Qwen3 8B is the pick; LFM2.5 8B-A1B fell 31 points

Pick of the week

An eighteen-month-old 8B climbs the trending list

Qwen3 8B · Alibaba Qwen · Q4_K_M 4.7 GB · Heat 65

Qwen3 8B by Alibaba Qwen takes the seventh pick: Heat Score 45 to 65 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 4.7 GB; at 8K context it needs 6.8 GB, EXCELLENT on every 12 GB card we track. At 32K it needs 10.9 GB, and 4.5 GB of that is context cache — the mirror image of last week's pick.

The pick on your card

How Qwen3 8B's memory grows with context

The model accepts up to 40,960 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 4090: fourteen days of hourly prices

The RTX 4090 made the week's largest move on the price board, +25 % 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

• 20 Oct · A100 at Clore.ai: +112 % ($0.33 → $0.71)
• 21 Oct · RTX 4090 at Vast.ai: +70 % ($0.54 → $0.92)
• 21 Oct · A100 at Vast.ai: +35 % ($0.73 → $1.00)

A number explained

The RTX 4090 hour rose 80 percent on one marketplace, and the board's cheapest reading did not move

Between Friday 17 and Wednesday 22 October, Vast.ai's daily median for an RTX 4090 hour went from $0.53 to $0.96. That is an 80 percent rise in five days on the most liquid class we track, and unlike last week's H100 story it was not a thin market changing shape: the market held 48 to 64 offers every day, and the whole distribution moved, the lower quartile from $0.46 to $0.67, the upper from $0.69 to $1.00. The cheapest verified offer stayed between $0.35 and $0.38 all week. The floor did not move; everything above it did. By Thursday the median had come back to $0.80, with the day still open.

The same five days on the other two sources: Clore.ai's RTX 4090 median went from $0.62 to $0.77 and RunPod's posted price stayed at $0.34, the cheapest 4090 type it listed, in every hourly reading of the week. On our price board the lime cell of the RTX 4090 row is the cheapest competing reading across all sources, and it read RunPod $0.34 on every one of these days. A renter reading the board saw no change at all; a renter reading one marketplace saw the price of a 4090 hour nearly double. That is the reason the board exists.

The RTX 5090 moved too, less: Vast.ai from $0.83 on the 17th to $1.07 on the 22nd, up 29 percent, Clore.ai between $0.83 and $0.95, RunPod flat at $0.69. Set the two cards side by side on Vast.ai on Wednesday and the 32 GB card cost $1.07 against $0.96 for the 24 GB card, eleven percent more for a third more memory. Per gigabyte-hour that is $0.033 for the 5090 against $0.040 for the 4090: for two days the newer card was the cheaper memory on that marketplace. On RunPod the same two cards sit at $0.69 and $0.34, double the price for the bigger card, and the gigabyte-hour favours the 4090. Which card is cheap depends on where you ask, and on which day.

The pick this week is Qwen3 8B, which needs 6.8 GB at 8K context and 10.9 GB at 32K, so none of this touches it: a 12 GB card already in your machine runs it at 8K and a 16 GB card at 32K. For the renters among you the lesson is the one the board is built on: a marketplace median is a reading of that marketplace at that hour, not the price of a GPU, and the cheapest reading across eight sources moved by nothing this week.

What the pick costs

Qwen3 8B, 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

• Qwen3.8 Flash-Next (Q4_K, 103.7 GB): fits no consumer card comfortably at 8K.
• DeepSeek-R1 (Q4_K_M, 376.7 GB): fits no consumer card comfortably at 8K.
• Qwen3 14B (Q4_K_M, 8.4 GB): fits a 16 GB, 24 GB card comfortably at 8K.
• Qwen3 8B (Q4_K_M, 4.7 GB): fits a 12 GB, 16 GB, 24 GB card comfortably at 8K.
• LFM2.5 2.6B (Q4_K_M, 1.6 GB): fits a 12 GB, 16 GB, 24 GB card comfortably at 8K.
• LFM2.5 1.2B (Q4_K_M, 0.7 GB): fits a 12 GB, 16 GB, 24 GB 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

2,104 rented hours

At last week's Vast.ai median of $0.76 an hour, the RTX 4090's launch list price of $1,599 (NVIDIA launch MSRP, Oct 2022) equals 2,104 rented hours, ≈ 5 years at 8 h/week. Electricity and resale value are left out on both sides.

Class of the week

RTX 5090

Last week's Vast.ai median: $0.96 an hour. Against 30 days ago: +10 %. Since we started reading it 46 days ago: −4 %. Cheapest source at the week's end: RunPod at $0.69 (list price). 29 of 46 tracked open models run comfortably on one at 8K context; the largest comfortable one is Qwen AgentWorld 35B-A3B.

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

• Since 23 October every memory table and derived price figure carries a mark that says whether a number is measured, computed or assumed; the fit API returns the same as a basis block. Correction to pick #2 (Gemma 4 26B-A4B, week 38): the 64K and 128K figures were computed before the engine read Gemma 4's global-attention head shape; the report now carries a dated correction with the recomputed figures, 20.3 GB and 23.6 GB.
• On cooldown and therefore not eligible as the next pick: Qwen3 8B (picked 2026-w43), LFM2.5 8B-A1B (picked 2026-w42), Ornith 1.5 9B (picked 2026-w41), Granite 4.2 8B (picked 2026-w40). The rule

From the methodology desk

What this issue does not say

Daily medians are the median of the counted hourly medians of the UTC day; a day with a hole in the hourly readings is marked in the data and the two marketplaces had none this week. RunPod's figure is a posted price for the cheapest listed type of the class, not a market reading; a change there means a different type was in stock, not a price change. Thursday 23 October is counted up to 18:00 UTC. The per-gigabyte figures divide a marketplace median by the card's memory and price memory only.