Modal
AI overview
Sign in and the AI will write an overview from our coverage.
Headlines · 11
- Modal details the fast filesystem stack behind its VM Sandboxes
Modal published a technical post on how it evolved the filesystem stack underpinning its VM Sandboxes. Its first approach routed guest VMs through virtio-fs on top of ImageFS, inheriting lazy loading and sub-second cold starts but making metadata operations very slow: a recursive chown went from four seconds to over a minute, pnpm install and bun install ran about three times slower, and ripgrep performance dropped sharply. Modal then moved to virtio-blk with ext4 inside the guest plus a new host daemon, ImageFS-blk, which writes ext4 metadata itself and serves blocks on demand; the post says a chown fell from about 99 microseconds over virtio-fs to about 2.1 microseconds.
Modal Blog · 🔥 8 - Meta releases Muse Spark Open, an 816B open-weight multimodal coding model
Meta released Muse Spark Open, an open-weight mixture-of-experts model with roughly 816 billion total parameters and about 42 billion active per token, accepting text, images, audio and video with a context window of up to one million tokens. Meta says it is built for agentic coding and long-horizon tasks and calls it “the most capable US open-weight model to date,” adding that its own benchmarks show it beating GLM-5.3 and Kimi K3 on agentic and visual-reasoning tasks. Weights are available on Hugging Face, with day-one hosting from Together AI and Modal, while Bloomberg reports Meta kept its most advanced models closed.
Hugging Face · 🔥 136 - Factory launches OpenCompute, a self-hosted agent sandbox orchestrator
On October 14, 2026, Factory introduced OpenCompute, a self-hosted orchestrator for agent sandboxes now in private preview that runs on the customer's choice of runtime, including Kubernetes, AWS Lambda MicroVMs, EC2 and in-house platforms. It connects to the Factory backend over an outbound, TLS-secured WebSocket with no inbound firewall ports required; the Go control plane records desired and observed state in a Postgres database inside the customer's environment, uses a declarative reconciler, and splits each provider into a standalone gRPC driver with a required Core conformance profile plus optional Suspend, Snapshot and Fork profiles. Factory says its first customer went from prototype to a first sandbox running in its own data center in six weeks, and the company plans to open-source OpenCompute, including its contracts, conformance tests and driver model.
Factory News · 🔥 0 - Modal raises $750M Series D at $15.75B valuation, tops $1B ARR
AI infrastructure company Modal said it raised a $750 million Series D at a $15.75 billion post-money valuation, led by Accel with Kleiner Perkins joining and existing investors including Menlo Ventures, General Catalyst, Redpoint, Lux Capital and Bain Capital Ventures participating. The company said it has crossed $1 billion in annualized run rate, that usage has tripled in the five months since its Series C and grown more than 15x year over year, and that its Sandboxes product alone runs at roughly a $300 million annual rate with close to 3 billion sandboxes started since 2023. It lists customers such as DoorDash, Ramp, Linear, Cognition and Runway, alongside robotics firms Physical Intelligence and 1X training robots on its platform.
Modal Blog · 🔥 3 - Theory 2026 AI survey: security tops pain points, Cloudflare leads agent sandboxes
Theory published the second annual edition of its AI in Practice Survey, fielded to 455 senior technical builders on 22 September 2026. Security and governance rose from sixth to first among pain points in twelve months: 43% rate it the most important pain point, with evals second at 22%; only 10% say security tooling delivers the most value today, but 18% expect it to lead in the future, the largest expected gain of any area. Routing is now standard practice, with 45% of teams routing queries across models — 23.7% on a router they trained and 21.5% on a proprietary one — while 54.7% still send every query to a single model. Among the 259 teams whose agents write and execute code, 62% host their sandboxes on Cloudflare, over three times the closest option (Other and Vercel at 17% each), followed by Modal 16%, E2B 12%, Daytona 9% and Sail 7%. Agent performance analysis is led by business intelligence tools at 56%, with coding agents at 48%, dedicated observability tools at 41% and spreadsheets at 30%.
Tomasz Tunguz · 🔥 0 - Modal: mixed-precision, mixed-accelerator samplers enable stable async RL
In a technical blog post, Modal reports that running reinforcement-learning samplers on mixed precisions and mixed accelerators can keep large-model RL stable while cutting cost. Six kinds of accelerators feed a single BF16 trainer on B200: B200 and B300 serve NVFP4 weights, H100 and H200 serve FP8 and BF16 weights, and A100 and RTX PRO 6000 serve BF16 weights, each running its own kernels. The run trains Qwen3.6-35B-A3B on agentic coding with up to 256K tokens of context, using the 2.7K repository tasks Xiaomi released with MiMo-V2.6, and evaluates on SWE-bench Pro V2 and its HARD-51 subset. The authors add masked importance sampling, score centering and top-p sampling with mask replay as off-policy corrections; uncorrected GRPO collapses soon after training starts on the mixed samplers, while SC+MIS with or without top-p trains stably, with pass@4 rising over roughly 100–140 steps and levelling around 65–75% on HARD-51 and rising from 89% to about 94% on the full set, at about 15% lower sampler cost per update for most of training.
