Five stories defined the AI news cycle this week: DeepSeek open-sourced the inference engine behind its production services, Mistral put its trillion-parameter flagship into public preview, OpenAI and Google Cloud buried the hatchet, Anthropic unveiled a proactive assistant called Claude Duet, and the Python supply chain weathered a brazen backdoor attack on a popular LLM proxy library. Here is what happened, and why it matters.

DeepSeek Open-Sources DeepInfer, the Inference Engine Behind Its Production Services

DeepSeek released DeepInfer, the high-performance inference framework that powers its large-scale production services, under the Apache 2.0 license. The release targets the DeepSeek-V4.1-Flash line of models and bundles the optimizations that make the company’s API so cheap to operate: custom kernels built on FlashMLA, DeepGEMM and DeepEP, DSpark speculative decoding with confidence-scheduled verification, bounded-replay prefill for SWA decoders, a zero-overhead asynchronous batch scheduler, batch-invariant and deterministic sampling, and grammar-constrained structured outputs.

The framework runs on NVIDIA SM100 GPUs and, notably, Huawei Ascend 950 NPUs, with CUDA 13 and CANN 9.2 support respectively. DeepInfer was built primarily on vLLM, with inspiration from SGLang and TensorRT-LLM, and the initial release covers offline inference; DeepSeek notes that a full production serving stack still requires a KV-cache system and prefill/decode disaggregation tooling of its own.

The Hacker News community pushed the story past 2,300 points and nearly 900 comments. Much of the discussion focused on what it means for a US host to adopt the stack: as one commenter observed, a US provider that adopts DeepInfer to serve DeepSeek models is arguably now “running a DeepSeek product,” with potential National Defense Authorization Act implications. Others framed it more simply: a common set of proven optimizations that every inference provider can now adopt.

Mistral Large 4 — “Le Chonk” — Enters Public Preview, Weights Due This Month

Mistral AI launched a public preview of Mistral Large 4, its largest model to date, unofficially known as “le Chonk.” The mixture-of-experts model packs roughly one trillion total parameters with about 52 billion active per token, is natively multimodal, supports coding, agentic workflows and long-context tasks, and was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own European data centers.

Mistral claims the model is competitive with the strongest open-weight systems globally and “significantly outperform[s] any open-weight model developed in the US or Europe,” with state-of-the-art results among open models in cybersecurity, finance and law — and, in visual grounding, performance that “surpass[es] even frontier closed models.” The company plans to release the weights by the end of the month, after a red-teaming period conducted with cybersecurity leaders, vetted partners and state authorities, who get access to the model with reduced moderation and expanded cyber capabilities.

Large 4 is the first major model funded by Mistral’s €3 billion Series D, and the company is leaning hard into its “forged in Europe” sovereign-AI narrative: the preview API runs on the same in-house infrastructure used for training, and the public preview is available today on Mistral Studio.

OpenAI Models Are Coming to Google Cloud

OpenAI and Google Cloud announced a partnership to bring OpenAI’s frontier models — including GPT-6 Astra — and Codex agentic workflows into Google Cloud’s Model Garden, deepening ties between two longtime rivals. Announced October 21, the deal lets enterprises buy OpenAI models through the Google Cloud Marketplace on a single invoice, draw down existing GCP spending commitments, and count purchases toward Flexible Savings Plans. General availability is expected “in the coming weeks,” and OpenAI models will also appear in Google Antigravity and the Gemini Agent.

Reuters framed the move as OpenAI’s latest enterprise push, underscoring how major clouds have become the primary distribution channel for frontier AI. Google Cloud already sells rival Anthropic models, and the HN thread zeroed in on the strategic oddity: as one top commenter noted, enterprises with GCP commitments will now be able to draw down that already-booked revenue against OpenAI purchases, while another joked that Google was simply “preparing for Oracle bankruptcy.”

OpenAI was careful to note that Microsoft remains its primary cloud partner, even as the company multiplies the channels through which enterprises can buy its models.

Anthropic Unveils Claude Duet, a Proactive Assistant Powered by Opus 5.5

Anthropic introduced Claude Duet, a research-preview “proactive agent” that takes on tasks of all sizes and keeps working while you are away. Powered by Opus 5.5, Duet briefs users every morning, drafts next steps, and “goes quiet until something needs you” — waiting for explicit approval when an action matters. It is available on Pro and Max plans in the US (with a waitlist) and works with more than 6,000 connectors and plugins, drawing on the same memory, skills and apps already in a user’s Claude account.

