The 8 DeAI stories that mattered this week

The week open AI's infrastructure question went from "who hosts the models" to "who owns the pipes": Nvidia agreed to buy Hugging Face for $12.93 billion, while NEAR AI Cloud shipped private inference you can actually verify. Eight stories mattered; here is what each one changes.

Key facts

• $12.93 billion — the price Nvidia agreed to pay for Hugging Face, the hub hosting 3M+ open-weight models.
• 552B parameters, MIT license — DeepSeek V4.1-Flash, the week's biggest open-weights release by reach.
• 375B parameters, Apache 2.0 — IFM's K2 Horizon flagship, a sparse MoE with a 512K context window.
• Hardware attestation, not policy — NEAR AI Cloud now signs every response from inside Intel TDX / NVIDIA TEE enclaves.
• €3 billion — Mistral's raise at a €21B+ valuation, explicitly framed around sovereign open-weight AI.
• $0.003 per million tokens — DeepSeek V4.1-Flash cached-input price, the week's lowest verified number.

1. Nvidia agrees to buy Hugging Face for $12.9B

The week's largest story by any measure. Nvidia signed a definitive agreement to acquire Hugging Face — 3 million models, 18 million developers, the default distribution layer for open weights — for $12,930,300,000, closing expected H1 2027 pending regulators. Nvidia says the hub stays open and multi-vendor; that is a corporate commitment, not a contractual guarantee outsiders can check yet. Why it matters: the neutrality of the model hub is what lets any lab publish weights to everyone, and a GPU vendor owning that layer is a structural conflict regulators will examine. Full story.

2. NEAR AI Cloud puts private inference in Intel and NVIDIA TEEs

The week's most consequential technical story. NEAR AI Cloud is now serving open-weight models inside Intel TDX confidential VMs with NVIDIA GPU TEEs on H100/H200 hardware — TLS terminates inside the enclave, each enclave holds a signing key that never leaves the hardware, and every response is signed so users can verify it against the attestation report. The privacy claims are still NEAR's own, but they are checkable claims now, which is a categorically different thing from a retention-policy promise. Why it matters: this moves "private inference" from marketing language to a verification path — the same shift our weekly deep dive tracks in detail. Full story.

3. IFM releases K2 Horizon open models, weights up to 375B

The Institute of Foundation Models released six Apache 2.0 models from 0.9B to a 375B-parameter mixture-of-experts (23B active per token) with a native 524,288-token context. Weights are on Hugging Face now; training data and code are promised but not yet public, and all benchmarks are IFM's own — treat the numbers as vendor claims until third parties replicate. Why it matters: a fully-licensed 375B-class model widens the top end of what you can self-host or run on a decentralized network without a lab's permission slip. Full story.

4. DeepSeek V4.1-Flash: MIT weights, 1M context, $0.003/M cached input

DeepSeek published V4.1-Flash — a 552B MoE with a causal encoder-decoder design, 8B active on prefill and 16B on decode — under MIT on Hugging Face, and flipped its API's flash model to the new checkpoint. Verified: the weights, the license, the pricing page. Claims: the benchmarks, which X spent the day arguing over (1.37M views in five hours) and which nobody has independently replicated. Why it matters: MIT licensing plus a cached-input price of $0.003 per million tokens resets the floor for cheap long-context serving. Full story.

5. DeepSeek retires V4-Flash in a silent swap, spares V4-Pro

The follow-through story, and arguably the more instructive one. A day after the V4.1-Flash launch, DeepSeek retired the legacy deepseek-v4-flash IDs — requests to those names are now served by V4.1-Flash with no version bump — and reversed its plan to shut down V4-Pro, both disclosed only in pricing-page footnotes. Why it matters: the model behind a pinned endpoint changed without notice; if you don't pin dated snapshots and re-run your own evals, your integration's outputs shifted this week and you may not have noticed. Full story.

6. MiniCPM5-2B: the new best sub-4B open model

OpenBMB shipped MiniCPM5-2B, a ~2.5B dense reasoning model with Apache-2.0 weights, ~550B+ tokens of training data, recipes, and an RL stack, with day-one support in vLLM, SGLang, llama.cpp, and Ollama. Artificial Analysis scored it 15 on Intelligence Index v4.2 — the highest for an open-weights model under 4B parameters — with known weaknesses in knowledge, coding, and long context. A widely-shared X figure of 23 does not match AA's published write-up; the verified number is 15. Why it matters: the edge-inference tier keeps closing on last year's mid-size models, and this one ships with the data to reproduce it. Full story.

7. Mistral raises €3B for sovereign open-weight AI

Mistral closed a €3 billion round at a valuation above €21 billion and published the sovereignty manifesto behind it — the same week the World Economic Forum framed open-weight AI as "owned infrastructure" rather than rented access. Why it matters: Europe's open-weight champion now has the capital to compete on frontier training runs, and its pitch is explicitly the anti-locked-in one this site covers. See the day's brief and the K2 Horizon brief for the sovereignty thread running through the week.

8. The 8 best Together AI alternatives in 2026, priced

Our reference-page update of the week: with Together's September 14 serverless deprecation deadline approaching, we re-priced the eight credible alternatives across serverless, fast-silicon, and decentralized options — Fireworks, DeepInfra, Groq, Cerebras, Chutes among them, with Morpheus ranked strictly on the stated criteria. Why it matters: a deprecation deadline is when switching costs get real; the honest comparison is the one that includes real competitors and real prices. Full page.

Also this week: the Friday hub, This Week in DeAI: Hardware-Attested Privacy Goes Mainstream, ties stories 1, 2, and 4 into the privacy-from-policy-to-proof arc, and the daily briefs from Sep 8, Sep 9, Sep 10, Sep 11, and Sep 12 carry the item-by-item sourcing.