April 2026 contained two releases that, taken together, describe where open-weight AI actually is. On the 8th Meta shipped its first frontier model without open weights. On the 24th DeepSeek shipped one with them. This briefing sets out what each is technically, what the licences on the Chinese models now actually say, where that leaves Europe, and what the AI Act does to the whole arrangement. It is written on 30 April with the sources available then.
Meta closes
Meta announced the Llama 4 family on 5 April 2025: Scout, Maverick, and a two-trillion-parameter Behemoth. Behemoth slipped repeatedly on capability concerns and was never publicly released, which left Meta without a genuinely frontier open model going into 2026. Muse Spark, released on Wednesday 8 April 2026, was the first release under Alexandr Wang's leadership of the AI effort, and it was closed. Meta's AI app climbed to number five on the App Store within a day, which TechCrunch attributed directly to the launch.[1] For eighteen months Llama had been the default answer to "which open model", including for a great many European firms that wanted weights they could host themselves. That answer expired quietly.
DeepSeek V4, and what is underneath it
DeepSeek's January paper set the direction. Engram, published on 13 January with a Peking University team under founder Liang Wenfeng, addresses a specific waste: transformers burn the same expensive attention on simple factual lookups as on hard reasoning. The researchers' own framing is that "the problem is that Transformer lacks a 'native knowledge lookup' ability".[2] Their fix runs hash-based constant-time lookups on short token sequences against very large embedding tables, with a gating layer that suppresses retrieved patterns contradicting the context, applied at some layers and not all. Reported gains: reasoning accuracy from 70% to 74%, knowledge retrieval from 57% to 61%, and a hundred-billion-parameter embedding table offloaded to host memory at under 3% throughput cost.[3]
V4 itself, per the day-zero engineering write-up published alongside the serving code, mixes three attention types per layer: sliding-window attention over the last 128 tokens, a top-512 sparse attention over a four-to-one compressed cache, and a dense attention over a 128-to-one compressed cache. Add a generalised residual connection for gradient flow and native four-bit expert weights, and the result supports a one-million-token context. The inference numbers are the striking part: decode throughput on a B200 falls only from 199 to 180 tokens a second across a range from 4,000 to 900,000 tokens of context, about a tenth, where a naive implementation would fall off a cliff.[4]
| Model | Released | Size | Licence |
|---|---|---|---|
| DeepSeek V4 | 24 Apr 2026 | 1M-token context, sparse MoE | Permissive |
| GLM-5.1 (Z.ai) | 9 Apr 2026 | 754B | MIT |
| Kimi K2.6 (Moonshot) | Early 2026 | 1T total, 32B active, 256K context | MIT plus attribution above 100m users or $20m/month |
| MiniMax-M2.7 | 18 Mar 2026 | Not disclosed | No commercial use without written permission |
| Mistral Large 3 | Dec 2025 | 675B | Open weights |
| Muse Spark (Meta) | 8 Apr 2026 | Not disclosed | Closed |
The hardware story is the one with geopolitical weight. V4 launched with same-day support on Huawei Ascend 950 silicon in both its variants, with Cambricon and Hygon also completing day-zero adaptation, and with Nvidia publishing its own guide to running V4 on Blackwell. By 29 April Reuters reported Chinese technology firms scrambling to secure Ascend chips after the launch, with DeepSeek itself said to have ordered 160,000 Ascend 950DT accelerators.[9] Export controls were designed to make Chinese frontier training hard. What they produced is a Chinese frontier model that runs on Chinese chips.
What the licences actually say
"Open" has stopped being one thing, and a business that reads the headline instead of the file will get this wrong. Moonshot's Kimi K2.6 uses a modified MIT licence: standard MIT plus a single added obligation that products derived from the software "must display 'Kimi K2.6' prominently in their user interface if they exceed either 100 million monthly active users or generate more than $20 million in monthly revenue".[6] For almost every company that is no constraint at all. MiniMax went the other way. M2.7, published on 18 March, is tagged neither MIT nor modified-MIT, and its licence file prohibits commercial use without separate, prior written authorisation, defined broadly enough to cover any paid product, any commercial API deployment, and any commercially deployed fine-tune; free use is limited to personal projects, self-hosted development and non-profit or academic research.[7] That is a binary gate, not a revenue threshold, and it is a material difference between two models that both appear on the same leaderboards under the same word.
Europe
Europe's April news was consolidation. On 24 April, Canada's Cohere agreed to acquire Germany's Aleph Alpha, with Germany's Schwarz Group leading a €500m structured financing into Cohere's Series E and Schwarz's Stackit sovereign-cloud service becoming the technical foundation, subject to shareholder approval and competition clearance.[10] Cohere's Aidan Gomez framed it as organisations "demanding uncompromising control over their AI stack"; Aleph Alpha's Ilhan Scheer spoke of "access to powerful, yet controllable AI they can truly own"; Forrester's Thomas Husson called the result "a unique transatlantic player designed to challenge the dominance of US giants".[10] The honest reading is that Europe's best-funded national champion sold, and the buyer is not European.
What Europe does still have is real, and it is not a frontier model. Mistral Large 3 shipped in December 2025 at 675 billion parameters with open weights and three small companion models. OpenEuroLLM, funded under the Digital Europe Programme, runs as a consortium of Charles University, AI Sweden, Aleph Alpha, AMD's Silo AI, the Barcelona Supercomputing Centre, Jülich, CINECA, CSC Finland and SURF, sharing training code and evaluation frameworks publicly.[11] And Stanford's AI Index, published on 14 April, reported that the US and Chinese model-performance gap had effectively closed, that the US invested $286bn in AI in 2025 and hosts more than ten times any other nation's AI data-centre capacity, while China leads on volume, industrial robot installations, citations and patents. Adoption, notably, is highest neither in the US nor China: 61% in Singapore, 54% in the UAE, around 28% in the US.[12]
The rules on top
The AI Act's treatment of open models is narrower than it is usually described. Article 53(2) exempts providers of models "released under a free and open-source licence that allows for the access, usage, modification, and distribution of the model", with parameters, architecture and usage information public, from two of the transparency duties. The same paragraph then closes it: "This exception shall not apply to general-purpose AI models with systemic risks."[13] And Article 51(2) presumes systemic risk whenever cumulative training computation exceeds 10²⁵ floating-point operations.[14] None of the models in the table above has published a confirmed operations count, but on their scale DeepSeek V4, GLM-5.1 and Kimi K2.6 are all plausible candidates. The practical consequence for a European business is that hosting an open Chinese model does not move it outside the Act, and the obligations that matter attach to the deployer as well as the provider.