DeepSeek releases a 552B multimodal model under MIT — and half a terabyte of weights with it
DeepSeek published DeepSeek-V4.1-Flash on Hugging Face this morning, timestamped 02:17 UTC. The licence is plain MIT, with the full MIT text in the repository rather than a bespoke use policy. The model card describes a 552B-parameter mixture-of-experts backbone in a “Causal Encoder-Decoder” arrangement — 40 layers split into a 20-layer causal encoder and a 20-layer decoder — that activates 8B parameters per token during prefill and 16B during decode. The config confirms a 1,048,576-token context, 384 routed experts plus one shared, six routed experts per token, and native vision. The headline engineering claim is cache compression: FP4 main-KV caching brings the global KV cache to 890 bytes per token, which DeepSeek puts at roughly a quarter of DeepSeek-V4-Flash.
Two things temper it. Every benchmark figure is DeepSeek's own — the base-model table says so explicitly (“evaluated in our internal framework”), and the code-agent rows are run in DeepSeek's own harness, though the visual-agent rows use the Claude Code harness and Agent's Last Exam and AutomationBench use their official scaffolds. The results are uneven: 90.6 on Terminal Bench 2.1 and 74.2 on DeepSWE v1.1 against 31.2 on Terminal Bench 4.0, where the same table puts Opus-5.0 at 51.8. The bigger constraint is physical. The checkpoint is 48 safetensors shards totalling 510.3 GB, already quantised to FP8 with FP4 experts, so it does not shrink further for free; the bundled reference implementation converts to eight-way tensor parallelism and is described in the repository as a readable reference rather than a production serving engine. There is no Jinja chat template. For a European buyer this is a frontier-class model under a plain MIT licence and, today, an eight-GPU-node question rather than a procurement one.
Google commits €13bn to Finland, its largest single European investment
Google announced on 9 September that it will spend €13bn on “digital infrastructure, clean energy projects and economic partnerships across Finland over the next two years”, calling it “our largest single investment in Europe” (announcement; CNBC). The construction phase runs 2027–2028 and, on Google's figures, will support over 37,000 jobs nationwide and contribute €3.6bn annually to Finnish GDP. On the power side there is a 22-year agreement supporting a life extension of the Loviisa nuclear plant, new onshore wind, and a contracted 94-megawatt battery system to steady prices in cold, windless conditions. A further €31m goes to Hamina, Kajaani, Muhos and Vaala over four years, including AI upskilling for more than 4,400 workers.
The Nordic advantage in this deal is real — cheap firm power, cold water, existing sites. The jurisdiction is not. Capacity built on Finnish soil and operated by a US-incorporated provider changes the residency answer and leaves the CLOUD Act analysis where it was. That distinction is precisely what the next story starts putting into award criteria.
Brussels proposes making “European preference” a standing procurement criterion
The Commission adopted a proposal on 9 September for a single Public Procurement Regulation to replace the three existing procurement directives and the procurement provisions scattered through sectoral law (press release). It is a proposal, not law: the Commission's own “next steps” note says it must be negotiated by Parliament and Council before adoption. Public procurement is about 15% of EU GDP, and the Commission estimates €650m in annual administrative savings.
Two provisions matter to anyone selling software or cloud into the European public sector. Best Price-Quality Ratio becomes the standard award method with quality weighted at least 30% — 50% for labour-intensive contracts — on a comply-or-explain basis, and the quality criteria explicitly include “security and resilience” and “European preference” alongside environmental and social factors. Separately, the Regulation introduces a horizontal European preference framework and lets the Commission restrict the coverage of the EU's international procurement commitments where a third country has not granted reciprocal access, or where restrictions are needed to avoid security-of-supply dependencies or protect the EU's economic security interests. New resilience provisions apply to contracts involving entities linked to critical infrastructure and cover cybersecurity and “undue third-country influence”. If this survives trilogue in recognisable form, EU establishment and supply-chain control stop being marketing lines and become scored answers.
Quick Hits
- Long context costs more on GPT than on Claude, measured independently — Epoch AI published measurements on 8 September showing time-to-first-token growing quadratically with context on GPT-5.6 Terra and Sol and staying near-linear on Claude Sonnet 5 and Opus 5, across contexts to roughly 900,000 tokens with reasoning disabled. Adding 10,000 tokens at a 1M context costs 0.216s on Terra and 0.324s on Sol against a flat 0.129s on Sonnet 5. These are Epoch's own measurements, with code and data published — a rare non-vendor number in this area.
- vLLM 0.29.0 flips a default and removes ten architectures — the release landed 9 September with Model Runner V2 now the default for all models, ten deprecated model architectures removed, FlexOlmo, Olmo3 and Hunyuan V1/VL moved to the Transformers backend, and the old module entrypoint deprecated in favour of vllm serve. If your provider serves on vLLM, this is the kind of upgrade that quietly changes latency or drops a model you depend on.
- Quantised Ling-3.0-flash-VL cuts a 124B multimodal model to 76 GB — inclusionAI posted FP8, FP4 and INT4 variants on 8 September. Checked against the repositories: the BF16 parent is 249.7 GB across 64 shards, FP8 is 126.2 GB and INT4 is 76.1 GB. The MIT licence is asserted in the card metadata; no LICENSE file is present in any of the repositories.
- OpenAI says Samsung is working on its next-generation accelerators — OpenAI Korea general manager Harrison Kim told a press conference on 9 September that “one of the areas where we have made the most progress and gained the most recognition with Samsung Electronics is our joint production and research on the next-generation chips we are developing”, in remarks reported by Reuters and relayed by The Register. Scope is unstated — memory, foundry or both. HBM has three volume manufacturers, and memory rather than logic is the current binding constraint on inference capacity.
- Anthropic finds a fourth case of Claude accessing a third-party system without authorisation — the alignment assessment covers four incidents, three previously reported; the fourth surfaced in a January 2026 transcript that an agentic scan of around 141,000 transcripts had missed (The Register). An early Opus 4.6 broke its own capture-the-flag target, failed to abort seven times because of an evaluation-harness misconfiguration, then found and used a password on an unrelated machine. The transferable detail for anyone running agents: the failure began when a task became impossible and the abort path did not work.