Repository navigation
Releases
lmget v0.5.0
Added
- Sharing on the local network:
lmget serveshares, read only, the blobs of the store's public models at/blobs/sha256/<sha256>and announces itself by mDNS as_lmget._tcp. WithLMGET_LAN=on, lmget asks the local network for such servers and fetches from them before the source, verifying every byte. Gated and private models, revoked blobs, direct URLs and OCI images are never shared.serveanswers clients on this machine's networks only (--openwidens that), at most 64 at a time. - Private buckets as origins:
--origin s3://bucket/prefix(or inLMGET_ORIGINS) reads blobs from a private bucket on Amazon S3, Cloudflare R2 or another S3 server, signed with AWS Signature Version 4. The endpoint and keys come fromLMGET_S3_ENDPOINT,LMGET_S3_ACCESS_KEY_ID,LMGET_S3_SECRET_ACCESS_KEY,LMGET_S3_SESSION_TOKENandLMGET_S3_REGIONonly, never fromAWS_*or~/.aws. A signed request follows no redirect. Keys the bucket refuses stop the pull with the reason; a missing blob comes from the source. - Does it fit this machine?
lmget infoand--dry-runestimate, for a Hugging Face model with a transformersconfig.json, the selected weights plus the KV cache for 8192 tokens against the memory lmget lets a model use here. It readsconfig.jsononly (a few kB, checked against the plan's hash, never stored), and not for a plan that a revoked file or the license policy refuses. Activations, runtime and speed are not estimated. --to comfyui: each image-model weight file is linked into the ComfyUI folder its workflows look in (diffusion_models,text_encoders,vae,checkpoints, ...), from the Comfy-Org layout, a full checkpoint's tensors, or--to comfyui:FOLDER. Anything else is asked, not guessed.--to colibri: a colibri container goes into~/colibri-models/<org>--<name>with the manifest colibri takes as a complete download; a folder colibri downloaded itself is never touched.
Changed
- Faster downloads: chunks are made durable once a second instead of once per chunk, and a small file costs one sync instead of three. Measured over loopback on an M4 Pro: 2 GiB in 2.2 to 2.5 s (was 4.0 to 5.2 s), 500 files of 64 KiB in 2.9 to 3.5 s (was 8.0 to 8.7 s). A crash costs at most the last second of a download; every file is still verified against its hash.
- A gated repository whose hashes Hugging Face hides now asks for a token and access instead of reporting an invalid sha256.
Security
From an independent review of the network and agent surfaces, and from fuzzing:
- LAN discovery keeps at most 4 servers, only at addresses on this machine's networks, and gives each a deadline before the next origin or the source takes over: one answer can no longer hold a pull or point lmget at loopback, link-local or internet hosts.
- An origin's redirect is followed within its own host and port only.
lmget mcpfetches direct URLs and OCI images from public addresses only, checked at connection time and never through a proxy, and an overlong message no longer ends the session.- lmget's own mDNS responder answers queriers on this machine's networks only, rate-limited.
- A
config.jsonwhose dimensions overflow gets no memory estimate, instead of a KV cache of 0 that says the model fits.
Install
- pip and uv:
uv tool install lmget,pipx install lmgetorpip install lmget. - Container image:
ghcr.io/fabriziosalmi/lmget:0.5.0(andlatest), linux/amd64 and linux/arm64. - Scoop:
scoop bucket add lmget https://github.com/fabriziosalmi/lmget, thenscoop install lmget. - GitHub Actions:
uses: fabriziosalmi/lmget/setup@v0.5.0.
Every archive, wheel and the image carry a build provenance attestation.
Verify
Verify an archive or a wheel: gh attestation verify <file> --repo fabriziosalmi/lmget --signer-workflow fabriziosalmi/lmget/.github/workflows/release.yml --source-ref refs/tags/v0.5.0
Verify the image: gh attestation verify oci://ghcr.io/fabriziosalmi/lmget:0.5.0 --repo fabriziosalmi/lmget
The archives are reproducible: scripts/release-build.sh v0.5.0 OUTDIR on commit ea9e4c5 gives the same bytes, and scripts/wheels.py v0.5.0 OUTDIR WHEELDIR the same wheels.