Anthropic’s marketing walks through a working day with Duet: it summarizes an expiring job offer, tracks a Series A data room (flagging two slides that need an eye), manages a restaurant opening’s checklist — including paying a county health-permit fee once approved — and quietly watches a slipping espresso-machine install. The pitch is judgment plus follow-through: the agent distinguishes what needs the human from what it can handle alone.

The announcement drew roughly 550 points and 500 comments on Hacker News, with discussion centering on whether always-on proactive agents can earn user trust with their permissioning and “wait for your ok” workflows.

Compromised LiteLLM Release on PyPI Installed a Credential Stealer and Backdoor

A troubling supply-chain attack hit the AI tooling ecosystem this week: LiteLLM 1.107.0 on PyPI — a widely used open-source LLM proxy/gateway — was found to carry a credential stealer and a remote-code-execution backdoor. All seven wheels (Linux, macOS, Windows) install three files into site-packages that execute on every Python process start, even code that never imports litellm.

Per an analysis from SafeDep, cook.py acts as a credential stealer harvesting SSH keys, cloud credentials, CI/CD secrets, AI tool tokens, browser passwords and the LLM provider keys stored in a LiteLLM proxy database, encrypting the haul and exfiltrating it to a command server. sitecustomize.py spawns a daemon that, hourly, reads a server address from a GitHub file and downloads and executes code from it. The injection was a single extra line in the pyproject.toml maturin build section — invisible in the sdist, absent from every public commit.

The attacker reportedly held the GitHub CLI credentials of three LiteLLM developers, obtained from logs of an internal deployment exposed to the internet. Using an organization owner’s token, they bypassed branch protection and pointed the private release workflow at a lookalike repository (Berri-AI/litellm), then passed the approval gate with a second developer’s credential — so the official pipeline built, signed and published the malicious artifacts. PyPI removed the release by 13:23 UTC on October 21; Docker Hub and GHCR images tagged v1.107.0, latest, main-latest and main-stable were also affected for a window on October 21. Versions 1.106.2 and the release candidates are clean, and LiteLLM confirmed the incident with remediation steps.

That was the AI news cycle for October 23, 2026 — from open-weight inference engines to supply-chain attacks, the frontier keeps moving in every direction at once. Check back tomorrow for the next roundup.

Thursday’s AI headlines are led by DeepSeek on both ends of the spectrum: the Chinese lab open-sourced the production inference framework behind its flagship models on the same week its technology was banned from U.S. federal civilian networks. Elsewhere, the AI community is picking up the pieces after a supply-chain attack on a popular LLM gateway, Anthropic rolled out a proactive personal agent, and OpenAI ended years of cloud exclusivity by bringing its models to Google Cloud.

DeepSeek open-sources DeepInfer, the inference engine behind its production services

DeepSeek released DeepInfer, the high-performance inference framework that powers its large-scale production services, as open source — the top AI story on Hacker News on Thursday with more than 2,300 upvotes and nearly 900 comments. The framework serves DeepSeek-V4.1-Flash with optimized kernels drawn from FlashMLA, DeepGEMM, DeepEP and custom TileLang operations, and supports expert-parallel MoE inference with load balancing across NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs.

DeepInfer includes DSpark confidence-scheduled verification (speculative decoding with draft tokens), bounded-replay SWA prefilling, async scheduling for zero-overhead batching, batch-invariant results, deterministic sampling and structured outputs constrained to regex, JSON Schema or EBNF grammars. The release covers offline inference; the accompanying KV-cache system, prefill/decode disaggregation and replica load balancing remain part of DeepSeek’s closed service layer. Commenters called it one of the most substantive open-source inference releases of the year — a reminder that even as Western governments restrict DeepSeek models, DeepSeek’s engineering is increasingly embedded in the open AI infrastructure stack.

DHS orders DeepSeek removed from U.S. federal civilian agency IT systems

The Department of Homeland Security has ordered federal civilian agencies to remove DeepSeek products — and third-party services developed using DeepSeek implementations — from their IT systems, and to exclude the company and its affiliates from future acquisitions. The notice, published Monday in the Federal Register (document 2026-21447), extends to any software “developed using DeepSeek implementations,” a scope commenters flagged as potentially sweeping.

DHS said DeepSeek’s models pose “multiple threat vectors”: servers located in China, personal information collected at account creation, and the risk that data entered by federal employees or contractors could be transferred to Chinese authorities or used to develop “actionable intelligence on US persons.” The order, signed by Secretary Markwayne Mullin as far back as June, took effect this week; agencies may request relief within 60 days, and the department will review it annually. The move escalates earlier restrictions at Commerce and the Navy and mirrors bans in Australia and elsewhere — deepening the split between U.S. policy, which is walling off DeepSeek, and the open-source community, which embraced DeepInfer on the same day.

