Skip to content
AI Compute Radar
IndexCollected 4 h ago

Gate: the smallest device it runs on easily at 32K context. Seat: the measured GGUF file.

Pick of the week — Week 46, 2026 · selected by rule, written by people

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

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.

Measured signals as of 2026-11-13 18:15 UTC · Hugging Face + Vast.ai · method on /methodology

Where it runs

Fit engine · measured GGUF + computed KV cache · usable memory after margins
Device8K context32K context
RTX 30708 GB · usable 7 GBSpills into RAM · 10.8 GBSpills into RAM · 15.3 GB
RTX 40608 GB · usable 7 GBSpills into RAM · 10.8 GBSpills into RAM · 15.3 GB
RTX 50608 GB · usable 7 GBSpills into RAM · 10.8 GBSpills into RAM · 15.3 GB
RTX 3060 Ti8 GB · usable 7 GBSpills into RAM · 10.8 GBSpills into RAM · 15.3 GB
RTX 308010 GB · usable 8.8 GBSpills into RAM · 10.8 GBSpills into RAM · 15.3 GB
RTX 2080 Ti11 GB · usable 9.7 GBSpills into RAM · 10.8 GBSpills into RAM · 15.3 GB
RTX 407012 GB · usable 10.6 GBSpills into RAM · 10.8 GBSpills into RAM · 15.3 GB
RTX 3060 12GB12 GB · usable 10.6 GBSpills into RAM · 10.8 GBSpills into RAM · 15.3 GB
RTX 507012 GB · usable 10.6 GBSpills into RAM · 10.8 GBSpills into RAM · 15.3 GB
RTX 4070 Super12 GB · usable 10.6 GBSpills into RAM · 10.8 GBSpills into RAM · 15.3 GB
RTX 4070 Ti12 GB · usable 10.6 GBSpills into RAM · 10.8 GBSpills into RAM · 15.3 GB
RTX 3080 Ti12 GB · usable 10.6 GBSpills into RAM · 10.8 GBSpills into RAM · 15.3 GB
RTX 408016 GB · usable 14.1 GBRuns · 10.8 GBSpills into RAM · 15.3 GB
RTX 508016 GB · usable 14.1 GBRuns · 10.8 GBSpills into RAM · 15.3 GB
RTX 4060 Ti 16GB16 GB · usable 14.1 GBRuns · 10.8 GBSpills into RAM · 15.3 GB
RTX 5060 Ti 16GB16 GB · usable 14.1 GBRuns · 10.8 GBSpills into RAM · 15.3 GB
RTX 5070 Ti16 GB · usable 14.1 GBRuns · 10.8 GBSpills into RAM · 15.3 GB
Mac mini M6 16GB16 GB · usable 10.4 GBToo big · 10.8 GBToo big · 15.3 GB
RTX 4070 Ti Super16 GB · usable 14.1 GBRuns · 10.8 GBSpills into RAM · 15.3 GB
RX 9070 XT16 GB · usable 14.1 GBRuns · 10.8 GBSpills into RAM · 15.3 GB
RTX 309024 GB · usable 21.1 GBRuns easily · 10.8 GBRuns · 15.3 GB
RTX 409024 GB · usable 21.1 GBRuns easily · 10.8 GBRuns · 15.3 GB
Mac mini M5 Pro 24GB24 GB · usable 15.6 GBRuns easily · 10.8 GBBarely fits · 15.3 GB
RX 7900 XTX24 GB · usable 21.1 GBRuns easily · 10.8 GBRuns · 15.3 GB
RTX 3090 Ti24 GB · usable 21.1 GBRuns easily · 10.8 GBRuns · 15.3 GB
RTX 509032 GB · usable 28.2 GBRuns easily · 10.8 GBRuns easily · 15.3 GB
Mac Studio M5 Max 36GB36 GB · usable 23.4 GBRuns easily · 10.8 GBRuns easily · 15.3 GB
Mac mini M4 Pro 64GB64 GB · usable 44.8 GBRuns easily · 10.8 GBRuns easily · 15.3 GB
Mac Studio M4 Max 64GB64 GB · usable 44.8 GBRuns easily · 10.8 GBRuns easily · 15.3 GB
Mac Studio M3 Ultra 96GB96 GB · usable 67.2 GBRuns easily · 10.8 GBRuns easily · 15.3 GB
Mac Studio M5 Ultra 96GB96 GB · usable 67.2 GBRuns easily · 10.8 GBRuns easily · 15.3 GB
NVIDIA DGX Spark128 GB · usable 112.6 GBRuns easily · 10.8 GBRuns easily · 15.3 GB

Or rent an RTX 4090 from $0.52/h →Check it against your own machine →

What it is for

The rule is unchanged: the largest seven-day popularity gain among models that fit a 24 GB card with margin and were not picked in the last four weeks; ties break by the higher score. Between the readings of 7 and 13 November, both at 18:15 UTC, Qwen3 14B gained more than any other model the rule counts. Three models gained more and were out. Granite 4.2 30B gained the most of all, but its 16.5 GB file barely fits a 24 GB card at 8K, which the rule does not count. Llama 3.2 3B has a gated config, so the engine has no attention shape and gives no verdict, and a model without a verdict is not a fit. Nemotron Nano 9B v2 runs easily on 24 GB, but it stays under the score floor, so its rise does not count. Qwen3-Coder 30B-A3B gained almost as much as the pick and is out for the same reason as Granite: a 17.3 GB file that barely fits. Gemma 4 26B-A4B is the runner-up, a 15.8 GB file that runs on a 24 GB card at 8K; it was picked in week 38 and is long off cooldown.