Modal Blog · 🔥 0 - Modal launches Endpoint Candidates for evaluating inference endpoints
Modal announced Endpoint Candidates, now generally available for all Modal workspaces. Users can create a candidate deployment for an existing Dedicated Endpoint, choosing an open model from Modal's catalog or bringing their own weights from a Hugging Face repo or a Modal Volume, then configure a traffic sample so the candidate mirrors a portion of production traffic. Production keeps handling every original request and returning the response users see, while the dashboard shows the candidate's latency and throughput metrics and supports load benchmarks; each candidate also has its own URL for running existing evaluations, and it can be promoted to production from the dashboard. Modal says mirroring is best-effort by design, with a separate bounded buffer per candidate so slow or failing mirrors are cancelled without holding up production, and candidates use standard Modal compute pricing including mirrored requests.
Modal Blog · 🔥 0 - K-Dense's Scientific Agent Skills library hits 177 skills, adds AI co-scientist app
K-Dense's Scientific Agent Skills collection — formerly Claude Scientific Skills — has been renamed and rebuilt around the open Agent Skills standard, so it now targets any compatible host including Cursor, Claude Code, Codex and Google Antigravity, not just Claude. Version 2.72.0 ships 177 skills spanning bioinformatics, cheminformatics and drug discovery, medicine, materials, physics, Earth science and scientific writing, with a database skill covering 80 sources and 70-plus version-aware Python package workflows. The project also launched K-Dense BYOK, a free open-source desktop "AI co-scientist" where users bring their own API keys, pick from 40+ models, keep data local and optionally scale to Modal, and it claims 250,000+ scientists already use the library.
GitHub Trending · Python · 🔥 0 - Cua releases Cua Spaces 0.1.0, giving agents full desktop machines
The open-source trycua/cua project has shipped Cua Spaces 0.1.0, a macOS app that gives agents full macOS or Linux desktops running on your Mac, other machines you own, or your own cloud account (AWS, Google Cloud or Modal), with shared human-plus-agent desktops, teleporting of signed-in apps, and sessions kept in a local encrypted Cua Keyvault. The repository also ships Cua Driver for driving native desktop apps and browsers across macOS, Windows and Linux via CLI, MCP or typed SDKs, Lume for local Apple Silicon VMs, the small specialized CUA-S1 'System 1' decision models whose weights are hosted on Hugging Face, and Cua Bench for building tasks, evaluating agents and exporting trajectories. Cua Spaces is source-available under FSL-1.1-MIT and becomes MIT two years after each release, while the rest of the repo is MIT; Spaces is free for individuals, with Pro and Teams plans listed as coming soon.
GitHub Trending（每日） · 🔥 0 - AI's simmering question: where is all the automation?
A The Information column frames the industry debate over why automation has not arrived even as OpenAI said its unreleased model solved or made major progress on hundreds of significant math and computer science problems. Cognition founder Scott Wu argued that ever-more-capable agents plus more data centers, memory and sandboxes will yield "virtual employees," while ex-OpenAI researcher Diogo Almeida, now at TypeSafe AI, maker of Jev, said today's LLMs are optimized for assistance and fail at unattended automation. Anthropic's Claude Code product lead Cat Wu conceded much automation stalls at a semi-automated stage where Claude does 80% and a human does 20%.
The Information · 🔥 0 - Omnigent open-source meta-harness orchestrates Claude Code, Codex and other agents
Omnigent is an open-source AI agent framework and "meta-harness" that, per its project documentation, mixes Claude Code, Codex, Cursor, OpenCode, Hermes, Pi and user-defined YAML agents in the same session, with sessions that follow you between terminal, browser, phone and a macOS desktop app. The README also claims support for first-party API keys or Claude/ChatGPT subscriptions, cloud sandboxes such as Modal, Daytona, E2B and Kubernetes, and policies that pause for approval, cap spend or limit tool access, plus session sharing for teammates. The repository is currently listed on GitHub Trending for Python.
GitHub Trending · Python · 🔥 0
Experience and discussion from the community
Share my Modal experienceAsk about Modal
Nobody has shared their experience with Modal yet.