Can useful AI run on equipment that smaller institutions control?
Direction of power
Leaning dispersion
Window
Opens
Confidence
Medium
Movement since August
Better evidence of deployable local capability and permissive alternatives
End-of-September assessment
September improves the case that useful AI capability can reach smaller institutions without continuous hyperscale inference. Taiwan's demonstration pharmacies combine a distinct 20B open-weight model with public drug information, local hardware and implementation support. This is stronger practical evidence than a benchmark alone, although it remains a field-deployment programme with outcomes still being evaluated.
MiniCPM5-2B and the smaller members of the IFM and Xiaomi families broaden the independently deployable model layer. Their licenses and artifacts matter because users can potentially run and adapt systems outside the original service. Large checkpoints, benchmark scores and multiple runtime options do not establish low total ownership cost, good performance in every domain or a maintained system for a small public body.
The accumulated direction remains leaning dispersion, with a clearer practical route than in July's more mixed assessment. The strongest September inference is a lower minimum infrastructure floor for some tasks. It is not a demonstrated general replacement for frontier models. Hardware, integration and evaluation can become the next dependency points even when weights are available.
Other evidence. The Global Open-Weight Model Landscape supplies licensing, runtime and availability comparisons. Google's Home MCP interface is a connection to a device ecosystem, not evidence that the model behind an assistant can be independently run. Public evaluation remains necessary to distinguish accessibility from demonstrated performance.
What would change this assessment. Independently replicated deployments with measured quality, maintenance cost and offline operation; portable runtimes on affordable hardware; public or nonprofit support for integration; or restrictive licensing, hardware concentration and maintenance burdens that make nominally open systems impractical.
Further deployment evidence. Edge0's native engines were released September 30 for a September 8 checkpoint. Its low-memory and speed claims come from the publisher's desktop test and cannot be generalized to phones. It remains an unscored deployment example pending independent quality, cost and performance evidence. See the unscored evidence register for dates, retained dispositions and source links.
Supporting scored events
The complete relevant event list follows. Primary and secondary references preserve one underlying event and score. Source holds and later commencement dates remain explicit; publication does not resolve them.
September scored actionSeptember 30, 2026 evidence cutoff
Qualcomm and ASUS launched a Pharmaceutical AI Agent field-deployment program for more than 50 demonstration pharmacies across southern Taiwan. The system uses the open-weight gpt-oss-20b model, public Taiwan FDA drug information, OCR and retrieval components, and local/on-premises inference on donated AI endpoint devices to support medication-safety and pharmacy workflows.
Limits and interpretation: Do not describe the deployment as compressing gpt-oss-120b into 20B. OpenAI released gpt-oss-120b and gpt-oss-20b as distinct models; the Taiwan program specializes and deploys the 20B variant. Field validation is beginning, demonstrated performance at scale remains limited, and the deployment still depends materially on Qualcomm hardware and ASUS/Taiwan implementation infrastructure.
OpenBMB released MiniCPM5-2B, a 2.517B-parameter dense model with a 131,072-token context window, under Apache 2.0. The released weights are usable through standard local and server runtimes including llama.cpp, Ollama, LM Studio, MLX, vLLM and SGLang, and OpenBMB released substantial associated pre-training, SFT/agent and RL data artifacts. Artificial Analysis independently scored the model 15 on its Intelligence Index v4.2, the highest current score among open-weight models below 4B total parameters.
Limits and interpretation: Benchmark strength is independently demonstrated, but broad real-world institutional deployment and sustained agentic performance on ordinary endpoint hardware have not yet been independently established. The open training artifacts also do not mean that every underlying dataset carries the same Apache license.
Abu Dhabi’s Institute of Foundation Models released the K2 Horizon family from 0.9B through 375B parameters with permissively licensed weights, training code, methods, intermediate checkpoints and substantial training-data artifacts or reproducible data recipes. The 0.9B model is already small enough for commodity local hardware, while IFM’s stronger smartwatch/glasses deployment claim remains unverified.
Action/date record: September 2026; processed September 16. Approved score: Dispersion +1. Primary lane: 5. Secondary lanes: 4.
Google exposed Home capabilities through MCP, allowing outside agents to interact through a standardized tool layer rather than only through Google’s own assistant. The score remains +1 because Google retains platform authorization, device access and execution control.
Source hold: The approved ledger refers to an alert source archive, but an operative URL was not recovered from the records materialized for this review. The score is preserved; actor, precise action date, text and scope need source-archive reconciliation before this becomes fully verified public evidence.
September scored actionSeptember 30, 2026 evidence cutoff
Xiaomi released full MiMo-V2.6 Pro-RL and Flash-RL checkpoints, together with a 9B distilled model, under the MIT License. The family supports multimodal inputs, tool use, a one-million-token context and independent deployment through common open runtimes.