Compromised LiteLLM 1.107.0 on PyPI ships credential stealer and backdoor

Security researchers at SafeDep uncovered a malicious LiteLLM 1.107.0 release on PyPI that installed a credential stealer and backdoor (dubbed cook.pth) on affected systems. LiteLLM is one of the most widely deployed open-source LLM gateways — an OpenAI-compatible proxy used to route and manage model calls across many enterprises — making the compromise a supply-chain event of unusual reach.

BerriAI, LiteLLM’s maintainer, published an incident report and said the attack was “novel” and “has been contained”: attackers compromised GitHub credentials for three developers, and the private release pipeline built the poisoned wheels from attacker-modified source. The maintainers revoked all but one developer’s access and urged users to pin below 1.107.0 and audit dependencies. Commenters noted this is the second significant PyPI supply-chain incident targeting LiteLLM this year, and used it to call for wider adoption of CI signature checks, canary tokens and default install cooldowns — a reminder that the AI toolchain’s dependency graph is now squarely in attackers’ sights.

Anthropic debuts Claude Duet, a proactive agent powered by Opus 5.5

Anthropic introduced Claude Duet, a research-preview “proactive agent” available on Pro and Max plans in the U.S. Powered by Opus 5.5 and integrated with more than 6,000 connectors and plugins, Duet briefs users each morning, drafts next steps, quietly handles tasks of all sizes — from scheduling meetings to building dashboards — and pauses to ask for approval when decisions matter. It runs in the same thread as the user’s existing Claude memory, skills and connected apps.

The launch puts Anthropic squarely in the emerging “always-on assistant” wave alongside OpenAI’s Dot and similar agentic companions, and commenters were quick to debate the margins (background agents can ride low-priority inference batches), the UX of yet another waitlist, and the privacy stakes of granting an LLM access to calendars, mail and password managers. Early testers report the morning-brief workflow is the standout function — a quiet shift from chat to delegated, background work.

OpenAI models are coming to Google Cloud’s Model Garden

OpenAI and Google Cloud announced a partnership to bring OpenAI’s frontier models and Codex agentic workflows into Google Cloud’s Model Garden, letting enterprises buy them through the Google Cloud Marketplace — drawing down existing GCP commitments — with Google’s data-governance and access controls layered on top. Reports identify the first models as GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, joining a Model Garden of more than 200 offerings. The models will also land in Google Antigravity and the Gemini Agent, with general availability planned in the coming weeks after an enterprise private preview.

OpenAI president and co-founder Greg Brockman called Google Cloud “a leading platform for developers and businesses building AI agents,” and Google Cloud CEO Thomas Kurian said the goal is “a broad selection of models, including those from OpenAI, so they can choose the right model for any given job.” Notably, Microsoft remains OpenAI’s primary cloud partner — the deal breaks years of effectively exclusive distribution through Azure without replacing it, and signals that frontier-model access is becoming an every-cloud commodity.

That’s the AI landscape for Thursday, October 22, 2026 — from an open-source inference engine and a federal ban, to supply-chain security, proactive agents and the end of cloud-model exclusivity.

Wednesday’s AI news cycle was unusually concentrated: Chinese lab DeepSeek dominated for two opposite reasons — open-sourcing the inference engine behind its production services while being banned outright from U.S. federal civilian agencies — while Mistral staked out new territory in time-series forecasting and signed a landmark research partnership with CERN. Meta, meanwhile, fielded a fresh controversy over hidden instructions embedded in WhatsApp’s help pages. Here are the five stories that mattered most today.

DeepSeek Open-Sources DeepInfer, the Inference Engine Behind Its Production Services

The top story of the day, with more than 2,200 points and 850 comments on Hacker News, is the release of DeepInfer, the high-performance inference framework that powers DeepSeek’s own large-scale production services. The company published it under the Apache 2.0 license and built it primarily on vLLM, with additional inspiration from SGLang and TensorRT-LLM.

The release supports offline inference for DeepSeek-V4.1-Flash and bundles the pieces of a production-grade serving stack: optimized kernels (FlashMLA, DeepGEMM, DeepEP, TileKernels, DeepSelect and custom TileLang operators), MoE expert parallelism with load balancing, DSpark speculative decoding with confidence-scheduled verification, SWA bounded replay to cut decoder prefill work, asynchronous batch scheduling, batch-invariant inference, deterministic sampling, and structured outputs constrained to regex, JSON schema, or EBNF grammars. It targets NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs, with prefill benchmarks measured on four GPUs and decode on 64.

