Issue 001: The 24 GB question

Published 13 October 2026. Prices checked 12 October 2026. Every figure links to its source; numbers from our own datasheets link to the datasheet, which cites the original.

Cover story: the 24 GB question

A year ago, the honest answer to “how much GPU memory do I need?” was “as much as you can afford”. In October 2026 it has a number. The two strongest open-weight models that a single consumer card can hold are both Apache-2.0 and both under 32 billion parameters: Alibaba's Qwen3.8-27B (27.8B parameters, released 5 August)¹ and Google's Gemma 4 31B (31.3B, released 11 March)².

On LMArena's text leaderboard of 8 October, Gemma 4 31B scores 1443.0 and Qwen3.8-27B 1441.1. That is within each other's confidence intervals, and within about 50 points of the top open-licence model.³ For comparison, Gemma 3 27B, last year's 24 GB workhorse and still the model on our datasheet, scores 1357.7 on the same snapshot.³

Now the arithmetic. At Q4_K_M, the usual quality and size trade-off, the Qwen3.8-27B file is 16.46 GB⁴ and the Gemma 4 31B file is 18.32 GB.⁵ On a 16 GB card neither fits before you add a single token of context. On a 24 GB card both fit, with roughly 5.7–7.5 GB left for the KV cache and the runtime. At 8-bit (29.05 GB and 32.64 GB⁴⁵) neither fits on 24 GB.

So 24 GB is the line. The trouble is what it costs to cross it.

GPU prices against launch

GPU prices on 12 October 2026 against launch MSRP
Card	Memory	Launch MSRP	Price, 12 Oct 2026	vs launch
GeForce RTX 3090 24 GB (used)	24 GB	$1,499⁶	$1,200 (used, eBay sold median)	−20%
GeForce RTX 4090 24 GB	24 GB	$1,599⁷	$4,395 (new, discontinued)	+175%
GeForce RTX 5090 32 GB	32 GB	$1,999⁸	$7,199.99 (new)	+260%
GeForce RTX 5060 Ti 16 GB	16 GB	$429⁹	$779.99 (new)	+82%

The cheapest new RTX 5090 we found on Newegg US was $7,199.99, 3.6× its $1,999 launch price.⁸ PC Tech Vault's tracker put the cheapest in-stock US listing at $6,995 this week, with a 90-day median of $4,900.¹⁰ The RTX 4090, now discontinued, lists at $4,395. Even the 16 GB RTX 5060 Ti, the card that cannot hold either model above, sells for 82% over its $429 MSRP.⁹

Against all that, the six-year-old RTX 3090 looks strange: its used price, about $1,200 (eBay sold median), is only 20% below its 2020 launch price of $1,499.⁶ It has kept its value because it has 24 GB and 936 GB/s of bandwidth. In llama.cpp's shared benchmark it generates 158.16 tok/s on Llama 2 7B Q4_0, against 186.21 for the 4090 and 290.02 for the 5090.¹¹ For running one 27–31B model at a time, it is the cheapest way across the line.

What about unified memory?

The 128 GB machines answer a different question. A Ryzen AI Max+ 395 mini PC ($3,499.99) and an NVIDIA DGX Spark ($4,699) will load 8-bit weights, several models at once, or a 120B mixture-of-experts. But their memory bandwidth (256 and 273 GB/s) is under a third of the 3090's, and decode speed tracks bandwidth (see why bandwidth is the speed limit). On the same Llama 2 7B test they reach 51.81 and 57.21 tok/s.¹²¹¹ The base Mac Studio M5 Max ($2,499, 36 GB, 460 GB/s) manages 119.92 tok/s¹³ and, with 36 GB of unified memory, also clears the line, though macOS reserves part of that memory for the system.

Our verdict for October: if one strong model is the goal, buy 24 GB of fast memory. A used 3090 is the cheapest route and a base Mac Studio the quietest. Buy 128 GB only if you need several models loaded at once, or models too large for 24 GB, and you can accept a third of the speed. We have not yet measured Qwen3.8-27B or Gemma 4 31B on our own bench. They are not in our model catalogue yet, and we will not publish speeds for them until we have measured them.

State of local AI, October 2026

Subscription break-even

Months until the hardware pays for itself: hardware price ÷ (plan price − electricity). This assumes 4 hours a day at the card or chip's rated power, at the US average residential rate of 18.34¢/kWh for June 2026.¹⁴ GPU prices are card only, so add a host PC if you don't have one.

