LLMs-local (curated list)

A community awesome-list GitHub repo (maintainer 0xSojalSec) that catalogs the tooling for running LLMs on your own hardware — platforms, inference engines, models, builder tooling, hardware, tutorials, and communities. Reached the wiki via a promotional tweet from Dan Kornas (@DanKornas, 2026-06-18) that frames it as "the map" for moving from "which local stack should I try?" to a usable shortlist by grouping resources by the job they do.

It is filed here as the wiki's first map of the on-device regime's ecosystem — the named tools that fill out the local-llm-stack that llama-cpp and quantization describe at the mechanism level.

What it catalogs

The list organizes the local-LLM world by task rather than by vendor:

Inference stack, split two ways. It deliberately separates platforms — turnkey apps like LM Studio, Jan, LocalAI — from engines — the runtimes underneath: Ollama, llama.cpp, vLLM, SGLang, MLX. This platform/engine split is the substantive idea the wiki takes from it (see local-llm-stack).
UIs. Local chat / front-end tools: Open WebUI, Lobe Chat, text-generation-webui, SillyTavern, Page Assist.
Model navigation. Explorers, benchmark links, providers, and per-use-case model sections (general, coding, multimodal, image, audio).
Builder tooling. Agent frameworks, MCP, RAG, coding agents, browser automation, memory, testing, evaluation, observability.
Learning path. Tutorials grouped by models, prompt engineering, context engineering, inference, agents, RAG, and misc.

Why it's filed, and its weight

Tier T4 — a single-maintainer curated link aggregation surfaced by a marketing tweet, with no benchmarks, no first-party measurement, and no analysis beyond the grouping itself. It does not move the spoke's standing open question (the which-lever-bought-what benchmark decomposition); its value is coverage of the edge ecosystem, which the spoke had only as two engine pages (vllm, llama-cpp) and a synthesis claim that llama.cpp is "the core of Ollama/LM Studio." The list names the layer above those engines for the first time. Several of its sections (agents, MCP, RAG, coding agents, evaluation) are out of this spoke's scope — they belong to agentic-tooling-wiki and llm-providers-wiki — and are recorded here only as context, not paged. Maintainer 0xSojalSec was deferred here as a thin creator and has since been paged in speech-audio-wiki as 0xsojalsec, on recurrence: the same curator produces the same artifact for audio (free-voice-clone-list — 53 open speech/music models, same table-plus- detail construction, same absence of any measurement). Dan Kornas noted inline rather than as entity nodes, matching the how-does-vllm-work precedent (low graph signal for an inference-mechanics spoke). Being a live list, its contents are volatile; claims here trace to the snapshot read 2026-06-21.

Tier: T4 → T3 (2026-08-25)

Graded under the catalogue clause QUALITY.md added on 2026-08-21. The list does organise the field — the platform/engine split above is a real editorial decision, and it is the thing this wiki takes from it — but the criteria for getting in are nowhere stated, so you can see what the maintainer collected and never what was rejected. That is the unstated-selection branch: T3, not T4, and not the T1 a catalogue earns only for exclusions it declares in its own text.

local-llm-stack · llama-cpp · vllm · quantization · llm-inference

Re-verified 2026-08-21 (staleness queue)

Repository live and unarchived: 1,083★ / 118 forks, read from the GitHub API. The number that matters is the other one — last pushed 2026-06-10, which is eleven days before this page was written and seventy-two days ago now. The list has not moved since the tweet that surfaced it.

That changes what the page is for. It was filed as a map of the on-device ecosystem, and a map of a fast-moving ecosystem that stopped being edited two months ago is a snapshot of June 2026, not a current shortlist. Anything a reader takes from it — which engines exist, which platforms are worth trying — should be checked against local-llm-stack rather than trusted from the list.

T4 unchanged, and correctly so under the corpus rule (../QUALITY.md → tier by the claim, not the URL): its carve-out puts curated link lists outside the first-party-artifact ladder entirely, and keeps a directory at T4 where no inclusion criteria are stated. None are stated here. The platform / engine split this page took from it stands on its own merits, and that borrowing is recorded on local-llm-stack rather than resting on the list's authority.

Re-verified 2026-09-04 (re-sent by the curator)

The same URL arrived again over Telegram. Nothing about the repository has moved: 1,138★ / 130 forks (up 55 stars and 12 forks since August), not archived, no license file — and still last pushed 2026-06-10. That is now 86 days without an edit, against 72 at the last check.

So the August reading holds and hardens. The list is a June-2026 snapshot of the on-device ecosystem, not a live map, and the gap between "how fast this field moves" and "how long this file has sat still" is the only fact on the page that keeps growing. Read it for the platform/engine split, which is an editorial idea and does not decay; take the tool names from local-llm-stack instead.

Tier stays T3 under the 2026-08-25 regrade above. (The 2026-08-21 note below it says "T4 unchanged" — that was written before the regrade and is superseded by it, not a live disagreement.)