Janas runs an 80-billion-parameter model on a laptop with 32 GB of RAM.

Janas (GPL-3.0, C, 229 stars, one contributor, repo created September 20, 164 points on HN Monday evening) is an inference engine in C with no dependencies beyond the C library. Its first component, Janas-LLM, keeps the weights on disk, streams the experts each token needs from an NVMe drive and keeps the ones that come back in memory. Linux and Windows x86-64 builds are on the releases page, and ./janas-chat lists the models it knows, downloads one and checks it. On the author's laptop (Core Ultra 9 185H, integrated Arc GPU, 32 GB, one NVMe SSD) it wrote 24 tokens a second from the 48 GB Qwen3-Next-80B against llama.cpp's 2.8, or 37 with the model's own draft block, and beat llama.cpp on all 13 models in its Q4_K_M table. The exception sits in a second table: unsloth's 3-bit Qwen3-30B reads prompts slower than llama.cpp (100 against 112) while writing three times as fast.

The README is unusually plain about method. Each figure is the mean of three rounds, the llama.cpp build and flags are listed, and the 80B comparison needed a llama.cpp build without weight repacking just to load. The draft-speed column was wrong, and too high, until September 26, and the README says so and explains why. janas-bench measures your machine, and the author asks for those reports because the engine has run on exactly one laptop. In the thread, one commenter said there are many stream-experts-from-SSD projects now and another pointed to Strata. A third argued that expert caching is model- and prompt-dependent and "doesn't look like a solved problem yet." A fourth read the catalog's older Qwen3 models as a sign the code was written by an agent that didn't know the newer ones.

Reach for it if you have an x86 laptop, a fast SSD and a mixture-of-experts model that doesn't fit in RAM. Delete the plan to buy 64 GB for a model you'd run twice a week. Tradeoffs: 62 tokens a second of prompt reading on the 80B is slow for agent work; the README warns the engine fills free memory and can push a busy desktop into swap, and that a GPU driver reset once took the desktop down; AMD CPUs, discrete GPUs and Apple Silicon are untested or unsupported; the command line changes several times a week; and janas-get pulls conversions from the author's Hugging Face account, though converting a GGUF yourself is documented.

billion-context (MIT, TypeScript, 1,131 stars, 216 of them today on GitHub trending, repo created August 6, v0.1.196 overnight) sits between your coding agent and the model API. npm install -g billion-context, then bili claude, bili codex or bili opencode starts the proxy and the client together, and bili plugin install <client> is the native route where a client allows one. The proxy adds four tools to the conversation. compress folds a message range into a summary, decompress restores it, search_context searches the summaries and acp_status reports what is still compressible. The model decides when to call them, in place of the host's compaction, which summarizes everything at a threshold and discards the source. Summaries are written so the cached prefix stays intact, and the README says a healthy session keeps a 95 to 97% prefix-cache hit rate.

The evidence is a preprint by the author, in the repo: four and a half months of daily use on three hosts, 174,327 model calls on the main one, and no context-window overruns on 204,800-token models. In a controlled pilot, recall probes tied an untruncated baseline while re-fetch calls dropped 2.7 times. The paper's limitations paragraph says what that adds up to: one user, observational, no randomized control, controlled runs on a single model (qwen3.8-27b), and probe suites that saturate. The production traffic ran on GLM, Qwen and GPT-class models. The "5x fewer tokens" on the README banner isn't one of the paper's measured figures.

Reach for it if your agent sessions hit the window or lose the thread after auto-compaction. Delete the routine of starting a fresh session and pasting in a handoff note. Tradeoffs: for clients with no plugin seam, the launcher intercepts their TLS with its own certificate, which the README calls cert-MITM; for Claude Code the native install writes a managed settings block and a SessionStart hook; it reached v0.1.196 in ten weeks, so expect churn; and all of the evidence is one author's own traffic.

Agent37 (YC F26, Launch HN on Monday, 141 points and 98 comments) hosts sandboxes for agents and their harnesses. OpenClaw, Hermes, Claude Code, OpenCode, pi and others ship as Docker templates, with an MIT-licensed gateway that puts one API across them. Rates are $0.80 per vCPU, $0.70 per GB of RAM and $0.09 per GB of disk per 730-hour month, metered by the minute. Compute stops when a sandbox sleeps; disk bills while you keep it. The founder's post explains the price. They rent from Hetzner and OVH rather than hyperscalers, run gVisor so sandboxes share host memory instead of each reserving a VM's worth, and buy hosts heavy on RAM, because agent harnesses sit idle on CPU and spike on memory for browser work.

The headline is the pricing page's own arithmetic. The $0.37 a month is 2 vCPU and 4 GB running 57 hours, 7% of the month, plus 4 GB of disk. Run that box all month at the same rates and it comes to about $4.76, still under the $6.49 Hetzner CX23 the page lists, which you would operate yourself. The page compares 31 providers at list price and says it compares prices, not throughput. Standard compute shares host resources; the Performance tier costs four times as much for compute. In the thread, the founder counted five or more companies doing something similar in the same YC batch. One user of several months signed up because their bank wouldn't process a payment to Hetzner. The free tier's default model, per the founder, is OpenAI Luna.