Break-even months by rig and plan
Rig	Power, electricity / mo	vs ChatGPT Plus ($20)	vs Claude Max (5×) ($100)	vs ChatGPT Pro 200 ($200)
GeForce RTX 5060 Ti 16 GB $779.99	180 W, $3.96	49 months (4.1 years)	8.1 months	4.0 months
GeForce RTX 3090 24 GB (used) $1,200	350 W, $7.70	98 months (8.1 years)	13.0 months	6.2 months
Ryzen AI Max+ 395 mini PC 128 GB $3,499.99	120 W, $2.64	202 months (16.8 years)	36 months (3.0 years)	17.7 months

This article excludes idle electricity and host power for GPU cards. The calculator instead uses assumed whole-system load at an editable rate; build pages also include idle draw. Tax, shipping and peripherals are excluded. Plan prices: OpenAI, Anthropic, OpenAI. Power is rated board or chip power from the datasheets (3090, 5060 Ti, Ryzen AI Max+), not measured at the wall. Try your own numbers.

Against a $20 plan, a used 3090 takes 98 months (8.1 years) to pay back. Against a $100–$200 plan, it pays back in 6–13 months. A local 27B model is not a frontier model, so the fair comparison is with what you actually use the subscription for.

Notable releases

Aleph Alpha Kolibri (3 Oct): an English–German mixture-of-experts with 78B total and 3B active parameters, up to 1M tokens of context, Apache 2.0.¹⁵ With so few active parameters it should decode quickly on 128 GB unified-memory machines; we have not measured it yet.
JetBrains Mellum2.1 (8 Oct): a 12B mixture-of-experts coding model with 2.5B active parameters, Apache 2.0. JetBrains says GGUF builds are “coming soon”.¹⁶
Liquid AI d1-3B and d1-omni-600M (7 Oct): “decision” models that return typed answers in one forward pass instead of generating text, with day-one llama.cpp support. The licence is free for commercial use below $10M in annual revenue.¹⁷
Reflection Beam (announced 5 Oct): a 501B mixture-of-experts reasoning model. Reflection says the weights will follow this month, and its benchmark claims have not been independently verified.¹⁸ It is far beyond home hardware.

Rank moves

Among open-licence models on LMArena's text leaderboard, between the 11 September and 8 October snapshots:³

Xiaomi's MiMo-V2.6-Pro entered at the top of the open models, scoring 1491.7 (9th overall). It has only 4,069 votes so far, so its interval is wide (1482.3–1501.0).
Z.ai's GLM-5.3-Max slipped from first to third among open models (1477.5 → 1471.2), and Kimi K3 Max held second (1475.5).
DeepSeek-V4.1-Flash-Max entered seventh among open models at 1462.4.
At home scale nothing moved: Gemma 4 31B (1441.7 → 1443.0) and Qwen3.8-27B (1439.9 → 1441.1) stayed within their intervals.

All the new entries at the top are frontier-scale models for data centres. Rankings on our site cover only models in our catalogue; see rankings.

Runtime news

Ollama 0.40 runs supported architectures on Apple's MLX by default on Apple Silicon.¹⁹ Version 0.40.2 then upgrades previously downloaded models in the background “for better performance and compatibility when running on llama.cpp”, keeping a backup copy on disk. If you are short of disk space, check for these backups.²⁰
vLLM 0.31.0 adds vllm preload, a daemon that keeps quantised weights resident in GPU memory across engine restarts. The release has 717 commits from 307 contributors.²¹
llama.cpp keeps shipping several builds a day. The latest at press time was b11581.²²

Featured use cases

A private voice assistant for the home

Wake word, speech recognition, a small language model and a voice, all on a box in the cupboard. Nothing you say leaves the house.

Real example: Home Assistant Assist, fully local. Three tiers, each linked to a build and checked against its usable memory.

Offline document search with RAG

Ask questions of your own PDFs, notes and manuals, with citations, on a machine that never phones home.

Real example: Open WebUI knowledge bases. Three tiers, each linked to a build and checked against its usable memory.

Corrections

None. This is the first issue.

Sources

Datasheet prices (Newegg, getpcparts, Apple and others) are linked inline and listed on each hardware page with their check dates.