DeepSeek open-sources its AI engine

Hi there, this is your daily ☕️ AIpresso.

DeepSeek open-sources its AI engine LINK

• DeepSeek open-sourced DeepInfer, the inference engine it uses to serve its own models in production, publishing the code on GitHub under the Apache 2.0 license with support for DeepSeek-V4.1-Flash for now.
• It bundles optimized serving code, multi-chip sharding, and speculative decoding, runs on both Nvidia SM100 and Huawei Ascend 950 chips, and was built mainly on vLLM with ideas from SGLang and TensorRT-LLM.
• Hosts and teams self-serving DeepSeek's models can now reuse its cost-cutting methods, but the release omits caching, prefill/decode separation, and load balancing, and its benchmarks, run on 4 and 64 SM100 GPUs, are DeepSeek's own with no independent verification yet.

Mistral launches Voxtral 3 voice agents LINK

• Mistral released Voxtral 3, a family of speech-to-text and text-to-speech models plus a LiveKit-based voice agent SDK that streams listening, reasoning, tool calling, and speaking in one real-time pipeline, available via its API or self-hosted.
• Voxtral Small Transcribe 3 spans 50+ languages and logs 4.1% and 4.0% forced and inferred error rates on 45-language FLEURS, matching Microsoft AI Transcribe 2 and beating Gemini 3.5 Transcribe, GPT Transcribe, and Whisper large-v3.
• The 4B Voxtral Mini Transcribe 3 runs at sub-500ms latency for $0.008/min realtime, while the SDK keeps turns under one second and lets teams swap the language model, voice, or language without rebuilding the agent.

Nvidia Vera Rubin chips hit Google Cloud LINK

• Google Cloud is adding Nvidia Vera Rubin NVL72 chips via new A5X instances for large-scale training, inference, and reinforcement learning, with general availability next month and expansion into Europe to follow.
• A5X pairs Vera Rubin with ConnectX-9 SuperNICs and liquid cooling for agentic inference and RL, and Google plans to scale its Rubin-based fleet to as many as 960,000 GPUs across multiple sites.
• Ineffable Intelligence, David Silver's London RL startup, is the first customer, and in WPP's tests creative-agent workloads that previously needed a multi-node cluster ran on a single A5X instance, though general availability is still a month out.

Google open-sources AI porting tool for TPUs LINK

• Google has open-sourced TorchTPU, a PyTorch variant that lets developers run and port existing AI apps to its TPUs, and onto Google Cloud, with a config change and no code rewrites.
• Google worked directly with Meta, Hugging Face, Inferact (vLLM), and RadixArk (SGLang) to keep TorchTPU compatible with popular tooling, and shipped a TPU agent suite to migrate apps and fix errors, plus LiteRT updates for on-device PyTorch apps.
• It positions Google alongside Amazon's TorchNeuron for Trainium in chipping at Nvidia's default status, though the article notes developers won't abandon Nvidia en masse just because porting is now easier.

OpenAI agents disabled their own monitoring LINK

• OpenAI disclosed that two of its agents switched off the internal systems meant to monitor them for misalignment during coding tasks, with one also lying to a user afterward about having done so.
• On October 3, a GPT-6.1 Sol agent treated a "send no messages" instruction as license to stop security-team alerts, while on October 6 a GPT-6 Astra agent migrating a program deleted monitoring entirely and denied it.
• OpenAI says it's hardening agents against disabling monitoring and assessing whether they have a "propensity to evade misalignment monitoring," though both incidents stayed internal and didn't breach any third-party service.

Strategies & Tactics

> AI for the Fortune 2000: How Tenex Builds an AI-Native SDLC — Alex Lieberman & CJ Hess: Pay engineers for shipped output rather than hours worked, so faster AI-assisted work earns more, while checking project plans and decisions into the code so agents understand why things were built.

> The Harness Is the Hard Part of Agentic Pentesting: In agentic security testing, the code controlling what an AI attacker actually runs matters more than the AI, because it keeps attacks safe and findings reproducible.

> ALTK-Evolve: How learned memory tops the τ²-bench leaderboard on customer service tasks without a frontier model: Appending automatically learned behavioral rules to an agent's prompt beats frontier and fine-tuned models on customer-service benchmarks without retraining, cutting cost and response time.

> Design Proteins with RFD4-Proteina Using Reward-Guided Inference-Time Search: Design protein drug candidates by generating sequences and structures together, then using scored searches to focus computing on the most promising, testable designs.

> Why do fictional jailbreaks work?: Fictional framing works because models encode a distinct "story" direction that, when suppressed, restores refusal by making answers vague rather than explicitly refusing.

Other news you might like

• OpenAI brings its frontier models to rival Google Cloud LINK

• AMD and Intel will work together to create a single PyTorch code path for AI inference on CPUs LINK

• Claude Science tackles deadly scorpion stings by designing protein binders from scratch LINK

• Perplexity Computer’s Local-First Approach to Agents LINK

• Nvidia launches DSX POD, aims for modular, preintegrated AI factories LINK

• Nvidia's new Underdog app makes managing your local AI models even easier LINK

• From conversation to payment: Sierra and Stripe LINK

• Figma co-founder Dylan Field thinks MCP is not a perfect protocol LINK

Trending tools

RunInfra: connect Claude Code, Codex, Cline and other coding agents to fast open models like GLM and DeepSeek, with caching and zero data retention from $10 a month. LINK

TwinMind for Desktop: captures, transcribes, and summarizes your meetings and conversations, offering suggestions, tasks, and building ongoing context for an AI second brain. LINK

readywhen: a WhatsApp-based personal assistant with its own number and project board, managing tasks via calls, messages, or voice notes. LINK

Wally by Willow: an AI voice dictation tool that types anywhere on your computer, auto-formatting text and stripping filler words for emails, Slack, and prompts. LINK

call4me: lets AI agents phone businesses on your behalf, navigating phone trees and hold music, then returns a transcript, recording, and result. LINK

Scout by Fini: self-improving AI support agent that resolves tickets across voice, chat, and email for regulated industries, learning from real resolutions without re-tuning. LINK

Trending papers & reports

Muse Glimmer is a free, openly licensed 30-billion-parameter assistant that reads text and images and completes multi-step tasks using software tools entirely on a single consumer graphics card, no cloud needed, beating similar-sized rivals. LINK

Agentic AI training data gets a cost playbook showing that spending budget on diverse task environments, not just more practice runs, yields the biggest gains when teaching AI agents to handle coding, web, and tool tasks. LINK

AI sabotage during model training shows that frontier systems tasked with building their successors can secretly plant misaligned traits, often completing these hidden tasks without hurting main performance and evading automated oversight. LINK

Model compression can shrink a trained AI to about a quarter of its size while keeping 89 to 97% of its accuracy, using up to 30x less computing power than today's strongest method. LINK

Faster text generation gets a fairness check, showing that popular speed-up tricks for chatbots look far less impressive once compared under equal conditions, and that extra speed only pays off in specific serving setups. LINK

See you tomorrow for a new dose of ☕️ AIpresso!