Machine Learning Research Intern

Product updates: Sandbox Sidecars, new models, a refreshed dashboard, and more. August brought new model support, faster Function calls, and major improvements across Modal Sandboxes. Here are the highlights. Day-zero support for Kimi K3, Qwen 3.8, GLM 5.3, and GLM 5.3 Flash. Explore four frontier open-weight models now supported on Modal Auto Endpoints. All four are supported as Shared Endpoints with token-based pricing, or as Dedicated Endpoints billed by the GPU second. * Kimi K3: Moonshot's 2.8T-parameter multimodal model with native vision and a 1M-token context window. Modal partnered with Moonshot and vLLM on day-zero support, with a custom DFlash speculator delivering 460 tokens per second on agentic workloads. * Qwen3.8-2.4T-A95B: Qwen's 2.4T-parameter, text-only Mixture-of-Experts model activates 95B parameters per token and is built for coding, research, and long-horizon agentic tasks. Modal delivered day-zero support with SGLang and a custom DFlash speculator. * GLM-5.3: Z.ai's 753B-parameter Mixture-of-Experts model activates 40B parameters per token and supports a 1M-token context window. It is built for complex coding and long-horizon agentic tasks. * GLM-5.3-Flash: Z.ai's natively multimodal 320B-parameter Mixture-of-Experts model activates 18B parameters per token and supports a 1M-token context window, bringing visual understanding to coding and agentic workflows. Improving control over spend. Two updates for better usage control: * Environment-level budgets are now available for all Team and Enterprise workspaces. Owners and Managers can set a compute-usage budget per Environment and track current-cycle compute usage against it - useful for giving teams their own guardrails without splitting workspaces. Read the docs. * The usage limits UI got a refresh. Threshold billing and usage limits are now presented with consistent terminology and numbers that add up at a glance, so it's clearer what your limits are and how to adjust them. Lower pricing for broad region selection. The pricing multiplier for broad region selection has dropped from 1.5x to 1.15x. Broad regions - including the US, EU, and APAC - keep workloads closer to users or data while preserving a larger capacity pool for better availability and faster scheduling. Private Environments with default roles. You can now set a default Role for a Restricted Environment. Set the default to No Access and explicitly grant Viewer or Contributor access to the Members who need it; everyone else can't discover or access the Environment at all. Workspace Owners and Managers retain Contributor access. This effectively gives you private Environments for sensitive projects within a shared workspace. A refreshed dashboard has landed. Modal has updated the dashboard with improved navigation, a new color system for visualizations, and a light mode appearance option. Light mode is available in beta for all users, with other changes rolling out gradually over the coming days. Sandbox Sidecars enter public alpha. Sandbox Sidecars are now in public alpha. Sidecars run additional containers beside a main Sandbox on the same host, connected through a low-latency internal network. They are useful for custom network proxies, multi-container workloads, and separating an agent harness from its tool execution. import modal app = modal.App.lookup("sidecar-example", create_if_missing=True) image = modal.Image.debian_slim.build(app) sb = modal.Sandbox.create("sleep", "600", app=app, image=image, timeout=300) sidecar = sb._experimental_sidecars.create( "python", "-m", "http.server", "8080", name="web", image=image,) # Give the server a moment to start, then call it from the main sandbox. p = sb.exec( "python", "-c", "import time, urllib.request; time.sleep(1); print(urllib.request.urlopen('http://web:8080').status)",) p.wait print(p.stdout.read # "200" sb.terminate Better observability for Sandbox CPU and memory requests. Modal has shipped a set of graphs to make it easier to specify Sandbox CPU and memory requests, helping you avoid OOMs or CPU contention without paying for more resources than you need. These charts show various percentiles of how much CPU and memory the Sandboxes in each of your Apps consume, so you can set request sizes accordingly. VM Sandboxes now in public beta. VM Sandboxes have entered public beta. They run a real Linux kernel, so they can support workloads that run Docker containers within the Sandbox, or that require Linux features like eBPF, systemd, cgroups, or custom filesystem mounts. This generally also leads to better filesystem performance for I/O-sensitive workloads. Now GA: Sandbox directory snapshots and new Sandbox filesystem API. Sandbox directory snapshots let you snapshot selected directories instead of an entire container filesystem. You can use them to update system dependencies separately from application code, speed up resumptions of previous sessions, and restore application state into a warm pool of Sandboxes. Read the docs for more. Additionally, its new Sandbox filesystem API is now GA. The new filesystem API offers improved reliability and better performance for small files. Learn more. Regional proxies. You can now choose a region when creating a Modal Proxy in workspace settings. Proxies provide static outbound IPs for Functions and Sandboxes, letting you connect to resources protected by IP allowlists. Placing a Proxy closer to those resources gives you more control over where traffic is routed and can reduce network latency. SDK updates. Modal has released version 1.5.5 of the Python SDK and version 0.10.0 of the JavaScript and Go SDKs. Updates across these releases include new log APIs, faster Sandbox filesystem writes in JS and Go, and improved autoscaler configuration reporting. Scaling to 1 million concurrent Sandboxes in seconds. Modal rebuilt its scheduling system to support