SIAgentNews.com
← All news

Linger says a 125B-parameter model can run on a 12 GB graphics card

A new open-source agent claims big-model performance on a mid-range gaming GPU, but the speed and quality figures are the project's own and untested by outsiders.

· First reported by explainx.ai

45.5claimed tokens per second on an RTX 3060 (12 GB)

The short version

  • Linger first shipped on November 10, 2026 and is at version 0.1.5, with frequent releases.
  • Its PolyStrata runtime spreads a model's parts across GPU, system RAM and SSD.
  • The README claims 45.5 tokens per second on an RTX 3060 (12 GB).
  • On the 12 GB tier, the first token can take about three minutes after a 62K-token prompt.
  • No independent tests have been published; many figures use simulated hardware.

Why it matters

If the approach holds up, a private assistant with no per-use fees could run on hardware many people already own. But the gap between a README table and your own machine can be large, so measure it yourself. Giving a proactive agent access to your mail and browser also deserves caution.

What Linger is

Linger is an open-source personal agent that runs on your own computer. As explainx.ai reported, the project, published as simplexai-labs/linger, drew about 1,800 stars on GitHub over the ten days after its first release on November 10, 2026. Its pitch is a "Claude-class brain" running locally: a 125-billion-parameter model on a 12 GB graphics card backed by 64 GB of ordinary RAM.

It is more than a model. The README describes five layers: a local model engine, an installer that picks a model for your hardware, a memory system built from your messages, a proactive assistant, and control over desktop apps and the browser. Meeting recording with live transcripts and a handoff to follow-up tasks is included too.

How a big model squeezes onto a small card

The trick rests on a type of model called mixture-of-experts. It is built from many specialist sub-networks, and for each word it produces, only a few of them do any work. So a model can be huge on disk while using a small slice at a time.

Linger's engine, called PolyStrata, sorts those specialists by how often they are used. The busiest stay on the graphics card, middling ones sit in system RAM, and rarely needed ones wait on the SSD (solid-state drive).

Simple arithmetic shows why this is necessary. At 4 bits per parameter, 125 billion parameters would need roughly 62 GB for the weights alone, far more than 12 GB. Any working setup has to shuttle data between memory tiers, and that usually costs speed.

What the README admits about speed

The project's benchmark table says one 125B-class model, Qwen3.8-Flash-Next, generates 45.5 tokens per second on an RTX 3060 with 12 GB and 179 on an RTX 5090 with 32 GB. A token is a small chunk of text, roughly a word fragment. Other listed models run from about 65 to 480 tokens per second, depending on the card.

The caveats matter. The authors say many of the results were produced on simulated hardware setups, so a real machine may be slower. They also say that on the 12 GB tier, the wait for the first token can reach about three minutes after a 62K-token prompt. For an agent that packs lots of context into each request, that delay is significant.

Read the table as a best case after warm-up, not a guarantee.

Is it really "Claude-class"?

The project rates its recommended models on its own SimplexAI score, built on a third-party scale in which 50 is the average and a gap of 3 points or less counts as equal. Local picks range from 35.4 to 57.3, depending on hardware. The README maps 51.9 to about Claude Opus 4.5, 54.0 to about Claude Sonnet 4.6 and 57.3 to about Claude Opus 4.6.

Two things weaken the headline. Those Claude versions are older than today's best models. And the 57.3 result belongs to the 125B setup, which the docs do not recommend by default for a 12 GB machine and which was measured on a simulated tier. A blended score can also hide the specific tasks where a local model fails.

Installing it and what to watch

You can use a desktop app for macOS, Windows or Linux, or run the server through Docker Compose. Release 0.1.4 added support for plugging in your own key for OpenAI-compatible and Anthropic services, so you can skip local models entirely. The README says the Mac app is not notarized and the Windows installer is unsigned, so check each download against the published checksum file before running it.

The larger risk is access. The assistant watches your inbox, calendar and Slack, can act on its own, and can operate a browser where you are already logged in. Start with a separate browser profile and limited accounts, and review what it does before you widen its reach.

The license is "Apache 2.0 with additional conditions", so read it before building a product on it. The repository shows 541 commits and 33 open pull requests, and at version 0.1.5 with releases every day or two, expect rough edges.

Where it fits

Most agent tools assume a hosted model somewhere. Linger bundles its own engine, model picker, memory and server, betting that designing them together lets modest hardware do more. The trade-off is that pulling experts from an SSD is slow when an agent keeps switching between mail, code and meetings. A hosted frontier model will probably stay faster and stronger on hard work. Linger's real appeal is privacy and no token bills.

What to watch

  • Whether independent testers reproduce the speed and quality numbers on real, not simulated, hardware.
  • Whether later releases bring signed installers and a clearer license position.
  • Whether the first-token delay on 12 GB cards shrinks for long prompts.

Try it

If you have a 12 GB GPU and 64 GB of RAM, install it in a sandboxed account, verify the checksum, and time a real task from start to finish.

Sources and links

First reported by explainx.ai. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news