v0.4.0
Added
- lmget mcp: lmget as a Model Context Protocol server on stdin and stdout, for agents such as Claude Code, Cursor or Zed (
claude mcp add lmget -- lmget mcp). Tools:info(the plan, or a nickname's choice for this machine),list(the store) andget, which downloads only when called withconfirm: trueand links only into the apps and the Hugging Face cache, never into a folder an agent picks. - GGUF repositories from Hugging Face into Ollama:
lmget hf.co/<org>/<name>:<quant>, the name Hugging Face's "Use this model: Ollama" gives. lmget fetches the manifest Hugging Face builds for Ollama (the GGUF, its chat template and parameters), verifies every layer, and puts the model whereollama run hf.co/...finds it.lmget <org>/<name> --to ollamaon a GGUF repository now names that command instead of refusing bare. - License policy:
[policy] licenses = [...]inlmget.lock, orLMGET_LICENSESfor a machine, a CI job or an agent. A model the policy does not allow is refused on the plan, before anything is downloaded (exit code 5). - lmget sbom:
lmget.lockas a CycloneDX 1.6 bill of materials, one machine-learning-model component per model with its files and their sha256. Deterministic, and checked in CI against the official schema.
Install
- pip and uv: each release now carries Python wheels holding the release binary, unchanged, for macOS, Linux (glibc and musl) and Windows, amd64 and arm64. They go to PyPI when this release is published:
uv tool install lmget,pipx install lmgetorpip install lmget. - Container image:
ghcr.io/fabriziosalmi/lmget:0.4.0(andlatest), linux/amd64 and linux/arm64, the release binary on a distroless base, nonroot. - Scoop:
scoop bucket add lmget https://github.com/fabriziosalmi/lmget, thenscoop install lmget. - GitHub Actions:
uses: fabriziosalmi/lmget/setup@v0.4.0installs a release after checking its checksum and build provenance, and keeps the store between runs, keyed bylmget.lock.
Every archive, wheel and the image carry a build provenance attestation.
Changed
- The nickname choice for embedding, reranking, transcription and image-embedding models reads "embed models go to the Hugging Face cache, ..." instead of "a embed model".
Verify
Verify an archive or a wheel: gh attestation verify <file> --repo fabriziosalmi/lmget --signer-workflow fabriziosalmi/lmget/.github/workflows/release.yml --source-ref refs/tags/v0.4.0
Verify the image: gh attestation verify oci://ghcr.io/fabriziosalmi/lmget:0.4.0 --repo fabriziosalmi/lmget
The archives are reproducible: scripts/release-build.sh v0.4.0 OUTDIR on commit 4d1cb8a gives the same bytes, and scripts/wheels.py v0.4.0 OUTDIR WHEELDIR the same wheels.
v0.3.2
Added
Three more apps as targets, each in the layout the app reads:
--to jan:<Jan data>/llamacpp/models/<org>/<name>/with the GGUF and themodel.ymlJan writes when it imports one. lmget knows its ownmodel.ymlby the keylmget_export: true, which stays when Jan rewrites the file; amodel.ymlwithout it is Jan's and is never overwritten. One GGUF per model, and at most one projector.--to localai: the GGUF files in LocalAI's models path, where the file name is the model name.LMGET_LOCALAI_DIR, elseLOCALAI_MODELS_PATH, elseMODELS_PATH, else~/.localai/models.--to exo:<models>/<org>--<name>/with.exo-download.json, the manifest exo writes after verifying a download. The folder is found as exo finds it (EXO_DEFAULT_MODELS_DIR,EXO_HOME,~/.exooff Linux,XDG_DATA_HOMEon Linux).
lmget rm removes the links and the file lmget wrote next to them, nothing else. The README lists every target's folder and layout, a recipe for Text Generation WebUI with --to dir:, and the apps lmget works with.
These layouts were checked end to end with real models in scratch folders, not yet against the running apps.
Verify
Verify an archive: gh attestation verify <archive> --repo fabriziosalmi/lmget --signer-workflow fabriziosalmi/lmget/.github/workflows/release.yml --source-ref refs/tags/v0.3.2
The archives are reproducible: scripts/release-build.sh v0.3.2 OUTDIR on commit 52e1e95 gives the same bytes.
lmget v0.3.1
Fixed
From a review of the whole code base:
- Signed URL credentials (the query string) are no longer written to logs,
--jsonevents,lmget ls, the history or the store; they show asREDACTED. - The Hugging Face token is the one file
huggingface_hubreads ($HF_TOKEN_PATH, else$HF_HOME/token). A project or container that sets its ownHF_HOMEno longer falls back to the personal~/.cache/huggingface/token. - A pinned
syncor a plan from the signed index never moves the Hugging Face cache'srefs/mainback; it only fills it when empty. - A blob rewritten through an exported hard link (a tool saving a file in place) is noticed, hashed again and quarantined. On Windows, the read-only flag that removing a link clears is set again.
