Beta9

license

AGPL-3.0(OSI, copyleft)

source

public

managed-tier

Beam (beam.cloud) fully-managed GPU cloud (proprietary)

core-gated

ungated(the whole AGPL-3.0 stack is in the repo - gateway, scheduler, worker, proto, deploy manifests and even the billing service - with no second license file and no license-key module, and Beam sells concurrency, seats and retention on the hosted cloud)

AGPL-3.0 (OSI copyleft) engine with the managed Beam cloud sold beside it. The 1,281-entry untruncated tree carries one governing LICENSE plus a single vendored one, no enterprise or ee/ directory, and no license-key module; the whole server stack ships - gateway, scheduler, worker, protos, Docker and Kubernetes manifests - and even pkg/gateway/services/compute/billing.go, the metering the hosted product runs on, is inside the AGPL tree. Beam’s pricing tiers concurrency, seats and log retention on the hosted cloud, names no capability the source lacks, and advertises a self-hosted product that "runs entirely in your own environment". That is a separate managed service around a complete open core, which the rubric says does not gate it.

• https://github.com/beam-cloud/beta9 recorded 2026-06-04

  AGPL-3.0 license; open source serverless AI runtime; engine powering managed Beam cloud (beam.cloud)

• https://api.github.com/repos/beam-cloud/beta9/git/trees/main?recursive=1 recorded 2026-08-12

  Full untruncated recursive tree, 1,281 entries. Two license paths: the root LICENSE and one vendored sdk/src/beta9/vendor/pathspec/LICENSE. No ee/, enterprise/, commercial/ or proprietary/ directory. The top level carries the complete server stack - cmd/, pkg/, proto/, deploy/, manifests/, docker/, sdk/ - and the paid-service plumbing is open too: pkg/gateway/services/compute/billing.go is in the AGPL tree.

• https://api.github.com/repos/beam-cloud/beta9 recorded 2026-08-12

  Repo metadata: language Go, license spdx_id AGPL-3.0, 1,741 stars, default branch main.

• https://www.beam.cloud/pricing recorded 2026-08-12

  Three tiers - Developer ($0 + usage), Team ($89/mo + usage), Growth (contact). They differ on concurrency, seats and log retention: 5 vs 50 vs 1,000+ concurrent GPU containers, 1 vs 3 vs unlimited seats, 30-day vs 1-year retention. No feature is named as absent from the open source, and the page states "We also offer a self-hosted product which runs entirely in your own environment, ensuring that no data leaves your VPC."

No PyPI or usage figure for the open source runtime has surfaced and Beam discloses no customer or usage count, so GitHub stars are the only signal: 1,742 stars and 156 forks on the one declared repo, which falls in the 1K-10K star band. A star is not a download, so a star-based reading is capped at level 3 and describes attention rather than volume; this sits at 2 given the small star base and the absence of any usage signal, and confidence is low. A star count moves daily, so the figure is a snapshot rather than a fixed number - the band depends only on its order of magnitude. One gap is worth naming: the README now installs the SDK with `pip install beam-client`, so a countable channel exists that this record does not yet measure.

• https://github.com/beam-cloud/beta9 recorded 2026-06-04

  1.7k stars; no published download/usage figures

• https://api.github.com/repos/beam-cloud/beta9 recorded 2026-08-13

  Repo metadata - 1,742 stars, 156 forks, 6 watchers, pushed 2026-08-13, not archived. Bands at 1K-10K stars, level 2 on the stars scale in signal_routing.yaml.

• https://raw.githubusercontent.com/beam-cloud/beta9/main/README.md recorded 2026-08-13

  Installation is `pip install beam-client`, and the page publishes no download, customer or usage figure of its own - only a stars badge.

Capable serverless container runtime with GPU support, volumes, autoscaling and sandbox primitives, but container-level rather than VM or microVM isolation places it below the VM-grade peers on isolation strength; mid-tier. The README claims container launches "in under a second" from a custom runtime, scheduler and embedded cache, fan-out to hundreds of containers, scale-to-zero, distributed volume storage, GPU support on 4090s and H100s or bring-your-own, and a Sandbox primitive for running model-generated code - but it describes containers throughout and names no VM, microVM or application-kernel tier where the category’s frontier-definers do, and that isolation ceiling is what holds the score at 3, two rungs below Firecracker at the frontier. Independent ComputeSDK TTI benchmarks of Beam (the managed beta9 cloud) corroborate the fast container launch at a 0.29s median, 4th-fastest by median of 24 providers, up from ~380ms and 8th of 19 on the earlier run; the band turns on isolation strength rather than latency, so a faster benchmark does not move it.

• https://github.com/beam-cloud/beta9 recorded 2026-06-04

  serverless runtime: rapid container launches, parallelize across hundreds of containers, GPU support, volume storage, sandboxes

• https://raw.githubusercontent.com/beam-cloud/beta9/main/README.md recorded 2026-08-13

  Features are fast cold starts ("Launch containers in under a second using a custom container runtime, scheduler, and embedded caching"), fan-out to hundreds of containers, scale-to-zero, mounted distributed volumes, and GPU support (4090s, H100s, or bring your own). A Sandbox primitive spins up isolated containers to run LLM-generated code. No VM, microVM or application-kernel isolation tier is named anywhere.

• https://www.computesdk.com/benchmarks/sandboxes/ recorded 2026-08-13

  ComputeSDK TTI leaderboard, run of 2026-08-07, 24 providers. Beam (the managed beta9 cloud) posts a 0.29s median, 0.42s P95, 0.43s P99 and 100% success for a composite of 96.6 - 4th-fastest by median.

• https://www.computesdk.com/benchmarks/sandboxes/ recorded 2026-07-09

  ComputeSDK TTI leaderboard (independent, 100 iters/day, 2026-07-09 run): Beam (managed beta9) median time-to-interactive ~380ms sequential / ~810ms burst @100 concurrent; 8th of 19 providers