TESIGN
SAVED

EN · KO

  1. TESIGN / RADAR
  2. Slotstream

Slotstream

Run a 105 GB model on a 48 GB Mac

WHO USES ITApple Silicon Mac owners · Privacy-conscious users · Local-LLM developers

Install to use

Run a 105 GB Qwen model on a 48 GB Mac: experts stream from SSD, Ollama-compatible API

★ 360▲ 240 HN

360 stars · checked on GitHub GitHub check 30 h late · no 7-day star increase observed (GH Archive) (as of ) · 240 Show HN points observed

BACK CATALOG

Terminal running slotstream doctor and slotstream run: the memory-plan check and a model reply

TESIGN TAKE

Streaming a model larger than RAM from SSD really works here, but it supports exactly one model and speed varies widely by machine.

INSTALL

AT A GLANCE

LICENSE
MIT
USAGE
Use, change and redistribute, commercially too. Keep the notice.
LANGUAGE
Swift
PLATFORM
macos · cli
ACTIVITY
last commit 1 day ago () · latest release v0.2.18 () · 26 releases
COMMUNITY
3 contributors · 7 open issues (incl. PRs) · made by: an individual
SOURCES
Show HN
OPEN SOURCE
YES
FIRST SEEN
CATEGORY
AI

A summary, not legal advice.

WHY IT MATTERS

Qwen3.8-Flash-Next is a 105 GB mixture-of-experts model (per the README); Slotstream keeps the shared weights in RAM and reads only the experts it needs from SSD into a cache. The developer's 48 GB M5 Pro measured 13.47 tok/s after warm-up (a 0.2.16 pre-release benchmark), while a community report on a 16 GB M2 Mac mini was 1.41 tok/s. After a one-time download it works offline with no Python or cloud account.

BUILD FROM THIS

  • Point Open WebUI or an agent such as Hermes or fx at its OpenAI/Ollama-compatible API and you have a personal AI server on one Mac that answers questions about files and pictures without an internet connection. A Swift library lets you embed it in a Mac app directly.

WHO IT'S FOR

Apple Silicon Mac owners
run a large model locally without a cloud account
Privacy-conscious users
keep document and image questions on the device
Local-LLM developers
study an MLX + Swift expert-streaming implementation

START IN 5 MINUTES

# curl -fsSL https://raw.githubusercontent.com/carloslfu/slotstream/main/install.sh | sh
# slotstream doctor
# slotstream run --prompt "Why is the sky blue?"

CAVEATS

  • MIT for the code; model weights carry a separate Qwen community licence.
  • Requires an Apple Silicon Mac with 16 GB+ RAM, macOS 14+ and about 110 GB of free SSD; 8 GB Macs refuse to start, and Windows, Linux and Intel Macs are not supported.
  • The model download can take hours, long prompts can take minutes before the first token, and only one generation runs at a time.

RECEIPT

FIRST SEEN
AT SOURCE
KEPT
CREATED → FIRST SEEN
11d
SOURCES
Show HN
◌ BACK CATALOG
Reconstructed from archive data, not a live discovery.

The same facts in machine-readable form — View as Markdown · JSON

SIMILAR TOOLS

Up to five from AI by ★ total: edited entries first, then repository cards; this entry is highlighted. Drawn from the same stored snapshot as the rankings — a comparison, not a recommendation.

IMAGENAME★ TOTAL▲ 7dLICENSEPLATFORMLAST PUSH
ECC README banner: headline 'The operating system for AI agent harnesses', harness chips, and skill, agent and command lists.
eccSkills, memory and security for agents★ 259k▲ +832MITpermissivewindows · macos · linux · cli
Hermes Agent landing: blue background, the headline 'The agent that grows with you' and an engraving-style winged figure.
Hermes AgentSelf-improving personal agent★ 246k▲ +375MITpermissivelinux · macos · windows · android · cli
DeepSeek Harness landing: blue background, the headline 'Everything is a plugin' and a quick-start box with the npx command.
DeepSeek HarnessPlugin agent harness for developers★ 225k▲ +1,356MITpermissiveweb · cli
Terminal running slotstream doctor and slotstream run: the memory-plan check and a model reply
SlotstreamRun a 105 GB model on a 48 GB MacTHIS ENTRY★ 360—MITpermissivemacos · cli
mattpocock/skillsCARD★ 261k▲ +872MITpermissive—

ALSO WORTH SEEING

OpenHunterAILocal AI red-team workspaceOpenHunterAI's homepage, labelled "AUTHORIZED EXTERNAL SECURITY TESTING," listing its five-stage process from scope verification to reporting.Self-host
★ 165▲ +15 / 7d
An alpha-stage local workspace that runs attacker-style security assessments — strictly against public targets you own or are authorized to test.MartinLoopExecution-control CLI for coding agentsMartinLoop's homepage headline: "Watch your AI burn $5.20. Then watch MartinLoop close the loop for $2.30."
★ 130▲ +21 / 7d
Wraps a coding agent with a spend cap and a verifier, so "I'm done" comes with an evidence-backed receipt instead of just a claim.Spec KitSpec-first process for coding agentsSpec Kit docs landing page — the headline 'Build with a spec, fix a bug, or assess an idea' and the uv tool install specify-cli commandInstall to use
★ 137k▲ +560 / 7d
Makes your coding agent write the spec first, then plan, build and verify against it.HindsightNarrative map for YouTube viewersHindsight landing page: thousands of video dots grouped into issue clusters, with a central search box for a topic, creator or YouTube link.Use in the browser
★ 0▲ 1 HN
Remembers 49,411 videos, maps 22,904 by issue, traces how a YouTuber's argument changed.

TIMELINE

  1. Repository created
  2. Show HN post ↗
  3. FIRST SEEN BY TESIGN ◌ BACK CATALOG