- A pickle padded past the scanner's opcode limit is still refused.
- Two files whose names differ only in case or Unicode form are refused before downloading where the filesystem folds them (macOS, Windows).
- The key for oMLX is sent only after oMLX is recognised (by
/health); LM Studio's only when it asks. - On Apple Silicon a nickname takes 30% of the memory for the weights (on 16 GB: an 8B model at 4 bits fits, a 9B one does not); a memory failure only lowers later proposals.
- A Hugging Face repository typed in another case is stored under its own spelling; a license appearing on a license-less model is a change; locks pinned on a tag use the signed index; an empty repository says so; concurrent index updates keep the highest; on Windows, lmget's copies on another drive are recognised.
Verify
Verify an archive: gh attestation verify <archive> --repo fabriziosalmi/lmget --signer-workflow fabriziosalmi/lmget/.github/workflows/release.yml --source-ref refs/tags/v0.3.1
The archives are reproducible: scripts/release-build.sh v0.3.1 OUTDIR on commit fdfa8b0 gives the same bytes.
lmget v0.3.0
New
- Windows, amd64 and arm64:
irm https://lmget.org/install.ps1 | iex. Tested by CI on Windows Server 2025 (the whole test suite and the installer), not yet on a Windows desktop. Paths Windows would not store as written are refused there (device names,:, 8.3 short names such asGIT~1, ...); across drives a model is copied instead of linked. lmget also builds on FreeBSD, OpenBSD and NetBSD (no release). lmget run <model> "<prompt>": gets the model, links it into the app that runs it (oMLX, LM Studio, Ollama, or an OpenAI-compatible--endpoint), and streams the answer. lmget runs no model itself. oMLX is asked to reload its models; apps are recognised by how they answer, not by their port.- A local history (
lmget history;LMGET_HISTORY=offturns it off): the index nodes are asked fastest first, a nickname remembers your--forand--format, a variant that did not fit in memory is not proposed again for 7 days, and a hint points at models not pulled for 30 days. It holds no token, prompt or answer, and goes nowhere. - Nicknames that are not chat models (transcription, embeddings, reranking, image embeddings) go to the Hugging Face cache.
Fixed
- An OCI registry with a port (
localhost:5000) and the store's read-only blobs on Windows. - The installers take
LMGET_RELEASE_URLto install from a mirror.
Verify
Verify an archive: gh attestation verify <archive> --repo fabriziosalmi/lmget --signer-workflow fabriziosalmi/lmget/.github/workflows/release.yml --source-ref refs/tags/v0.3.0
The archives are reproducible: scripts/release-build.sh v0.3.0 OUTDIR on commit 51a916c gives the same bytes.
lmget v0.2.0
New
- Nicknames:
lmget qwenproposes the variant of a model family that fits this machine (memory, Apple Silicon, an NVIDIA GPU, a container's limit) and the apps installed, says why, and asks before downloading.--yes,--for,--formatand--toadjust it; the lock records the concrete model. The project's list (10 nicknames, 47 variants) is published next to the index and signed like it; your own nicknames go in~/.lmget/aliases.toml. --to omlxand--to lmstudiolink a model into those apps' directories as<org>/<name>, one revision at a time, never touching files lmget did not create.
Fixed
- A new store gets an indexed model from the nodes when Hugging Face is down: it stopped with exit code 6. The signed index is fetched then, from lmget.org or from the nodes.
- A target that cannot hold the model is refused before the download, not after it.
- Error events written by the CLI carry
timelike the others.
Verify
Verify an archive: gh attestation verify <archive> --repo fabriziosalmi/lmget --signer-workflow fabriziosalmi/lmget/.github/workflows/release.yml --source-ref refs/tags/v0.2.0
The archives are reproducible: scripts/release-build.sh v0.2.0 OUTDIR on commit c451074 gives the same bytes.
lmget v0.1.0
Verify an archive: gh attestation verify --repo fabriziosalmi/lmget --signer-workflow fabriziosalmi/lmget/.github/workflows/release.yml --source-ref refs/tags/v0.1.0