Gate: the smallest device it runs on easily at 32K context. Seat: the measured GGUF file.
Pick of the week — Week 47, 2026 · selected by rule, written by people
The first model picked twice: same 6.6 GB file, half the downloads
Gemma 4 12B by Google takes the eleventh pick and is the first model to be picked twice: it was the first pick, in week 37. Between the readings of 14 and 20 November it had the largest popularity gain among models that run comfortably on a 24 GB card and are not on cooldown, and it is the smallest gain that has won a pick so far. Its Q4_K_M file measures 6.6 GB, read from Unsloth's GGUF repository. At 8K context it needs 8.1 GB: it runs on the 12 GB cards and runs easily from 16 GB up. At 32K it needs 9.2 GB and still runs on 12 GB.
Measured signals as of 2026-11-20 18:15 UTC · Hugging Face + Vast.ai · method on /methodology
Where it runs
Fit engine · measured GGUF + computed KV cache · usable memory after marginsOr rent an RTX 4090 from $0.49/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 14 and 20 November, both at 18:15 UTC, nine models gained more than Gemma 4 12B and two gained as much, and none of the eleven counts. GLM-5.2 gained the most of all; its Q4 file is 433.8 GB. Gemma 4 E4B had the second largest gain and is on cooldown as the pick of week 44, and Qwen3 8B, with the same gain as this pick, is on cooldown from week 43. KAT-Coder V2.5 and Ornith 1.5 35B-A3B have files of 19.9 GB and 20.2 GB that spill into RAM on a 24 GB card at 8K, and Qwen3.8 2.4T-A95B and DeepSeek-V4-Pro are far beyond any consumer card. Devstral Small 2 24B, Mistral Small 3.2, LFM2.5 1.2B and Granite 4.2 8B fit, but they stay under the score floor, so their rise does not count. The runner-up is gpt-oss 20B, a 10.8 GB file that needs 12.2 GB at 8K and runs on the 16 GB cards; it has never been picked. Qwen3 4B gained as much as gpt-oss 20B and stands behind it on the tie-break.
The gain is small and the report should say what it is made of. It is less than half of the smallest gain that had won a pick before, and the model's own signals fell over the window: its 30-day downloads went from 1,676,006 on 14 November to 1,579,383 on 20 November, down 5.8 percent, and its Hugging Face trending score went from 9 to 8, which moved it from a tied sixteenth to a tied twentieth place among the 49 models we track. What improved is the direction, which the score counts on its own. On 14 November the model stood 9.2 percent below its downloads of a week earlier, and its trending score had fallen from 13 to 9 in those seven days; on 20 November it stood 6.7 percent below the week before, and its trending score was 8 against 7. A fall that slows ranks higher among models that are falling faster, and 31 of the 49 lost 30-day downloads over the window. The score still stands below its level of 7 November. The first pick, published on 11 September, froze 3,083,702 downloads in 30 days, a trending score of 19 and 1,547 likes; today's card freezes 1,579,383 downloads, a trending score of 8 and 1,753 likes. By 30-day downloads the model is eighteenth among the 49. It was released on 23 May 2026 under Apache 2.0.
What we measure is the file and the shape, and neither has changed since September. The model carries 12.0 billion parameters, dense, and the Q4_K_M weights weigh 6.6 GB. The config publishes 48 layers, and only 8 of them cache the full context, each with a single KV head of 512 dimensions; the other 40 are sliding-window layers that keep 1,024 tokens. That is why the cache stays small as the context grows: 0.44 GB at 8K, 0.81 GB at 32K, 2.31 GB at 128K and 4.31 GB at the full window of 262,144 tokens. At 8K the engine computes 8.1 GB in total. It runs easily on the 16 GB and 24 GB cards and runs on the 12 GB cards, which have 10.6 GB usable, and on the 16 GB Mac mini with 10.4 GB usable; the 10 GB RTX 3080 barely fits it, and the 8 GB cards, with 7.0 GB usable, spill into RAM. At 32K the total is 9.2 GB, and the 12 GB cards and the 16 GB Mac mini still run it. At 128K it is 13.7 GB: the 24 GB cards run it easily, the 16 GB cards barely fit it with 14.1 GB usable, and the 12 GB cards spill. At the full window the total is 19.7 GB, of which 8.8 GB is our assumed runtime overhead at that length: the RTX 5090 runs it easily, the 24 GB cards barely fit it, and for the 24 GB Mac mini, with 15.6 GB usable, it is too big. Qwen3 14B, last week's pick, needs 15.3 GB at 32K. This model, with a file 1.8 GB smaller, needs 9.2 GB, because the 14B caches the full context in all 40 of its layers and this one in 8 of its 48.
Renting is not needed for this pick at ordinary context lengths: a 12 GB card that is already in your machine runs it at 8K and at 32K. The case for a rented card starts at the long windows, where 128K wants more than 16 GB and the full window wants 24 GB or more. The pick card freezes an RTX 4090 hour on the Vast marketplace at $0.49, the 18:00 UTC hourly median across 64 offers, and RunPod's posted price for the card was $0.34; when this model was first picked on 11 September, the same card froze $0.67. The RTX 3090 class has the same 24 GB: at 18:00 UTC the Vast marketplace median for an RTX 3090 hour was $0.25 across 64 offers, and RunPod's posted price was $0.22. Two versions made by other people are listed on the model's page, each measured on its own: an abliterated version by huihui-ai with a 6.9 GB file, and Winnow-12B, which is only offered as an 8-bit file of 11.8 GB.
- Maker
- Parameters
- 12B
- Measured GGUF
- Q4_K_M · 6.6 GB
- Max context
- 256K tokens
- Curated for
- general, research
- Repo created
- 2026-05-23
- Trending score
- 8
- 7-day downloads
- -6.7%
Model pageHugging Face repo (External link)GGUF file (External link)
Downloads, daily readingsdaily last · UTC
2026-11-07 · 1.8M2026-11-20 · 1.6M
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 $0.495 and last week's $0.503. Hourly medians move with which offers happen to be listed.
- This is the smallest gain that has won a pick, and the model's own signals fell over the window: 30-day downloads lost 5.8 percent and the Hugging Face trending score went from 9 to 8. The gain comes from the momentum parts of the score, where a fall that slows ranks above falls that continue, in a week in which 31 of the 49 tracked models lost 30-day downloads. The popularity score is percentile-based by design.
- A model can be picked again once four weeks have passed since its last pick, and this is the first time the rule has done so. The report and the numbers of the first pick, in week 37, stay frozen on their own page and were not changed.
- Popularity measures attention on the Hub (downloads, trending, likes), not quality. We publish no benchmarks and no speed figures for this or any model.
- The totals include runtime overhead (1.0 GB at 8K, 1.8 GB at 32K, 4.8 GB at 128K, 8.8 GB at 262,144 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 2026-11-14 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 2026-11-20 18:40 UTC
Issues
- Week 47, 2026 · Gemma 4 12B2026-11-20
- Week 46, 2026 · Qwen3 14B2026-11-13
- Week 45, 2026 · Granite 4.2 3B2026-11-06
- Week 44, 2026 · Gemma 4 E4B2026-10-30
- Week 43, 2026 · Qwen3 8B2026-10-23
- Week 42, 2026 · LFM2.5 8B-A1B2026-10-16
- Week 41, 2026 · Ornith 1.5 9B2026-10-09
- Week 40, 2026 · Granite 4.2 8B2026-10-02
- Week 39, 2026 · Mistral Small 3.22026-09-25
- Week 38, 2026 · Gemma 4 26B-A4B2026-09-18
- Week 37, 2026 · Gemma 4 12B2026-09-11
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.