Janas: Run LLMs larger than RAM by streaming experts from SSD

URL SCAN: Janas: Run LLMs larger than RAM by streaming experts from SSD
FIRST LINE: 37 tok/s from Qwen3-Next-80B (a 48 GB model) on a laptop with 32 GB of memory, 24 without drafts, where llama.cpp manages 2.8; reading a prompt, 62 tok/s against 16. Janas-LLM was faster than llama.cpp on every model compared on the same laptop (Debian 13 Linux, Intel Core Ultra 9 185H, its integrated Arc GPU, 32 GB, an NVMe SSD; 9-11 October 2026; the whole machine below):

The Dissection

This is a technical README functioning as a benchmark report, launch document, contributor funnel, and access narrative. Its real demonstration is not that an 80-billion-parameter model has become small. It is that the memory wall can be converted into an SSD bandwidth, cache-locality, latency, thermal, and wear problem.

The text carefully establishes credibility through hardware details, caveats, test methodology, failure modes, licensing notes, and the explicit admission that Janas is an experiment rather than a product. It is trying to make local high-capability inference appear ordinary before it is operationally ordinary.

The Core Fallacy

The core fallacy is technological access being mistaken for economic sovereignty. Fitting a model into 32 GB of RAM proves that a laptop can execute it. It does not prove that the operator owns valuable model capital, controls distribution, possesses customer access, or has an indispensable position in a productive workflow.

Janas reduces one entry barrier; it does not abolish scarcity. The bottleneck moves into disk latency, cache behavior, power, heat, SSD endurance, model quality, reliability, maintenance, and concurrent serving. The laptop becomes a cheaper execution node, not automatically a Sovereign.

Under the Discontinuity Thesis, this result is more threatening to labor than protective of it. Cheaper local inference expands the number of machines capable of cognitive automation. It supports P1, does nothing to defeat P2, and can accelerate P3. A tool that lets more people run an 80B model does not preserve the wage-to-consumption circuit if the model is being used to remove economically necessary human work.

The benchmark claim itself is narrower than the rhetoric around it. It uses one laptop, limited rounds, selected configurations, and comparisons that are explicitly not identical in measurement method. The special llama.cpp build and options used for the largest model make the headline comparison informative but not a universal performance verdict.

Hidden Assumptions

• Sparse expert access will remain cache-friendly. Workloads with shifting prompts or many concurrent users may trigger much more disk traffic and worse tail latency.
• NVMe throughput, power use, thermals, and SSD wear are acceptable substitutes for RAM capacity.
• The quantized models retain sufficient accuracy, context handling, and tool reliability for real work; tokens per second alone do not establish that.
• Results from one Intel laptop generalize to other CPUs, GPUs, disks, operating systems, and memory configurations. The text itself admits they do not yet.
• A model that can run locally is equivalent to a production service. It is not; deployment, security, updates, observability, uptime, and integration remain unsolved costs.
• Open or downloadable weights equal control of AI capital. They do not guarantee durable ownership of a valuable workflow or market.
• Hardware democratization produces bargaining power for workers. More often it gives employers and small operators a cheaper tool for substitution.
• The novelty will persist. Once this technique becomes routine, its advantage becomes infrastructure and its economic rent collapses.

Social Function

Classification: partial truth, prestige signaling, and transition management; secondarily, copium when interpreted as individual liberation.

The engineering result is real and useful. The copium enters when local execution is treated as a return of productive agency. Janas normalizes the next stage of automation by making large-model inference look like a laptop configuration problem. It also lowers the hardware moat, which broadens participation in the transition without changing who captures the resulting value.

This is not a rescue mechanism. It is a friction-reduction mechanism for cognitive automation.

The Verdict

Janas is a competent bottleneck hack with asymmetric consequences. It weakens one centralized hardware advantage while strengthening the general substrate for AI substitution. The text does not refute the Discontinuity Thesis; it supplies another implementation path for it.

The decisive question is not whether a 32 GB laptop can run an 80B model. It is whether humans remain economically necessary after such capabilities become cheap, local, and ubiquitous. This document provides no evidence that they will. Its value is transitional: useful for verification arbitrage, local deployment, and workflow intermediation before those advantages commoditize. As a claim of system emancipation, it is a polished illusion. As an automation enabler, it is exactly the kind of progress that makes the old order more terminal.