Gate: the smallest device with an EXCELLENT fit at 32K context. Seat: the measured GGUF file.
Pick of the week — Week 37, 2026 · selected by rule, written by people
The 12B people came back to
Gemma 4 12B is this week's pick by rule: the largest counted Heat Score rise among tracked models that run comfortably on a consumer card, from 45 to 62 since the evening of 7 September. Its trending score is back where it stood in late August, its measured Q4_K_M file is 6.6 GB, and it runs with room to spare on any 16 GB card at 8K context.
Measured signals as of 2026-09-11 18:00 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 $0.67/h →Check it against your own machine →
What it is for
Why now: the climb is the model's own. Gemma 4 12B's trending score had slid from 20 at the end of August to 11 in the first week of September, then recovered to 21 on 10 September and 19 today, while 30-day downloads eased from 3.3 to 3.1 million. The Heat Score followed, from 45 on 7 September to 62 at selection. The rule does not ask why; the readings describe a model people drifted away from for a week and came back to. Granite 4.2 8B (+14) and Qwen3 8B (+7) were the runners-up.
Where it runs: the measured Q4_K_M file is 6.6 GB. With the context cache and the runtime overhead the engine puts the total at 8.5 GB at 8K context and 10.7 GB at 32K, computed from the published attention shape rather than guessed from the parameter count. That is EXCELLENT on every 16 GB and 24 GB card we track, on every Mac and on the DGX Spark, and GOOD on the 12 GB cards at 8K. If you own nothing suitable, an RTX 4090 hour rented at the median cost $0.67 today.
What it is for: a 12B dense model with Google's open weights, curated here for general and research work, with a 256K context window on paper. The window is the reason to look at it and the thing to plan around: the cache grows with every token you actually use, and the fit ladder shows where a 12B stops being small.
- 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
- 19
- 7-day downloads
- -3.5%
Model pageHugging Face repo (External link)GGUF file (External link)
Downloads, daily readingsdaily last · UTC
2026-08-29 · 3.3M2026-09-11 · 3.1M
Watch out
- The rise is a recovery, not a breakout: the trending score fell to 11 in the first week of September and is back at 19 to 21, the level of late August. Thirty-day downloads are still 3.5% lower than a week ago.
- 12 GB cards are GOOD at 8K only; at 32K the engine already asks for offloading.
- The 256K window is a spec, not a plan. At 64K the engine estimates 13.7 GB in total (EXCELLENT on 24 GB, TIGHT on 16 GB); at 128K it is 19.7 GB, TIGHT even on a 24 GB card and comfortable only on 64 GB of unified memory.
- One measured GGUF, Q4_K_M from unsloth. Other quantizations change every number above.
- No benchmarks and no speed figures: whether it is good at your task is yours to test. We only know that it fits and that people are coming back to it.
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 2026-09-07 18:15 UTC. 45 scored models were considered; the full list with reasons is stored with the issue.
The rule on the methodology page · rule v1.2 · selected 2026-09-11 18:20 UTC
Issues
- 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.