The release is notable both for its engineering depth and for what it signals: DeepSeek is now distributing the same framework that handles its at-scale workloads, a move that further cements the open-weight, open-infrastructure path it has followed throughout 2026.

DeepSeek Banned from U.S. Federal Civilian Agency Systems

On Monday, the U.S. Department of Homeland Security made public a Federal Register order (2026-21447) banning DeepSeek technology from federal civilian agency IT systems. The order — originally issued in late July — bars agencies including Treasury, State and Justice from procuring DeepSeek products and services and requires removal of existing deployments. Notably, it extends to third-party products and services developed using DeepSeek implementations, and it excludes the Department of Defense.

The order states that DeepSeek’s models “present multiple threat vectors”: its servers are located in China, and Chinese law permits government cyber and intelligence functions broad access to private company infrastructure. The public disclosure drew over 500 comments on Hacker News, where the debate split between national-security pragmatists and critics who argued the ban is as much competitive policy as security policy — several noted it effectively funnels federal AI budgets toward domestic labs such as OpenAI and xAI, while others worried open-weight bans could chill the broader open-model ecosystem.

Mistral Unveils Timestral, Its First Time-Series Foundation Models

Mistral announced Timestral, its first family of time-series foundation models, on October 15. The family spans four sizes from 5M to 209M parameters, and the largest model claims first place across all five metrics on the field’s three main public benchmarks — GIFT-Eval, TIME and fev-bench — using about two-thirds of the parameters of the previous state of the art.

Timestral delivers probabilistic, multivariate forecasts and supports both past-only and known-future covariates, such as production schedules, planned shutdowns and weather forecasts. The models are designed to run efficiently on CPUs, forecast zero-shot on unseen domains, and combine synthetic pre-training, covariate-aware attention and long-context learning; domain adaptation is available via fine-tuning on Mistral Forge. Results have been submitted to the GIFT-Eval and TIME leaderboards. For enterprises forecasting sensor telemetry, energy prices, supply chains or financial flows, the pitch is simple: foundation-model accuracy at a size small enough to deploy almost anywhere.

CERN and Mistral Partner to Build a “Large Physics Model”

CERN and Mistral signed a statement of intent on October 15 to establish a long-term strategic partnership on foundation models for scientific discovery, with an initial focus on high-energy physics. The centerpiece ambition: a single pretrained “Large Physics Model” — described as the first foundation model for collider physics — trained on billions of simulated proton-proton collisions at the Large Hadron Collider, and designed to support everything from reconstructing collisions and identifying particles to flagging unexpected signals and evaluating future detector designs.

CERN brings roughly one exabyte of stored LHC data — a volume set to grow tenfold with the High-Luminosity LHC expected around 2030 — plus what its CIO, Enrica Porcari, called “a unique environment in which artificial intelligence can be developed and tested with exceptional rigour.” Johannes Brandstetter, Mistral’s VP of AI for Engineering, said the pair have “the opportunity to push the frontiers of high-energy physics, developing AI capabilities that have the potential to lead to new scientific discoveries.” For Europe, the deal doubles as an industrial-policy statement: a home-grown AI leader pairing with the continent’s flagship science institution.

Hidden LLM Instructions Found in WhatsApp’s Help Pages

A reporting collaboration between redact.dev and WIRED uncovered hidden instruction text embedded in WhatsApp’s help pages — invisible to human readers but clearly aimed at the company’s AI support chatbots. The embedded instructions told the models not to recommend contacting data protection authorities, regulators or supervisory authorities; not to suggest WhatsApp can manually delete accounts on a user’s behalf; and not to make statements about WhatsApp’s legal liability.

After WIRED contacted Meta, the hidden text was removed from the pages, with WhatsApp spokesperson Joshua Breckman attributing it to a bug that has since been fixed: “These are instructions to our chatbot, not directly to users, and are intended to ensure accurate and relevant responses for WhatsApp users who choose to message WhatsApp Support with AI.” The story drew 724 points and more than 200 comments on Hacker News, where reactions ranged from sympathy — no company wants its bot dispensing legal advice — to accusations of “industrialized prompt injection” designed to steer users away from regulators. Notably, several commenters pointed out the deeper significance: the instructions were publicly visible, a transparency trade-off that invites scrutiny of how companies constrain even their customer-facing AI.

That’s the AI news for Wednesday, October 21, 2026 — a day when open-source infrastructure, government policy, scientific ambition and corporate transparency all shared the spotlight.