Two Open-Weight Models Shipped on Thursday, and One of Them Needs Meta's Permission First

Two open-weight models shipped on Thursday, and one of them needs Meta's permission first

Meta published Muse-Spark-Open-816B-A42B on Hugging Face on 15 October: 816 billion total parameters, 42 billion active, a mixture-of-experts transformer with 256 experts and eight routed per token, a 1.8B perception encoder, and a 1M-token context at inference. Muse Spark has been proprietary since its debut in the spring, and Meta's only previous open Muse release was Glimmer, a 30B Apache 2.0 distillation of Spark itself. Two things qualify the word "open" here. The repository is set to manual approval, so the weights need Meta's sign-off before anyone can download them. And the model card says that "to run on your own system, the model requires a GPU with at least 700GB of VRAM for NVFP4 and 2TB for BF16", adding that it "is also optimized to run on two Mac Studios with M5 Ultra using MLX".

Then there is the licence. Section 3 of the Muse Model License says that where a licensee offers model access as a product or service and "the aggregate gross revenue of the Licensee and its affiliates exceeds twenty (20) million US dollars" during "any consecutive twelve (12) month period", the licensee "must notify Meta and enter into a separate written agreement with Meta before offering the Software or any Derivatives as part of that Model as a Service business". The threshold is group revenue, not AI revenue, the clause carries no geographic carve-out, and "or any Derivatives" catches a host serving its own fine-tune. Internal use, research, and using the model as a component of a product that is not model-as-a-service are all excluded — so this bites hosts, not adopters. Nvidia's release the same day has no equivalent gate: Nemotron 3.5 Super VL is 120B total and 12B active, licensed under OpenMDW-1.1 per its model cards, and all four of its Hugging Face repositories — base, post-trained, NVFP4 and the quantisation-corrected variant — are ungated, with training recipes, data and reinforcement-learning environments published alongside. Neither release appears in Artificial Analysis's changelog, which has no entries for 14, 15 or 16 October, so on the evidence available neither has a third-party score: the Intelligence Index figure of 39 on Meta's card, its xhigh reasoning-effort configuration, carries no attribution, and Nvidia's claim that "on a single GPU, this model delivers up to 5x higher throughput for agentic tasks compared to other open models in its size category" names its comparator only in a figure's alt text, as Inkling Small, with no GPU, precision or concurrency stated.

Microsoft finished a 2.8-trillion-parameter base model and will license it to "trusted partners"

Microsoft AI said on Thursday it has completed pre-training of MAI-Base-2, an MoE foundation model with "80 billion active and 2.8 trillion total parameters", trained on "16k GB200 GPUs over 38 days in August and September 2026". Against DeepSeek-V4-Pro-Base, which it calls "the most capable base model with open weights", Microsoft puts its recipe at "4.2–12.3x more FLOPs-efficient" across four held-out evaluations, and says MAI-Base-2 "outperforms every open-weight model we tested on our held-out evaluations". That comparison set also includes Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL — though Microsoft notes that "only post-trained checkpoints are available for these models" and reports their results "with the caveat that post-training may affect these evaluations". Only the DeepSeek comparison is base model against base model.

The comparators are named, which is more than most announcements of this kind offer. The evaluations are not: they are held out, which is sound methodology for measuring your own progress and no use to a buyer trying to compare two vendors. And the model does not exist as a product. It is in post-training, with no availability date and no pricing, and Microsoft's stated plan is to "license our model weights to trusted partners through Foundry" — partners rather than customers, on terms it has not published. What a European buyer can take from Thursday is the disclosed cost of entry: 16,000 GB200s for 38 days is the current price of a frontier-scale base model, published by a company arguing it paid less than the alternative.

Barclays says the enterprise AI server boom is memory prices, not AI adoption

Dell fell 11% on 14 October, with HPE down 8% and NetApp down 7%, after Barclays downgraded all three; Oracle lost 9% and CoreWeave 15% over the same week, though ITPro gives no dates for either. The argument behind the downgrade is narrow and checkable. Barclays puts enterprise spending at just 5% of IT budgets on AI, says it does not "see a meaningful AI inflection yet", and concludes that "the strong enterprise revenues are more of a function of memory costs than a move to enterprise AI at this time". That is a claim about what the revenue is made of rather than a forecast of where it goes: the same server sells for more because memory costs more, and that is not the same event as a customer deciding to run AI.

Gartner published the supporting arithmetic the same week. AI providers are expected to spend "nearly $1 trillion on AI infrastructure" in 2026 and "an additional $6 trillion between 2027 and 2030". To cover that announced capital spending plus financing costs and a 12% return on invested capital, Gartner reckons AI data-centre owners "would need to generate a cumulative $8.5 trillion in revenue" — against which it forecasts total enterprise and consumer AI revenue attributable to AI data centres of "no more than $3.6 trillion" for the period, and does not expect the 12% return "until the end of 2032 at earliest". Senior director analyst Will Sommer said Gartner's research "indicates that enterprise and consumer AI revenue will be insufficient to support profitable returns on all planned AI capital expenditures during this period". For anyone signing a multi-year inference commitment this quarter, the relevant detail is that the capacity is being priced on the assumption that the demand arrives late.

Quick Hits

• Bank of America put roughly 1,000 contractors on flaws AI models found, on top of about 3,000 of its own staff. Chair and CEO Brian Moynihan said on 14 October, after third-quarter results, that the outside contractors came in during Q2 — "not in the headcount", because they were hired as consultants — to repair software weaknesses exposed by AI models including Anthropic's Mythos, work about 3,000 of the bank's own people were already assigned to. He put the quarter's remediation cost at "a couple hundred million dollars plus", which also covers anti-money-laundering control repairs; TNW, which reports the figures, does not give a combined total.
• Several hundred researchers want cyber-capable models air-gapped as a first precaution. An open letter published this week and signed by Geoffrey Hinton, Yoshua Bengio, Stuart Russell and Herbert Bos of Vrije Universiteit Amsterdam asks developers to disconnect models with advanced hacking ability from the internet, to adopt containment systems that EdTech Innovation Hub describes as designed to give a high level of assurance that models cannot act outside permitted environments, and to create an international body that standardises containment practice and receives reports of containment failures. Reported signatory counts range from about 400 to 420.
• An agent-building platform exposed its control API to the network with no authentication, and the agent behind it runs with permissions bypassed. CVE-2026-111322, published 15 October, covers Archon 0.2.13 through 0.11.1, where /api/* binds to 0.0.0.0 by default. The record is explicit about where it ends: attackers "send messages to the AI coding agent, which runs with bypassPermissions, to execute commands on the host as the Archon user". VulnCheck scores it 9.3 on CVSS 4.0 and 9.8 on 3.1; NVD's own status is still "Received", so neither score is NVD's.
• A JetBrains coding agent could execute code from a project's own MCP configuration. CVE-2026-111091, published 15 October, says code execution "was possible" in Junie CLI before 26.8.3 via an untrusted project's MCP server config. JetBrains scores it 7.8 on CVSS 3.1 with a local vector requiring user interaction, so someone has to act on an untrusted project — the record does not say which action.
• An inference server can be killed by a field it uses internally, and its health check will not notice. , published 15 October, lets unauthenticated callers crash a LightLLM worker through 1.2.0 by passing pd_kv_trans_params to /generate. One request is enough, and the inference thread dies "while /health stays up and all later requests hang until restart" — the failure mode a load balancer cannot route around. VulnCheck scores it 8.7 on CVSS 4.0 and 7.5 on 3.1; NVD has deferred its own analysis.