The signals behind the gain are a recovery more than a surge, and the report should say so. The baseline of 7 November was this model's lowest reading since 28 October: its score had fallen on all but one day until then, turned on 8 November, and today still stands below where it was on 31 October. Its 30-day downloads went from 12,952,853 on 7 November to 13,011,347 on 13 November, up 0.5 percent; its Hugging Face trending score went from 1 to 4, which moved it from a tied fortieth to a tied twenty-eighth place among the 49 models we track. The popularity score ranks every component against the other tracked models, and 32 of the 49 lost 30-day downloads over this window, so a model that held its downloads and came off the bottom of the trending column rose past many that slipped. By 30-day downloads it is the second most downloaded model we track, behind Qwen3 0.6B. Released on 27 April 2025 under Apache 2.0, the same day as Qwen3 8B, it is the second Qwen3 pick after the 8B in week 43; the two share the oldest release date among the ten picks.

What we measure is the file and the shape. The model carries 14.8 billion parameters, dense, and the Q4_K_M weights, read from Qwen's GGUF repository, weigh 8.4 GB. The config publishes 40 layers that all cache the full context, with 8 KV heads of 128 dimensions, and a window of 40,960 tokens, which is where our readings end: 1.25 GB of cache at 8K, 5.0 GB at 32K, 6.25 GB at 40,960 tokens. At 8K the engine computes 10.8 GB in total. It runs easily on the 24 GB cards, which have 21.1 GB usable, and runs on the 16 GB cards with 14.1 GB usable; the 12 GB cards have 10.6 GB usable, miss it by 0.2 GB and spill into RAM, and for the 16 GB Mac mini, with 10.4 GB usable, it is too big. At 32K the total is 15.3 GB: the 24 GB cards run it, the RTX 5090 runs it easily, the 16 GB cards spill 1.2 GB into RAM, and the 24 GB Mac mini barely fits it with 15.6 GB usable. At the full 40,960 tokens the total is 16.8 GB, of which 2.1 GB is our assumed runtime overhead at that length: it runs on 24 GB and runs easily on the RTX 5090. Qwen3 8B, the pick of week 43, has the same heads and 36 layers instead of 40, so its cache at 32K is 4.5 GB against 5.0 GB here. The difference between the two is the file, 4.7 GB against 8.4 GB: the 14B at 8K needs 10.8 GB, about what the 8B needs at 32K, 10.9 GB.

Renting has a case here that the last two picks did not have: 32K and the full window want 24 GB, and many machines stop at 12 or 16. The RTX 3090 class has the 24 GB: at 18:00 UTC the Vast marketplace median for an RTX 3090 hour was $0.24 across 57 offers, and RunPod's posted price was $0.22. The pick card freezes an RTX 4090 hour on the Vast marketplace at $0.52, the 18:00 UTC hourly median across 64 offers; this week's median for the class was $0.50 and last week's $0.48, and RunPod's posted price for the card was $0.34. If the job fits in 8K, a 16 GB card that is already in your machine runs it and no rental is needed.

Maker
Alibaba Qwen
Parameters
14.8B
Measured GGUF
Q4_K_M · 8.4 GB
Max context
40K tokens
Curated for
general, coding, research
Repo created
2025-04-27
Trending score
4
7-day downloads
+1.2%

Model pageHugging Face repo (External link)GGUF file (External link)

Downloads, daily readingsdaily last · UTC

2026-10-31 · 13.0M2026-11-13 · 13.0M

Watch out

  • The rent figure on this card is the 18:00 UTC hourly Vast marketplace median for the RTX 4090 (64 offers); this week's median of daily medians for the class was $0.50 and last week's $0.48. Hourly medians move with which offers happen to be listed.
  • This week's gain is a recovery from a low baseline: the reading of 7 November was the model's lowest since 28 October, and today's score is still below its level of 31 October. Downloads rose 0.5 percent and the Hugging Face trending score went from 1 to 4, while 32 of the 49 tracked models lost 30-day downloads over the window. The popularity score is percentile-based by design, and a quiet week for the field lifts the models that merely held.
  • Popularity measures attention on the Hub (downloads, trending, likes), not quality. We publish no benchmarks and no speed figures for this or any model.
  • Our readings end at 40,960 tokens, the window the config publishes, so there is no 128K figure for this model. The totals include runtime overhead (1.1 GB at 8K, 1.9 GB at 32K, 2.1 GB at 40,960 tokens), which is an assumption of our engine, not a measurement; the weights are measured, the cache is computed from the config.
  • Fit verdicts assume llama.cpp-style inference with the measured Q4_K_M GGUF and an f16 cache; other runtimes, quantized caches and other quantizations change the numbers.

How we chose it

Pick of the week is the model with the largest Heat Score rise over the last seven days among tracked models that run comfortably (GOOD or EXCELLENT) on at least one consumer card of up to 24 GB at 8K context. A rise counts only when it is at least three points, the model's Heat Score is at least 50 today, and the model's own direction is not falling — a low score that climbs while the models around it fall has only moved because they moved. Ties break by the higher Heat Score, then by 30-day downloads. A model cannot be picked again within four weeks. While the heat history is younger than a week, the rise is measured over the readings that exist and the report says so; when no rise counts, the rule falls back to the highest Heat Score and records that.

This issue was selected on the heat rise. Rise measured since 2026-11-07 18:15 UTC. 49 scored models were considered; the full list with reasons is stored with the issue.

The rule on the methodology page · rule v1.2 · selected 2026-11-13 18:42 UTC

Issues

A new pick every Friday. The model is computed, the text is written and released by a person, and every number links back to where it was measured.

RSS feed of every issue · JSON: /api/v1/pick

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