DeepSeek Open-Sources Its DeepInfer Serving Stack as Amazon Opens Trainium to PyTorch and Microsoft Expands Sovereign Cloud

DeepSeek Open-Sources DeepInfer, the Stack That Serves V4.1-Flash

DeepSeek has published the software it uses to run its own models in production. In DeepSeek’s official post introducing DeepInfer on X, published at 2:30 AM PT on Tuesday, October 20, the lab calls DeepInfer its in-house inference framework, says it is now fully open source, and says it powers DeepSeek-V4.1-Flash at scale with high throughput and low latency. The stated goal is to help the community lower compute costs and reduce energy use. This follows the V4.1-Flash refresh we covered on October 18, but the release is new: this time DeepSeek is sharing the serving code, not another set of weights.

"DeepInfer is the high-performance inference framework that powers DeepSeek's large-scale production services." — DeepSeek

The DeepInfer repository and README on GitHub spell out what is and is not included. The release supports one model, DeepSeek-V4.1-Flash, and defaults to the DeepSeek-V4.1-Flash-1017 checkpoint. It bundles DeepSeek’s own kernels, including FlashMLA, DeepGEMM and DeepEP, along with expert parallelism for its mixture-of-experts design, draft-token decoding with confidence-scheduled verification, an asynchronous batch scheduler, and structured outputs. Batch-invariant inference and deterministic sampling are available on NVIDIA hardware only. Supported hardware is NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. DeepSeek is clear that this is an offline inference release: a KV cache system, prefill and decode disaggregation, and load balancing across data-parallel replicas are all needed for a complete service and are not included. The code is under the Apache License 2.0, and DeepSeek credits vLLM as the original base, with inspiration from SGLang and TensorRT-LLM. When we fetched the page, the repository showed about 7,100 stars.

Amazon’s TorchNeuron Lets PyTorch Code Run on Trainium Without a Rewrite

Amazon is going after the software moat around Nvidia. Amazon’s announcement of TorchNeuron for AWS Trainium, posted Monday, October 19, says existing PyTorch code now runs natively on its Trainium chips with a single configuration change and no rewrite. TorchNeuron plugs into PyTorch’s built-in mechanism for registering new hardware, so standard testing, debugging and scaling tools work unchanged. TorchNeuron supports Trainium2 and Trainium3 and is tested against hundreds of thousands of models on every build. Amazon says it will be released under the Apache 2.0 license at PyTorch Conference North America on October 20. Hugging Face has also enabled native Trainium support in its tools, and early users include Cartesia and the Lieber Institute for Brain Development.

"The goal is to make PyTorch more hardware-agnostic for the entire developer community, not just Trainium users." — Amazon

TechCrunch’s analysis of TorchNeuron as Amazon’s biggest shot at Nvidia frames the launch as a direct challenge to CUDA’s deep integration with PyTorch, and notes that PyTorch users can already port to AMD GPUs or Apple chips with a configuration change. Colin Brace of AWS Annapurna Labs told TechCrunch that Trainium pairs price-performance with one platform for training and inference, citing Deepgram, which he said processes 70% more audio per dollar after moving to Trainium with native PyTorch support.

Microsoft Expands Sovereign Cloud With Local AI, Key Control and EU Access Oversight

Microsoft used a Paris event to widen its sovereignty offering. Satya Nadella’s post announcing the Paris Sovereign Cloud news describes new choice and control for organizations running their most sensitive workloads. The details are in Microsoft’s October 20 blog post on advancing digital sovereignty by Douglas Phillips, president and CTO of Microsoft Specialized Clouds. On the public cloud side, a Sovereign Control Panel enters public preview at Microsoft Ignite. For EU and EFTA customers, Data Guardian adds European-resident oversight by a trusted EU partner of qualifying remote access by Microsoft engineers, and Production Touch Ledger keeps tamper-evident records of that access. External key management with Azure Managed HSM is now generally available, letting customers hold the root encryption key in hardware they own.

