A week stacked with flashy releases — Muse Wonder and Wonderful, Gemini 4 Argon — and yet the two unassuming Thursday announcements almost nobody is hyping may matter most. Microsoft's MAI Base 2 and Meta's Muse Spark Open won't top the benchmark discourse, but they speak directly to how businesses are actually going to use AI: owning the intelligence layer, tuning their own models, and refusing to hand all their dependencies to one lab.
The quietest model announcements of the week are the biggest deal for enterprise AI.
Microsoft's MAI Base 2 and Meta's Muse Spark Open aren't going to dominate the benchmark conversation, but they answer the two questions enterprises are actually asking: cost and sovereignty. With frontier tuning on a genuinely frontier-scale base and the most capable US open-weight model to date, companies that don't want to hand their data and dependencies to OpenAI or Anthropic suddenly have far more choices than they did a few months ago — and the US could soon have six companies competing at the frontier, up from two.
Thursday's model announcements from Microsoft and Meta won't change state-of-the-art benchmarks, and Artificial Analysis won't dedicate pages to them. But for reasons that have everything to do with how business usage of AI is changing — and nothing to do with the hype discourse on X — they may be the most significant releases of a crowded week.
A 2.8-trillion-parameter mixture-of-experts model with 80 billion active parameters — Kimi K3's scale — from a completely fresh pre-training run with zero distillation. It outperformed every open model over a trillion parameters on pre-training benchmarks, with a claimed 12x pre-training efficiency over DeepSeek V4 Pro, trained on 16,000 GB200 Blackwells in 38 days.
— Mustafa Suleyman, Microsoft AI CEO, to staff at an internal town hall. Suleyman's town-hall comparison relied on proprietary information Microsoft gets through its OpenAI investment deal and wasn't shared publicly. Either way, it's a remarkable claim from a company that spent years as OpenAI's dependent partner.
Buy-not-build looked prophetic after the $13B OpenAI investment — until the Altman ouster made clear that fiduciary responsibility demanded independence. Hiring Suleyman in March 2024 produced a year of reported acrimony and little else, and even with all its OpenAI access, Microsoft struggled across 2025 before its real strategy emerged this year.
Microsoft's first MAI models were only okay on benchmarks — adequate commodity models deployed across Copilot. But the real product was the business model: Frontier Tuning, letting customers fine-tune their own models on a Microsoft base. Without a state-of-the-art base model, that program could only go so far. MAI Base 2 is what makes it matter.
Running all intelligence needs through the most state-of-the-art subscription turned out to be nowhere near cost-efficient, especially as the agentic era put code to work on problems far beyond engineering — and that was before true mass-scale adoption. The result: token caps as policy, and enterprises becoming model omnivores who move between models by task rather than locking into one vendor ecosystem.
Satya Nadella has been hammering that enterprises must own the intelligence layer around the models rather than hand it to their vendors. While Anthropic and OpenAI answer the cost question with faster, cheaper models, Microsoft's frontier tuning implicitly answers sovereignty too. And if MAI 2's post-trained version approaches state of the art, six US companies could be competing at the frontier — up from two a few months ago.
Meta's 816-billion-parameter model (42B active, million-token context) is built for coding and agentic work — long-horizon tasks, hard software engineering, multi-step reasoning. The smart money thinks it's a bigger deal than the flashy Wonder and Wonderful releases earlier in the week.
Spark Open scores a bit behind Kimi K3 on coding but claims open state-of-the-art on agentic benchmarks like JobBench and MCP-Atlas. Artificial Analysis scored it 39 — seventh among open models, seven points off the open state of the art, but 13 points ahead of Nemotron and Inkling. Unlike Reflection's Beam, it ships with full weights across major inference providers.
Once synonymous with open models, Meta cratered with Llama 4, never shipped the two-trillion-parameter Behemoth, and tore down its AI division. After hundreds of billions committed to the build-out, many assumed Alexander Wang's pledge to return to open models was just talk. It wasn't — Meta is back atop the US open rankings.
Open model researchers note US models had been behind Chinese releases from as early as February 2026; Spark Open is the first US open-weight model with near-frontier agentic capabilities, and US models historically get outsized adoption. Box's Aaron Levie argues this accelerates diffusion: enterprises can mix frontier models for orchestration with low-cost post-trained models for repeatable workflows — dramatically increasing total token consumption.
