cautioning that AI reports often exaggerate severity and that even good ones, in volume, swamp volunteer maintainers.

AI's science gets harder to score

Thomas Bloom has frozen new proof claims and comments on erdosproblems.com, removed every problem's open or solved status along with the solved count, and dropped credit-giving language for future results, human or AI, writing that the main way people now use the site publicly is to advertise AI-generated proofs, often without explanation; he notes much recent progress came from renewed human effort, and will promote expositions and link Lean formalisations once they are verified. The most public scoreboard for AI mathematics has removed its score, so claims of capability will have to rest on verified repositories or papers. In materials, a post on the Vals AI blog reports that a team of Claude Opus 5.5 agents proposed two candidate room-temperature magnets for spintronic memory, a new compound, YBaMnFeO5, and KV[Cr(CN)6], first made in 1999. Both are predictions, with the calculations, code and caveats public on GitHub: the new compound's required atomic ordering breaks down below the temperatures standard synthesis needs, and the 1999 material exists only as a wet powder on which the two simulation methods disagree, so the more interesting result may be that the agents found an overlooked match in a 27-year-old paper.

Read all 12 signals →