A 2 GB file that runs on everything we track

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 31 October and 6 November, both at 18:15 UTC, Granite 4.2 3B gained more than any other eligible model. Qwen AgentWorld 35B-A3B gained almost as much, but its 20.6 GB file spills into RAM on a 24 GB card at 8K, so it is out. GLM-4.7-Flash is the runner-up: its 17.1 GB file runs on a 24 GB card at 8K. Qwen3.8 27B, a 15.3 GB file that also runs on 24 GB, gained the third most, Ling 3.0 Tiny the fourth. LFM2.5 8B-A1B was picked in week 42 and sits on cooldown until week 47; Qwen3 32B gained but stays under the score floor.

The signals behind the gain are a trending move more than a download move, and the report should say so. Its 30-day downloads went from 35,342 on 31 October to 37,286 on 6 November, up 5 percent; its Hugging Face trending score went from 4 to 11 in the same week, which put it in the top ten of the 48 models we track. The popularity score ranks every component against the other tracked models, and 31 of the 48 lost 30-day downloads over this window, so a small model that held and trended rose past larger ones that merely held. Released on 7 August 2026 under Apache 2.0, it is the first Granite pick since Granite 4.2 8B in week 40, and the smallest model picked so far.

What we measure is the file and the shape. The model carries 3.66 billion parameters, dense, and the Q4_K_M weights, read from IBM's GGUF repository, weigh 2.1 GB. The config publishes 40 layers that all cache the full context, with 8 KV heads of 64 dimensions, a narrow cache for a dense model: 0.6 GB at 8K, 2.5 GB at 32K, 10.0 GB at the full 131,072-token window. At 8K the engine computes 3.6 GB in total, which runs easily on everything we track, from the 8 GB cards and the 16 GB Mac mini up. At 32K the total is 6.3 GB: the 8 GB cards run it with 7.0 GB usable, and from the RTX 2080 Ti at 11 GB upward it runs easily. At the full 128K window the total is 16.8 GB, of which 4.7 GB is our assumed runtime overhead at that length: it runs on the 16 GB cards and runs easily on 24 GB. Last week's pick, Gemma 4 E4B, is a 4.6 GB file that needs 7.2 GB at 32K; this one needs 6.3 GB from half the file, because its cache is larger and its weights are smaller.

Renting for this model is the wrong tool almost everywhere: an 8 GB card you already own runs it at 32K. If the job is the full 128K window on a card you do not have, the RTX 3090 class has the 24 GB it wants: at 18:00 UTC the Vast marketplace median for an RTX 3090 hour was $0.23 across 64 offers, and RunPod's posted price was $0.22. The pick card freezes an RTX 4090 hour on the Vast marketplace at the 18:00 UTC hourly median; this week's median for the class was $0.48 and last week's $0.53.

Watch out

• The rent figure on this card is the 18:00 UTC hourly Vast marketplace median for the RTX 4090; this week's median of daily medians for the class was $0.48 and last week's $0.53. Hourly medians move with which offers happen to be listed.
• This week's gain is a trending-score move more than a download move: downloads rose 5 percent while the Hugging Face trending score went from 4 to 11, and 31 of the 48 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.
• The 128K figure includes 4.7 GB of runtime overhead, which is an assumption of our engine for that context length, 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.