Xiaomi ships a 1.02T omnimodal model, and an independent evaluator scores it
XiaomiMiMo published MiMo-V2.6-Pro-RL on 21 September: a sparse mixture-of-experts model with 1.02 trillion total parameters and 42B activated per token, a 1M-token context, and text, image, video and audio in. It runs 384 routed experts with eight active, 70 layers split 60 sliding-window to 10 global attention, a five-layer speculative decoder, and ships FP8-quantised. A Flash variant is 309B total and 15B activated at the same context length. The card's headline table — Terminal-Bench 2.1 at 89.9, OSWorld-Verified at 82.0, CyberGym at 94.0 — is Xiaomi's own, run by Xiaomi, against a comparison set Xiaomi chose.
The independent read arrived the same day: Artificial Analysis scored MiMo-V2.6-Pro at 46 on its Intelligence Index on 21 September, the same value it assigns Grok 4.7, released by xAI that day. That is the number to quote. One thing to settle before planning around any of it: all three repositories — Pro-RL, Flash-RL and MiMo-V2.6-Distill-Qwen-9B — declare a licence of MIT in the card metadata, and the LICENSE path returns 404 on each. A licence tag with no licence text behind it is an assertion, not a grant, and it is not something a procurement review can accept.
Europe's data-protection board writes down how a fine gets decided
At its plenary in Brussels on 21 September the European Data Protection Board adopted Guidelines 04/2026 on the power to impose administrative fines relative to the other corrective powers in the GDPR. The method has five steps: establish the fineable infringement, identify the liable party, assess whether it was intentional or negligent, weigh aggravating and mitigating factors, then test whether the result is effective, proportionate and dissuasive. Fourteen worked examples are attached. These replace the old Article 29 Working Party guidelines and sit next to Guidelines 04/2022, which cover calculating the amount rather than the decision to fine at all.
They are draft. Public consultation runs to 13 November 2026, so nothing binding has changed this week. What has changed is that the reasoning a regulator will publish is now written down in advance, and step three — intentional or negligent — is the one that matters for anyone running models over personal data. Documentation of what a deployment was checked for, and when, is the material a regulator weighs at that step. The Board also adopted the final version of its guidelines on the interplay between the DSA and the GDPR, pending linguistic checks.
Salesforce tries three ways to charge for agents, a week after a seven-hour outage
Salesforce is testing seat licences, a consumption model it calls Flex Credits, and outcome-based fees — charging per resolved support case rather than per user or per token — The Register reported on 21 September, alongside multi-year Agentic Enterprise License Agreements and spend commitments branded Salesforce Commit. Deputy CFO Mike Spencer, speaking at an investor conference last month, described the combination as “an anxiety-filled architecture” and said consumption revenue needs three to five years before it becomes material to the company's numbers.
The timing is the interesting part. A global Salesforce outage ran for more than seven hours on 16 September, during the second day of its own Dreamforce conference. Outcome pricing moves delivery risk onto the vendor, which is attractive until you ask what it does when the outcome does not arrive — whether an unresolved case during an outage is simply unbilled, or whether the contract says something more. That clause is worth reading before the pricing model is, and it is the kind of detail that does not appear in a pricing page.
Quick Hits
- vLLM 0.30.0 adds watermarking and a weight cache that cuts cold starts. The release, published 22 September with 762 commits, adds Gumbel-max watermarked generation and detection behind a keyed pseudorandom function with per-request opt-out, and a per-GPU weight-cache daemon holding post-quantised sharded weights mapped over CUDA IPC. The project measures engine init falling from 28.9s to 8.2s on an H200 after freezing garbage collection during CUDA-graph capture — a separate change from the weight cache. New model support includes DeepSeek-V4.1-Flash, GLM-5.3-Flash, K2-Horizon and Cohere Compass.
- Grok 4.7 gains two index points and spends three times the tokens getting there. Artificial Analysis benchmarked the model on 21 September at 46 on its Intelligence Index, up two on Grok 4.6, with AA-Briefcase Elo up 111 — and regressions on AA-LCR and AutomationBench-AA. AA's own measurement puts it at roughly 81,000 output tokens per index task against 36,000 for Grok 4.6 and 27,000 for GPT-6 Astra. xAI's own release notes price it at $2 in and $6 out per million tokens below a 200k prompt, doubling above it.
- Brussels proposes a rating scheme for data centres and opens a second consultation under it. The Commission published the proposal on 21 September for a common EU rating scheme covering energy use and grid integration. It is a proposal, not adopted law, and builds on Delegated Regulation (EU) 2024/1364, which set the first-phase reporting KPIs for data centres above 500 kW. In parallel DG ENER opened a 12-week consultation on minimum performance standards under the Energy Efficiency Directive, closing 14 December, with a legislative proposal planned for Q2 2027.
- Eutelsat books another 229 satellites from Airbus for about €1bn. The order, reported 21 September, takes the order book to 669 new craft with launches running to 2034, and Eutelsat frames it around IRIS², the EU's sovereign connectivity programme, whose Ku-band constellation it expects operational in mid-2032. The 69.5% growth in low-earth-orbit revenue cited alongside it is the company's own figure for its year to June 2026.
- Hugging Face claims a 3–30× faster tokenizer, measured on its own harness. The , written up on 21 September, replaces the regex pre-tokenizer with SIMD bitstream splitting, adds a thread-local word cache and an allocation-free merge loop, and reports encoding 3× to 30× faster than v0.23 single-threaded across ten model families, scaling at 76% of linear over eight workers, with identical token IDs. The benchmarks are Hugging Face's own, on Apple silicon; 1.0.0 has not shipped.