"Digital sovereignty should give organizations greater control and security over the technology they depend on, while continuing to innovate." — Douglas Phillips, Microsoft

The bigger AI story is on Azure Local, Microsoft’s customer-operated private cloud. Foundry Local becomes generally available at Ignite for connected and disconnected operations, with a catalog that includes Microsoft, partner, open-weight and open-source models, plus local inference and agentic retrieval over enterprise data. GitHub Enterprise Local also reaches general availability at Ignite, and developers can use GitHub Copilot with locally hosted models. Confidential virtual machines on Azure Local are generally available, the ability to move workloads from connected to disconnected operation enters public preview, and Microsoft highlighted National Partner Clouds such as Bleu in France.

OpenAI Says Its Newest Models Can Now Migrate COBOL While Preserving Behavior

OpenAI is pitching its latest models at the oldest code in the enterprise. In OpenAI’s research post on coding in low-resource programming languages, dated October 19, the company says COBOL supports more than 40% of online banking systems, 80% of in-person credit card transactions and 95% of ATM transactions. It built two evaluations. MultiLangBench asks a model to build a SQLite-compatible database engine in each of 20 languages; at maximum reasoning effort GPT-6 Astra averages an 87.4% test pass rate, up from 26.8% for GPT-5.6 Sol, with COBOL rising from 10.9% to 72.1%, Fortran from 21.4% to 83.5% and Visual Basic from 27.5% to 89.4%. OpenAI is releasing the MultiLangBench setup and test suite on GitHub.

The second evaluation migrates COBOL programs to C++, Java, Kotlin, Python or Rust and scores them on hidden tests that demand byte-for-byte parity on observable outputs. OpenAI notes that COBOL’s fixed-point arithmetic discards excess decimal digits unless rounding is requested, so a rewrite that rounds can run cleanly and still produce different financial results. GPT-6.1 Sol scores 86.6% at low reasoning effort, against 5.4% for GPT-5.6 Sol at the same setting and at 61% of the API cost, and 97.9% at higher effort. OpenAI’s official post on legacy-language coding on X pitched the work at banks, governments and businesses, and Engadget’s report on OpenAI’s COBOL migration claims highlights the company’s own caveat that production use still needs expert review and testing against business rules and dependencies.

"A migration is considered successful only if all tests pass." — OpenAI

Tenstorrent Raises $1.5 Billion to Scale Open-Weight Inference Servers

The money keeps flowing to Nvidia challengers. Tenstorrent’s Series E announcement, dated October 19, says the company closed $1.5 billion at a $6.2 billion post-money valuation in a round led by Trace Capital and Innovation X, with participation from investors including ARK Invest, Fidelity, Morgan Stanley and Tether Investments. Tenstorrent will use the capital to deploy its Galaxy servers at scale for open-weight LLMs and video generation, and to advance the AI and CPU intellectual property it licenses to chip designers.

"The future of compute is open, scalable, and owned by the people building on it." — Jim Keller, Tenstorrent

TechCrunch’s report on Tenstorrent’s $6.2 billion valuation says the round more than doubled the $2.6 billion valuation from late 2024 and comes five months after Intel and Qualcomm reportedly showed acquisition interest at valuations of up to $5 billion. CEO Jim Keller told Bloomberg, according to TechCrunch, that the company chose to stay independent and raise more after showing that its Galaxy servers could run DeepSeek models.

Dream Labs Emerges With a Robot Model That Controls Unseen Machines

Robotics got its own API launch. Dream Labs’ research post introducing Dream-1 describes a World Action Model that, given a new two-armed robot with parallel grippers, can start performing tasks through the company’s API with no fine-tuning and no demonstrations from that robot. Pricing is $3 an hour for Dream-1 and $2 an hour for Dream-1-flash; each call returns 1.6 seconds of actions. The team says it has tested zero-shot control on robots ranging from Franka arms to the 18-year-old PR2.

"It will make mistakes, and it has not been validated for safety-critical use." — Dream Labs

Dream Labs’ Business Wire press release on its $90 million seed round, via The National Law Review says the round was led by B Capital along with Kindred Ventures. Founded in 2026 by former leads of NVIDIA’s GR00T robotics research, the company works with more than 20 hardware partners, supports robots that cost roughly $2,000 to $10,000, and will spend the money mainly on compute. Dream-1 currently handles tabletop tasks such as pick-and-place and folding laundry. SiliconANGLE’s profile of Dream Labs quotes co-founder and CEO Joel Jang: “The hardware is ready, but what’s been missing is a brain any robot can use.”