Skip to content
AI Compute Radar

The Tower Report · Week 46, 2026

Week 46, 2026 · 9 Nov – 15 Nov

Qwen3 14B is the pick; Qwen AgentWorld 35B-A3B lost the most

Compute weather
H100 at the Vast.ai median on 15 Nov: $2.85 per hour · the cheapest big cloud (AWS) listed 2.4× that · 14 price moves of 10 % or more · 5 popularity moves of 5 points or more. All prices

Name your graphics card or Mac once and every issue opens with what it runs this week. Choose your card

Pick of the week

8.4 GB on disk, and the 12 GB cards miss it by 0.2 GB

Qwen3 14B · Alibaba Qwen · Q4_K_M 8.4 GB

Qwen3 14B: the pick of the week

Qwen3 14B by Alibaba Qwen takes the tenth pick: the largest popularity gain between the readings of 7 and 13 November among models that run comfortably on a 24 GB card and are not on cooldown. Its Q4_K_M file measures 8.4 GB, read from Qwen's own GGUF repository. At 8K context it needs 10.8 GB: it runs easily on 24 GB, runs on the 16 GB cards and spills into RAM on the 12 GB cards, which miss it by 0.2 GB. At 32K it needs 15.3 GB and runs on 24 GB.

Read the report and its caveats →

The pick on your card

How Qwen3 14B's memory grows with context

ContextMemoryComfortable
8K tokens 10.8 GB from 16 GB
32K tokens 15.3 GB from 24 GB
64K tokens beyond the 40,960-token window this model publishes 17.5 GB from 24 GB
128K tokens beyond the 40,960-token window this model publishes 19.5 GB from 32 GB

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.

Run it against your own card →

What to run this weekend

The largest and the hottest model that fit each card class

Card classRuns comfortably at 8K context
12 GB
RTX 3060 12GB · 15 of 49 fit
Largest: Gemma 4 12B (8.1 GB)Hottest: LFM3 1.7B (2 GB)
16 GB
RTX 4060 Ti 16GB · 17 of 49 fit
Largest: gpt-oss 20B (12.2 GB)Hottest: LFM3 1.7B (2 GB)
24 GB
RTX 4090 · 24 of 49 fit
Largest: GLM-4.7-Flash (18.9 GB)Hottest: Qwen3.8 27B (17.1 GB)

"Runs" means our verdict for the measured Q4_K_M file at 8K context is "Runs easily" or "Runs". Largest = most parameters; hottest = highest popularity score. Check your own card

Departures board

Who gained and who lost this week

Model9 Nov15 NovMove
Qwen AgentWorld 35B-A3B
Alibaba Qwen · trending 4 → 3 · 9,478 downloads in 30 days (−7.1 % w/w)
53 27 −26
Nemotron Nano 9B v2
NVIDIA · trending 1 → 1 · 130,180 downloads in 30 days (+24.3 % w/w)
18 43 +25
GLM-5.2
Zhipu AI · trending 7 → 4 · 309,982 downloads in 30 days (−4.4 % w/w)
67 43 −24
KAT-Coder V2.5
Kwaipilot · trending 3 → 2 · 37,045 downloads in 30 days (+557.8 % w/w)
23 45 +22
Qwen3 8B
Alibaba Qwen · trending 13 → 7 · 6,050,582 downloads in 30 days (−13.0 % w/w)
64 42 −22

The score behind these moves is the one the board shows: a published composite of Hub trending, download growth, 30-day downloads and likes. On the pages it reads as popularity. The full board

Price board

What an hour cost last week

GPUVast.ai medianvs. week beforeCheapest sourceCloud vs. market
RTX 3090 $0.24−4 % $0.17
Clore.ai median
—
RTX 4090 $0.50−3 % $0.34
RunPod list
—
RTX 5090 $0.69−2 % $0.62
Clore.ai median
—
L40S $0.800 % $0.79
RunPod list
2.3×
AWS
A100 $0.840 % $0.62
Clore.ai median
3.4×
AWS
H100 $2.94−9 % $2.59
RunPod list
2.4×
AWS
H200 $5.20−8 % $3.59
RunPod list
1.5×
AWS
B200 $10.00−6 % $5.98
RunPod list
1.3×
Oracle Cloud

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

H100: fourteen days of hourly prices

The H100 made the week's largest move on the price board, −9 % against the week before.

H100 on Vast.ai, daily median per hour, 2 Nov $2.85 to 15 Nov $2.85

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

  • 13 Nov · H100 at Vast.ai: +22 % ($2.87 → $3.50)
  • 15 Nov · A100 at RunPod: +19 % ($1.00 → $1.19, list price)
  • 10 Nov · H100 at RunPod: +17 % ($2.29 → $2.69, list price)

A number explained

3 list changes this week, all made by the rule

On 16 November the rule put Gemma 4 26B-A4B in place of Ornith 1.5 9B on the Chat list: meets the entry rule with more than 2× the 30-day downloads of Ornith 1.5 9B, the least downloaded entry. On 16 November the rule put Gemma 4 26B-A4B in place of Ornith 1.5 9B on the Long documents list: meets the entry rule with more than 2× the 30-day downloads of Ornith 1.5 9B, the least downloaded entry. On 16 November the rule put Ornith 1.5 9B in place of Granite 4.2 8B on the Code list: meets the entry rule with more than 2× the 30-day downloads of Granite 4.2 8B, which was the least downloaded entry when the swap was proposed on 4 November. Correction, 16 November: the mailed issue called Granite 4.2 8B the least downloaded entry without that date. At the last reading before the change, on 15 November at 18:15 UTC, Devstral Small 2 24B stood lower, at 26,373 30-day downloads against 59,920. Since 16 November the rule checks a swap again on the day it applies and lets it expire when its reason no longer holds.