Apple updated MLX so Spark Open runs smoothly on Apple hardware, Zuckerberg posted a video of Muse on Mac Studios crediting John Ternus, and Ternus congratulated the team in return. Wang credits former Apple exec Daniel Gross with smoothing over past grievances: 'We're excited for mommy and daddy to be playing nice again.' For Apple's new CEO, it's quite a way to announce things will be different.
Full-size Spark Open needs a terabyte of memory — a pair of tricked-out M5 Ultra Mac Studios running around $40,000. Out of reach for prosumers, but not for small startups that care about private inference. And because it's open, a quantized version can cram into a Mac Mini if you accept the trade-offs.
The idea that most enterprises will abandon easy hand-me-a-solution subscriptions is unrealistic in the short term. But a growing portion — especially those with technical capability — will design architectures that move between models based on performance, capability, cost, and efficiency needs. This week, both Microsoft and Meta threw majorly upgraded models into that fray, and American businesses are better off for it.
Following Dario Amodei's pledge to integrate outside evaluators, a six-person team from Meter now has what Anthropic calls employee-like access — company computers, office space, internal tooling. SecureBio and Accenture are next, with Hugging Face in early discussions. Notably, Elon Musk responded on X that SpaceX will probably do the same.
Anthropic says it doesn't fund Meter, but the two share major backers aligned around the effective altruism space — same philosophical roots, so largely the same funders. That context explains why Anthropic is moving fast on this partnership while OpenAI appears more hesitant, and why critics are calling the arrangement security theater.
Anthropic can require redactions only in narrow predefined categories — customer data, legally privileged information, and model architectures — and Meter is free to publicly characterize whether redactions were justified. A three-tier system governs access: Meter alone sits in the top tier, while SecureBio starts in the bottom tier, where publication requires approval.
If Meter is the only evaluator, it's like JPMorgan picking its own bank examiners. Even Meter's own team notes this is an assessor under a voluntary agreement, not an auditor, because there are no set safety standards to audit. The constraint right now: almost nobody qualified to do this work actually works for a truly independent organization.
After the token-maxing feast of Q2 and the token-austerity famine that followed, Cognition says both are wrong: Devin now scores how efficiently engineers use their token budget — estimating equivalent human hours per session — and automatically raises limits for good stewards, the way credit card limits grow. The bet: in hyper-competitive agentic coding, efficiency tools build more trust than encouraging maximum spend.
Cognition's new AI system gives engineers feedback on common agent-usage errors, like ambiguous requests or using an expensive model for a mundane task. Developers hated it — 'software engineering is now a factory role,' 'something very Soviet about this' — and critics note that measuring equivalent human hours incentivizes building the most bewildering thing, not the most valuable one. Still, others concede token allocation discipline is simply a feature of the world now.
The first meeting commissioned a safety report due in 120 days covering cybersecurity, biosecurity, and model control; set up a working group to implement the White House accord signed by the frontier labs and NVIDIA; and assigned security-incident reporting protocols, an FTC-DOJ enforcement committee, and critical-systems hardening plus Department of Labor work on AI's workforce impact.
The White House accord is only 'morally binding,' as the president put it. But a White House official told Semafor that one stick for non-compliance could be the threat of export controls — the same lever used against Anthropic earlier this year.
Analyst Dan Levy sees a possible all-stock deal — roughly 0.7 Tesla shares per SpaceX share — announced before Tesla's December 9 shareholder meeting, driven by the joint Terrafab chip plant and leaving Musk with 37% ownership and ~56% voting power. No bankers hired, nothing filed, and prediction markets put the odds at just 24% — but this is a deal between Elon and Elon, and Elon is already in.
YouTuber Max Blade's Kong Royale — huge island, vehicles, giant gorilla protagonist — was built in a few weeks with Opus 5.5 and a couple of Claude Max accounts using 'hive coding': agents monitor his live stream chat, and overnight they review ideas, select the best changes, and ship a morning build. A year and a half after Peter Levels' plane game, crowdsourced streamer-built games are starting to look like a genuine genre.