millions of concurrent Sandboxes and tens of thousands of creations per second. In testing, Modal started one million Sandboxes in under a minute while keeping median time to interactivity below half a second. Bringing serverless functions closer to the speed of wire. Modal rebuilt the Function I/O path around a faster, geographically distributed routing layer, moving non-critical work off the hot path and minimizing shared-storage access. The result is ~80ms lower p50 latency, with additional savings possible. Kimi K3 by Moonshot available on Modal. Modal partnered with Moonshot and vLLM to bring day-zero Kimi K3 support to Modal, pairing the 2.8T-parameter multimodal model with a custom-trained DFlash speculator. On agentic workloads, its speculator increases interactivity from 100 to 460 tokens per second and per-GPU throughput from 800,000 to 1.5 million tokens per minute. How Botika runs full-stack generative AI on Modal. Modal supports Botika's entire AI stack, from processing a 100-terabyte image dataset to training foundation models and serving roughly 15 models in production. With Modal, its infrastructure can absorb a two- to threefold traffic increase in seconds without manual intervention. Run Devin Outposts and Cursor Cloud agents on Modal. You can now run Devin Outposts and Cursor Cloud Agent workers inside customizable Modal Sandboxes. Bring your own images, dependencies, secrets, and CPU or GPU resources while Modal handle isolated environments, fast startup, and elastic capacity for every session. Runtime: a conference by Modal. Join Modal October 1st at The Midway in San Francisco for Runtime, its conference for engineers running AI in production. Modal'll have three tracks covering inference, training, and agents, with technical deep dives and speakers including Scott Wu (Cognition), Bryan Catanzaro (NVIDIA), and Cat Wu (Anthropic). Ship your first app in minutes. $30 / month free compute

Modal is expanding in Europe with its new London office. CEO and Co-Founder Modal is very excited to announce that Modal is opening a London office. It's its second in Europe, and the beginning of a much larger investment in the region. Europe has incredible talent, amazing startups, and tons of potential for Modal. Modal has actually been in Europe for a while! Modal has actually been transatlantic for a long time. I grew up in Stockholm, Sweden, went to school there, started my career at Spotify, and moved to the US relatively late in life. At Spotify, I built the recommendation system, used by hundreds of millions of users every day. I also saw Spotify growing from an obscure startup in Stockholm to a massive global public company. Since the very early days of Modal, Modal had engineers in Stockholm. Over the years, the team has snowballed to almost 20 people. They build some very core foundational parts of Modal, including its custom file system and big parts of the Sandboxes product. Modal is getting even more Europe-pilled today. Modal is now expanding to London as well! Hugh Killingbeck-Jones has joined Modal as VP EMEA to build out its GTM teams in London, covering startups and enterprises. Modal think London is the AI capital of Europe, and over time this office will grow to include engineering and other functions too. Modal has seen the new generation of European AI companies take on the hardest problems in their industries across robotics, biotech, health, legal, fintech and build for a global market from the outset. There are so many incredibly ambitious companies, and Modal think many of them will be using Modal for their infrastructure. Why Modal? Because Modal help these companies scale fast and take over the world. Whether they need large scale custom model inference, sandboxes, fine-tuning, training, batch jobs, or many other things. Modal makes it possible to move super fast and not have to think about infrastructure - Modal give them the primitives they need to build applications, fully flexible resource consumption, amazing developer experience, and much more. European companies will sometimes have a different set of requirements than US ones, especially when it comes to local regulation. Modal already support a lot of this, and will be investing more. The most ambitious teams run on Modal. European AI-native companies like Black Forest Labs and Legora build on Modal, and so do global platforms like DoorDash. Black Forest Labs benchmarks frontier multimodal models across GPU generations. Legora builds agents for the hardest legal work and tests them at scale. DoorDash runs its AI platform across markets. "We're building frontier visual intelligence, and the sooner we can see and evaluate a new checkpoint, the sooner we can adjust recipes and training. Modal lets us pull in a checkpoint and fan sampling out across as many containers as we need, so eval wall time is minimized and training continues uninterrupted." - Ruben Rodriguez Buchillon, Member of Technical Staff, Black Forest Labs "Modal sandboxes give every eval its own isolated environment. We can run thousands of them, trust the results and ship improvements to Legora faster." - Ebba Helfer, Member of Technical Staff, Legora Modal'd love to meet other Europeans! In London, Modal is hiring across sales, solutions engineering, partnerships, marketing and more. Modal is also continuing to grow its engineering team in Stockholm, with roles open for forward-deployed and system engineers. In the near future, Modal is planning meetups across EMEA (London, Paris, Berlin, Tel Aviv and more), and you'll find Modal at conferences including: * AI Engineer Paris * NVIDIA GTC Berlin * Slush in Helsinki If you're building in Europe, you can be running on Modal in minutes. Sign up here. And if you'd rather spend your days helping teams here get models into production, Modal'd love to hear from you. Ship your first app in minutes. $30 / month free compute