Afternoon edition — Tuesday 20 October 2026

DeepSeek published DeepInfer on GitHub under the Apache 2.0 licence and announced it on 20 October. The repository describes it as the high-performance inference framework that powers DeepSeek's large-scale production services. Inference software runs a finished model on chips and turns requests into answers, so it largely decides how fast and how cheaply a model can be served.

This release supports one model, DeepSeek-V4.1-Flash, on Nvidia SM100 GPUs and Huawei Ascend 950 NPUs. Listed features include optimised kernels such as FlashMLA, DeepGEMM and DeepEP, attention and mixture-of-experts parallelism, draft-token decoding with confidence-scheduled verification, batch-invariant inference, deterministic sampling and structured outputs. The batch-invariant and deterministic modes work on Nvidia hardware only.

The public code covers offline inference. DeepSeek says a complete online service also needs a KV cache system, prefill and decode disaggregation and load balancing across data-parallel replicas, and none of these is included. The README says DeepInfer was originally built primarily on vLLM, with additional inspiration from SGLang and TensorRT-LLM.

Contributors to llm-d announced version 1.0 on 20 October. Red Hat, IBM, Google Cloud, CoreWeave and Nvidia founded the project in 2025, and the Cloud Native Computing Foundation now governs it. It is a Kubernetes-native stack for serving large language models across many accelerators.

According to the announcement, 1.0 adds priority- and fairness-aware flow control so several tenants can share one fleet, router high availability with rolling updates, prefill and decode replicas managed as one unit, a v1 configuration API, and session-aware routing that keeps multi-turn agent conversations close to their cached context. The project says it has more than 500 contributors from over 25 organisations.

Google Cloud says it is integrating llm-d 1.0 into Google Kubernetes Engine across its TPU and Nvidia GPU portfolio. Nvidia says it will bring llm-d to its Vera Rubin systems.

Linux Foundation Research published a report on 20 October, Navigating Open Source Software Development in the Generative AI Era. Red Hat sponsored it, with support from Futurewei, OpenSSF and LF AI and Data, and it is being presented at the All Things Open conference in Raleigh.

59% of maintainers report growing contribution volume since generative AI arrived, rising to 69% in large projects. Maintainers estimate that 25% of incoming pull requests are low-quality AI slop, and 31% in large projects. Respondents estimate that AI cut code-authoring time by 52%, but 21% say high-complexity tasks take longer with it.

Governance lags behind use. Only 35% of projects have a formal policy permitting generative AI, while 32% have no policy and no plans for one. 61% of contributors say they review and fully test AI-generated code before submitting it, but disclosure of AI help is inconsistent.

Google Public Sector unveiled Gemini for government services on 20 October at its Public Sector Summit. It is a packaged agent product for US federal, state and local governments, available first in preview. Google says the State of Missouri and the US Department of Health and Human Services already use it.

The package bundles pre-built skills for tasks such as analysing regulations, reviewing procurement bids and redacting Freedom of Information Act releases; connectors to Google Workspace, Microsoft SharePoint, Outlook 365, Salesforce, ServiceNow and Oracle applications; and third-party agents from partners including Accenture Federal Services, Deloitte and SAIC. Google says agency data, prompts and outputs stay within the organisation's cloud perimeter.

Chris Hein, field CTO at Google Public Sector, told Nextgov/FCW that Google will prescreen the partner agents and that agency IT leaders, not individual users, will switch them on. He said general availability is expected later this year, with no target date.

Salesforce announced a definitive agreement on 20 October to acquire Braze, a customer engagement platform, for about $5.66 billion in enterprise value. Holders of Braze Class A shares would receive $47.50 per share in cash.

Salesforce says marketers now need agentic tools that act on customer behaviour in real time, and that Braze's real-time engagement and AI decisioning would complement its Data 360, Customer 360, Agentforce and SIForce layers. Braze serves about 2,770 customers, including Canva, Instacart and Venmo.

The deal is expected to close in Salesforce's fiscal year 2028, subject to approval by Braze shareholders and regulators. Salesforce shares fell 2.6% in premarket trading, according to a Reuters report published by CNBC.

Arda, the startup co-founded by former OpenAI chief research officer Bob McGrew, has raised $230 million at a $1.5 billion valuation, the company told CNBC on 20 October. Altimeter Capital, Accel and Khosla Ventures led the round.

Arda makes AI software that helps manufacturing engineers set up and change assembly lines. McGrew told CNBC: "It's a user interface and an agent that does for manufacturing engineers what Claude or Cursor does for programmers." He said Arda's models are already deployed with defence-technology company Anduril, and that robotics is next on the roadmap.

Arda runs a mix of its own models, open-source models and some frontier models. Its co-founders include Augustus Odena, who helped start Adept, and former Palantir engineers Alex Mark and Jakob Frick.