The "What should I use?" lists are kept by a published rule, checked every Monday. A model enters when it has open weights, runs in a common tool, counts among the most downloaded of its kind on the Hub and was released within twelve months. A successor from the same maker in the same size replaces its predecessor at once; a newcomer with more than twice the downloads of the least downloaded entry replaces it after a seven-day veto window; three Mondays of fading in a row propose a removal.

Each list holds at most five models and the page shows three, ordered for the reader's own card by the measured file and the fit word. There is no score on the page: the order says what fits, not what is better.

A change is not a verdict. A replaced model keeps its page, its measurements and its place in the pick rule; it has only left a list that the rule keeps short on purpose. The rule itself is on the methodology page, and every change carries the reason that made it.

What the pick costs

Qwen3 14B, three ways, per day

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

WayPer dayAssumptions
Through an API$0.08OpenRouter list price: $0.12 in, $0.24 out per million tokens
On your own RTX 4090$2.01$1,599 launch price over 3 years plus 550 W at $0.25/kWh
Rented RTX 3090$0.97$0.24 an hour at the Vast.ai median, 4 hours a day
Recalculate with your own usage →

Newly measured

Files measured this week, and where they fit

  • Qwen3 32B (Q4_K_M, 18.4 GB): fits no consumer card comfortably at 8K.
  • GLM-4.7-Flash (Q4_K_M, 17.1 GB): fits a 24 GB card comfortably at 8K.
  • Gemma 4 31B (Q4_K_M, 17.1 GB): fits no consumer card comfortably at 8K.
  • Granite 4.2 30B (Q4_K_M, 16.5 GB): fits no consumer card comfortably at 8K.
  • Gemma 4 26B-A4B (Q4_K_M, 15.8 GB): fits a 24 GB card comfortably at 8K.
  • Qwen3.8 27B (Q4_K_M, 15.3 GB): fits a 24 GB card comfortably at 8K.

Memory from the measured GGUF file plus the computed context cache; "runs" means our verdict is "Runs easily" or "Runs". Check your own card

First reading

Models scored for the first time

  • LFM3 1.7B (Liquid AI): first scored on 9 Nov, now at 70.

Rent or buy, one number

3,198 rented hours

At last week's Vast.ai median of $0.50 an hour, the RTX 4090's launch list price of $1,599 (NVIDIA launch MSRP, Oct 2022) equals 3,198 rented hours, ≈ 8 years at 8 h/week. Electricity and resale value are left out on both sides.

The break-even, with your own hours →

Class of the week

H100

Last week's Vast.ai median: $2.94 an hour. Against 30 days ago: +26 %. Since we started reading it 46 days ago: +17 %. Cheapest source at the week's end: RunPod at $2.59 (list price). The cheapest big cloud, AWS, listed 2.4× the marketplace median. 34 of 49 tracked open models run comfortably on one at 8K context; the largest comfortable one is gpt-oss 120B.

Everything on the H100 →

Maker board

Which labs hold the top-10 slots

LabTop-10 slotsMean Heat7-day change
Alibaba Qwen
best placed: Qwen3-Coder 30B-A3B
3 55 −3
Meta
best placed: Llama 3.1 8B
2 66 +7
Ornith AI
best placed: Ornith 1.5 35B-A3B
2 62 +10
OpenAI
best placed: gpt-oss 120B
1 70 +6
Zhipu AI
best placed: GLM-5.3-Flash
1 56 −4

Counted from the popularity 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.

What should I use

The lists this week

  • Write & chat: Gemma 4 26B-A4B in place of Ornith 1.5 9B (meets the entry rule with more than 2× the 30-day downloads of Ornith 1.5 9B, the least downloaded entry), 2026-11-16.
  • Long documents: Gemma 4 26B-A4B in place of Ornith 1.5 9B (meets the entry rule with more than 2× the 30-day downloads of Ornith 1.5 9B, the least downloaded entry), 2026-11-16.
  • Code: Ornith 1.5 9B in place of Granite 4.2 8B (meets the entry rule with more than 2× the 30-day downloads of Granite 4.2 8B, the least downloaded entry), 2026-11-16.

Kept by a published rule, checked 2026-11-16; a person looked over every change. The lists on your card

From the tower

What changed on the site

  • On cooldown and therefore not eligible as the next pick: Qwen3 14B (picked 2026-w46), Granite 4.2 3B (picked 2026-w45), Gemma 4 E4B (picked 2026-w44), Qwen3 8B (picked 2026-w43). The rule

From the methodology desk

What this issue does not say

List changes are the rule's output after a person's veto window. What the rule cannot judge, which is how good a model is at the job, it does not judge.

Forward this to someone with a graphics card. They can subscribe at aicomputeradar.dev/newsletter. Replies reach a person.

The Tower Report · every Monday

Monday's board, in your inbox.

Prices, the pick, what newly fits your card. Measured, never forecast. One mail a week, one click to leave.

Name your card and every issue opens with what it runs this week. You can change it by subscribing again with the same address.

Double opt-in; you confirm by one click. Details in the privacy policy · what is in an issue

Keep reading