LFM2.5 8B-A1B is the pick; Qwen3 8B fell 33 points

Pick of the week

Eight billion parameters that fit an eight-gigabyte card

LFM2.5 8B-A1B · Liquid AI · Q4_K_M 4.8 GB · Heat 56

LFM2.5 8B-A1B by Liquid AI takes the sixth pick: Heat Score 28 to 56 in seven days, the largest gain among models that run comfortably on a 24 GB card and are not on cooldown. Its Q4_K_M file measures 4.8 GB; at 8K context it needs 5.9 GB, which our engine rates GOOD on every 8 GB card we track.

The pick on your card

How LFM2.5 8B-A1B's memory grows with context

The model accepts up to 128,000 tokens of context. Memory = measured weights + computed context cache + runtime, on an RTX 4090's 21.1 GB usable.

What to run this weekend

The largest and the hottest model that fit each card class

"Comfortably" is our EXCELLENT or GOOD verdict for the measured Q4_K_M file at 8K context. Largest = most parameters; hottest = highest Heat Score. Check your own card

Departures board

The largest Heat Score moves of the week

Heat Score: a published composite of Hub trending, download growth, 30-day downloads and likes. The full board

Price board

What an hour cost last week

Week median = the median of the week's daily medians, glitch days excluded. "Few offers" marks a median over fewer than five offers. Cloud vs. market = cheapest big-cloud list price ÷ the same day's Vast.ai median. What these numbers are

Chart of the week

RTX 5090: fourteen days of hourly prices

The RTX 5090 made the week's largest move on the price board, +38 % against the week before.

Daily median of Vast.ai's verified offers, UTC days, glitch days excluded by the same rule as the price board. Gaps are days without a reading.

Price moves

The week's largest day-over-day moves

• 14 Oct · H100 at Vast.ai: +34 % ($2.54 → $3.40)
• 14 Oct · RTX 3090 at Clore.ai: +33 % ($0.25 → $0.33)
• 14 Oct · H100 at RunPod: +30 % ($1.99 → $2.59, list price)

A number explained

Two H100 jumps of 30 % on the same day, and neither was a price rise

On Tuesday 14 October our price alerts fired twice for the same card. Vast.ai's daily H100 median went from $2.54 to $3.40, up 34 percent against Monday. RunPod's listed H100 went from $1.99 to $2.59, up 30 percent. Two sources, one GPU, one day: it looked like a market move. The hourly readings behind the two numbers say something else.

Vast.ai first. Monday's median stood on four verified offers, and only 8 of the day's 23 hourly readings had enough offers to count at all. That is a thin market, and our board marks it as one. On Tuesday the market filled up again, 22 of 24 readings counted, median of 8 offers an hour, but the offers that arrived were expensive ones: the upper quartile of the day sat at $5.39, with hours at $6.80 to $8.50 in the afternoon. The cheapest verified H100 hour was $1.79 on Monday and $1.79 on Tuesday. The floor never moved; the middle did, because the top half of the list changed. By Wednesday the median was back at $2.47.

RunPod is a list price, not a marketplace, and its pricing page shows the cheapest H100 variant it has in stock at that moment: PCIe at $1.99, NVL at $2.59, SXM at $2.69. On Sunday and Monday the PCIe line was the cheapest one shown in 14 and 13 of 24 hourly readings, so the day's median read $1.99. On Tuesday the NVL line led in 12 of 24 readings and the median read $2.59. The posted price of every variant was identical all week. What changed was which variant RunPod had available, flagged "Low" stock throughout.

What did not move at all: AWS at $6.88, Azure at $6.98 and Oracle at $10.00 per H100 hour, every day from 8 to 16 October. A 30 percent swing on the open marketplaces lives inside a 2.6 to 3.9 times gap to the big clouds, and that gap is the number that decides where you rent. The lesson for reading our boards: a median is only as strong as the offer count beside it, and a list price can change without anyone changing a price.

What the pick costs

LFM2.5 8B-A1B, two ways, per day

For 500,000 tokens a day over 4 active hours, the calculator's "regular" preset:

Newly measured

Files measured this week, and where they fit

• Hy3 (Q4_K_M, 166.3 GB): fits no consumer card comfortably at 8K.
• Ling 3.0 Flash (Q4_K_M, 70.1 GB): fits no consumer card comfortably at 8K.
• GLM-4.5-Air (Q4_K_M, 68 GB): fits no consumer card comfortably at 8K.
• gpt-oss 120B (Q4_K_M, 58.5 GB): fits no consumer card comfortably at 8K.
• Qwen3-Coder Next (Q4_K_M, 45.1 GB): fits no consumer card comfortably at 8K.
• Llama 3.3 70B (Q4_K_M, 39.6 GB): fits no consumer card comfortably at 8K.

Memory from the measured GGUF file plus the computed context cache; "comfortably" is our EXCELLENT or GOOD verdict. Check your own card

Rent or buy, one number

2,665 rented hours

At last week's Vast.ai median of $0.60 an hour, the RTX 4090's launch list price of $1,599 (NVIDIA launch MSRP, Oct 2022) equals 2,665 rented hours, ≈ 6 years at 8 h/week. Electricity and resale value are left out on both sides.

Class of the week

RTX 4090

Last week's Vast.ai median: $0.60 an hour. Against 30 days ago: −14 %. Since we started reading it 46 days ago: +22 %. Cheapest source at the week's end: RunPod at $0.34 (list price). 22 of 46 tracked open models run comfortably on one at 8K context; the largest comfortable one is GLM-4.7-Flash.

Maker board

Which labs hold the top-10 slots

Counted from the Heat Score ranking when this issue was built, Monday 05:45 UTC; the 7-day change is the mean per-model change. The full board

New on the radar

Added this week

• Radar Cards claimed: 5 in total. The next number of the RTX 4090 card is #2.

From the tower

What changed on the site

• The Tower Report now carries seven more sections: the pick's context ladder, what runs this weekend per card class, the readings behind every Heat move, a chart of the week, what the pick costs three ways, the maker board and what is new on the radar.
• On cooldown and therefore not eligible as the next pick: LFM2.5 8B-A1B (picked 2026-w42), Ornith 1.5 9B (picked 2026-w41), Granite 4.2 8B (picked 2026-w40), Mistral Small 3.2 (picked 2026-w39). The rule

From the methodology desk

What this issue does not say

Vast.ai medians with fewer than five verified offers are shown but marked "few offers", and hours with too few offers are left out of the daily median. Clore.ai had one to three H100 offers all week and therefore no H100 median at all; it is absent from the H100 row on purpose, not missing.