Reach for it if you keep a personal agent running that mostly waits, or need many short-lived sandboxes and don't want to run gVisor yourself. Delete the always-on VPS you rent for one Hermes instance. Tradeoffs: gVisor sandboxes can't run Docker-in-Docker or Android emulators, so those need the pricier full-VM tier; commenters expect a race to the bottom in a crowded category; and over 90% of new signups are on the free plan, which the founder puts at $0.05 to $0.10 a user a month to host.

xtool (MIT, Swift, 8,587 stars, 936 of them today on GitHub trending, maintained since November 2024) replaces the parts of Xcode you need to get an app onto a phone. xtool new scaffolds a SwiftPM project, xtool dev builds and runs it, and the device subcommands install, launch and uninstall over USB through usbmuxd. On Windows that means WSL plus USB passthrough. XKit, the library underneath, talks to Apple Developer Services from your own Swift code, and IDE support runs through a Build Server Protocol server. Version 1.20 in September moved builds onto SwiftBuild with Swift 6.4 and SwiftUI macros, which the release notes say brings behavior much closer to Xcode's. 1.21.0 followed on October 5.

The setup docs say where the Mac-free claim stops. You still download Xcode 27's .xip from Apple in a browser, logged in with an Apple ID, and xtool extracts an iOS SDK from it: about 7 GB on disk since 1.20, or --slim for CI. Login comes in two forms. An App Store Connect API key needs a paid developer membership and uses Apple's public API. Password login works with any Apple ID but goes through private Apple APIs, and the docs suggest a throwaway Apple ID "to be extra cautious." They also say credentials go only to Apple and invite you to build from source to check.

Reach for it if you write Swift on Linux or Windows and want your app on your own iPhone without borrowing a Mac. Delete the cloud Mac you rent for device builds. Tradeoffs: there is no iOS Simulator off macOS, so testing means a real device; the password route rests on private APIs Apple can change; and 1.20 changed the SDK layout and the macOS sign-in backend, which meant an SDK reinstall on Linux and a fresh login on macOS.

SolidCraft (Apache-2.0, Rust, 1,414 stars, repo created last Wednesday, v0.6.0 and v0.7.0 three hours apart overnight, 73 points on Show HN) does sketch, constrain, extrude, fillet and shell on a feature timeline: change a parameter and the model rebuilds. Installers cover Windows, macOS and Linux, and there is a browser build. It imports STEP and IGES and exports STEP AP242, IGES, 3MF, STL and OBJ. The B-rep kernel underneath is Truck, an existing Rust kernel. claude mcp add solidcraft -- solidcraft-cli mcp gives an agent tools to run commands, measure, inspect topology, set parameters, export and take a screenshot that returns the image. Headless, it runs its own session; pointed at a running app with --control, you can watch the agent model.

The screenshot tool answers a complaint from the thread. One commenter found that an agent on FreeCAD's MCP didn't build parts from sketches and extrusions, so they were hard to modify, and that OpenSCAD agents lacked "visual reasoning" until given an image tool. SolidCraft ships the second fix; the first still depends on the agent. On accuracy, the README says 75 of 76 public reference parts match their analytic or FreeCAD-computed volume, area, center of mass and bounding box, with the other a tracked bug, and adds that this doesn't establish full equivalence. One commenter said Truck still handles only single-edge fillets. The author's solidcore, created Monday, is an experimental Rust port of the SMLib B-rep kernel subset NVIDIA released.

Reach for it if you print or prototype simple parts and want an agent to change dimensions for you. Delete the OpenSCAD file you regenerate from a prompt every time. Tradeoffs: six days old and, in the README's words, "early but usable for simple parts"; one author made 540 of the commits; downloads aren't signed, so you click past SmartScreen or Gatekeeper; and the kernel is the hard part, still in motion.

Verification notes: star, license, creation, contributor and release data are from the GitHub API on 2026-10-13; trending counts are from today's github:trending scrape. HN points and comments are from the Algolia API. Janas's speeds, method, machine, stability warnings, draft-column correction and model sources are from its README; the prefill, Strata, expert-caching and model-age comments are from HN item 50063717. The five-minute figure is 20,000 tokens at the README's 62 tokens a second. billion-context's install, tools, cache-hit figure and client modes are from its README; the production study, pilot and limitations are from the preprint in its paper/ directory. Agent37's rates, the $0.37 basis, the 31-provider comparison and the Hetzner row are from agent37.com/pricing, checked by Agent37 on October 1; the $4.76 always-on figure is those rates times 2 vCPU, 4 GB and 4 GB of disk. The cost techniques, gVisor limits, free-plan figures, batch count, Hetzner-payment user and Luna model are from HN item 50059096. xtool's commands, setup, login modes and SDK size are from its README, installation docs and the 1.20.0 and 1.21.0 release notes. SolidCraft's features, MCP tools, reference-part checks and unsigned downloads are from its README; the FreeCAD/OpenSCAD and Truck comments are from HN item 50064840. Pipeline health: today's pool held 79 rows, against the ~280 the compose prompt assumes. github:trending returned 10 rows across all five feeds, and Reddit returned zero again, 157 days since its last nonzero day on May 9.