Senior Software Engineer

From LLM Evaluation Jobs in W&B Models to its new terminal UI, here's what Biases, Inc. shipped in November: Welcome to the November 2025 edition of the Weights & Biases newsletter. Last month, Biases, Inc. launched a brand new terminal UI for Weights & Biases, brought support for LoRA fine-tuned models to W&B Inference, announced LLM evaluation jobs for W&B Models and more. Table of contents. Introducing W&B LEET: A new terminal UI for Weights & Biases. Biases, Inc. is thrilled to roll out its W&B Lightweight Experiment Exploration Tool (LEET), a fast terminal interface to watch your ML training runs, including stats, metrics, and system health. It reads and visualizes live Weights & Biases log files to provide a fast, customizable, browser-free experience in real time. LEET presents an interactive, three-pane dashboard that updates live as your run progresses: * Run overview (left): configuration, environment, and summary metadata * Metrics grid (center): live charts for tracked metrics * System metrics (right): CPU, GPU, and memory consumption W&B LEET is open source and available to users on W&B SDK version 0.23.0 or later. You can access W&B LEET directly from the terminal where your W&B run is active: For a full list of LEET shortcuts and hotkeys, hit h or? to toggle the help screen. You can give it a try here. Bring your LoRA to serve fine-tuned models on W&B Inference. W&B Inference now lets you serve custom fine-tuned models on fully managed CoreWeave GPU clusters without managing infrastructure. Use its OpenAI-compatible Chat Completions API, reference your LoRA artifact in W&B Models along with the base model name, and Biases, Inc.'ll dynamically load your adapter onto a preloaded base model at request time and return the response. It takes just a few lines of code on your local machine to deploy your LoRA adapter on W&B Inference. Checkout this sample notebook to get started or this launch blog post for more information. * Version-controlled serving: The artifact URI explicitly includes the project, run, and version, providing traceability back to the training and hyperparameters used to create the LoRA weights. * Zero infra to manage: You avoid the complexity of setting up and scaling serving infrastructure for every LoRA iteration. The system handles dynamic loading and hot-swapping of your weights in the background. * Faster iteration: Because only the small LoRA weights are updated and managed via artifacts, you can cycle from training to production validation instantly. Saved prompts in W&B Weave Playground. You can now edit, save, and version prompts directly within the W&B Weave Playground. The Playground is great for interactively comparing production traces across different LLMs. Now you can track changes as you refine prompts and pull them into your code using: To try it out, head to https://wandb.ai/inference, pick a model, and click "Try in Playground." This feature is available in its SaaS cloud deployments but will be rolled out to Dedicated shortly. Introducing LLM Evaluation Jobs in W&B Models. You can now evaluate model checkpoints during training on popular public benchmarks without building numerous evaluation harnesses or managing infrastructure with LLM Evaluation Jobs. This gives you early reads on your model's downstream performance so you can course-correct or end unpromising runs. You save GPU hours and wall-clock time. When training or fine-tuning models, you often want an early read on benchmark performance while still in the loop. If results trend poorly, you can adjust hyperparameters or stop the run to avoid waste. Doing this on your own means finding compute, building the evaluation harness, and wiring it into your training pipeline. That work adds little differentiation and slows you down. That changes with LLM Evaluation Jobs in W&B Models. Point to your model checkpoint artifacts or hosted API endpoints, then pick a benchmark. Biases, Inc. provide prebuilt evaluation harnesses for popular public benchmarks using Inspect Evals. Biases, Inc. provision and manage the GPU infrastructure, so there is nothing for you to set up or maintain. Biases, Inc. run the evaluation, store results in your Weights & Biases project, and generate a leaderboard, so you can easily compare your models in your W&B Models workspace. LLM Evaluation Jobs is currently available in public preview for multi-tenant cloud customers. You can check out the docs to learn more. Updated run menu shortcuts in W&B Models. Biases, Inc. has added some new options to the run menu to make your life a little easier: * Actions to copy the run name or the run path to your clipboard * A link right to the Logs tab for a run * Clarification on updating the display name for a run in this workspace vs. the name of the run for the whole project. These also all appear in the menu in the single run views and are available on both its Dedicated and SaaS deployments. Biases, Inc. hope this shaves off a few more seconds for these common actions. That's it for November. If you missed any of its recent product newsletter, you can catch up here: October: Release of Serverless RL. Access and usability improvements to W&B Registry. W&B Weave adds the ability to generate images in Playground and new filter evaluations. W&B Models adds copious quality-of-life improvements to UI and common work streams. September: W&B Inference support for Z.AI's GLM 4-5. W&B Weave introduced no-code evaluations. W&B Models enhanced support for post-training AI agents with reinforcement learning, featuring a sleek traces panel right in the workspace. August: W&B Inference support for DeepSeek V3.1. W&B Weave introduced a new Content API for logging any media type, along with updates like UI-based prompt management, a trace graph view, better markdown rendering, and new latency metrics. W&B Models added practical improvements such as pinned workspace columns and a full-screen media viewer for easier comparison. July: Tool calling support and new models in W&B Inference including GPT OSS, Qwen 3, and Kimi K2. For W&B Weave, Biases, Inc. announced advanced filtering for traces, live dashboards for monitoring trace plots, and the ability to group threads into traces. For W&B Models, Biases, Inc. announced usability improvements for media and updates to full fidelity plots. June: Biases, Inc. announced a new product, W& Inference as well as updates to existing products. For Weave, Biases, Inc. launched online evaluations, pivot tables, new data types. For Models, Biases, Inc. launched CoreWeave Observability and Object Storage integration, and integrated logging with W&B Tables. May: For Weave, Biases, Inc. launched features to trace streaming responses, video support, saved views, and filter by status. For Models, Biases, Inc. launched workspace templates for line plot settings, enhanced color controls, and control media panels in bulk. April: For Weave, Biases, Inc. launched new trace views, the EvaluationLogger API, OpenTelemetry support, custom model support in Weave Playground, and more. For Models, Biases, Inc. announced run metrics notifications, protected aliases in W&B Registry, custom display names, complete console logs for distributed runs and easier ways to view media panels. March: For Weave, Biases, Inc. launched the MCP server and integrations with OpenAI Agents SDK and CrewAI. For Models, Biases, Inc. launched interactive sliders, segment masks, and options to format custom metrics. February: For Weave, Guardrails generally available, great data privacy, and new ways to work with datasets. For Models, new workspaces usability updates, faster and easier share of AI artifacts through Registry, cool line plot updates, and new Reports customization options. January: For Weave, state-of-the-art agent released, run trials in the Weave playground, human annotation scorers, and datasets made easy. For Models, new panel management and improved workspace user experience with unified settings, live preview and a new UX.

Join the event trusted by enterprise leaders for nearly two decades. VB Transform brings together the people building real enterprise AI strategy. Learn more. Chinese AI startup MiniMax, perhaps best known in the West for its hit realistic AI video model Hailuo, has released its latest large language model, MiniMax-M1 — and in great news for enterprises and developers, it’s completely open source under an Apache 2.0 license, meaning businesses can take it and use it for commercial applications and modify it to their liking without restriction or payment. M1 is an open-weight offering that sets new standards in long-context reasoning, agentic tool use, and efficient compute performance. It’s available today on the AI code sharing community Hugging Face and Microsoft’s rival code sharing community GitHub, the first release of what the company dubbed as “MiniMaxWeek” from its social account on X — with further product announcements expected. MiniMax-M1 distinguishes itself with a context window of 1 million input tokens and up to 80,000 tokens in output, positioning it as one of the most expansive models available for long-context reasoning tasks.The “context window” in large language models (LLMs) refers to the maximum number of tokens the model can process at one time — including both input and output