An eighteen-month-old 8B climbs the trending list

What it is for

The rule is unchanged: the largest seven-day Heat Score 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 17 and 23 October, both at 18:15 UTC, Qwen3 8B rose from 45 to 65. Qwen3 4B gained the same 20 points and lost the tie on score, 50 against 65. Three models moved more and were out on the rule: DeepSeek-V4-Flash gained 28 and Ornith 1.5 397B gained 23, neither fits a 24 GB card at 8K; Granite 4.2 3B gained 24 but stands at 43 today, below the 50 a counted rise requires.

The gain is a trending gain, not a download gain, and the report should say so. The Hub's trending score for Qwen3 8B went from 6 on 17 October to 14 on 23 October, which put it in the top two percent of tracked models on that component. Its 30-day downloads fell over the same window, from 9.34 million to 8.08 million, down 13 percent over the window. The model was released in April 2025; what moved this week is attention on the Hub, and the rule counts attention. We do not know why, and the data does not say.

What we measure is the file and the shape. Qwen3 8B is a dense model of 8.2 billion parameters, and the Q4_K_M weights weigh 4.7 GB. Its 36 layers all cache the full context, 8 KV heads at 128 dimensions each, so the cache grows with every token: 1.1 GB at 8K, 4.5 GB at 32K. At 8K context the engine computes 6.8 GB in total: TIGHT on the 8 GB cards (6.8 of 7.0 usable), GOOD on the RTX 3080 and RTX 2080 Ti, EXCELLENT from 12 GB upward, the RTX 3060 12GB, 4070 and 5070 included. At 32K the total is 10.9 GB, and the 12 GB cards miss it by less than half a gigabyte (10.9 against 10.6 usable): OFFLOAD_REQUIRED on everything up to 12 GB, GOOD on the 16 GB cards and the Mac mini M5 Pro 24GB, EXCELLENT from 24 GB. The config publishes a window of 40,960 tokens. Last week's pick, LFM2.5 8B-A1B, needs 0.38 GB of cache at 32K where this one needs 4.5; the two files are within 0.1 GB of each other.

Renting a 24 GB card for a model that needs 6.8 GB at 8K is the wrong tool, and at 32K it is still twice the memory the job needs. The pick card froze an RTX 4090 hour on Vast.ai at $1.00, the 18:00 UTC hourly median across 37 offers; this week's median for the class was $0.60 and last week's $0.53. A 12 GB card that is already in your machine runs it at 8K, a 16 GB card at 32K. If you do rent for 32K work, the RTX 3090 class has the same 24 GB: at 16:00 UTC RunPod posted $0.22 per hour and the Vast.ai median was $0.35 across 35 offers.

Watch out

• The rent figure on this card is the 18:00 UTC hourly Vast.ai median for the RTX 4090 (37 offers); this week's median for the class was $0.60 and last week's $0.53. Hourly medians move with which offers happen to be listed.
• This week's rise came from the Hub's trending score (6 to 14) while 30-day downloads fell 13 percent; the Heat Score weights both, and a trending rise alone can move it by this much.
• Heat Score measures attention on the Hub (downloads, trending, likes), not quality. We publish no benchmarks and no speed figures for this or any model.
• The 40,960-token window is the model's published config. The maker's model card describes a longer window with YaRN scaling, which we do not measure; the cache figures above stop at 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.