Five gigabytes of weights, context at full price

What it is for

Granite 4.2 8B is an 8.8-billion-parameter model from IBM with open weights, curated here for general, coding and agent work. The repository dates from 7 August 2026, IBM's own Granite organisation publishes the GGUF build we measure, and Ollama carries it in its library as granite4.2. On 2 October it stood at 127,613 downloads over the last 30 days, up 16 % week on week, with 92 likes on Hugging Face and a trending score of 8.

The move is a rebound. The Heat Score had slid from 50 on 20 September to 35 on the 26th while the trending score fell to 1; then it turned: 3, 4, 5, 5, 8 and 8 on the following evenings, with downloads going from 112,993 to 127,613 inside the window. That took the score to 57. The whole family moved in the same days: Granite 4.2 3B rose 12 points to 46, short of the 50 a model needs for its rise to count, and Granite 4.2 30B rose 5 to 39 but fits no consumer card. The runners-up were Qwen3.8 27B and LFM2.5 2.6B with a rise of 11 each. GLM-5.3 gained 13 points to 86 and Gemma 4 31B 8 to 84, but neither runs comfortably on a card of up to 24 GB. Last week's pick, Mistral Small 3.2, sits at 45, seven points below where it was picked.

The weights are the small part. The Q4_K_M file is 5.0 GB, but all 40 layers keep a full attention cache, so context is charged at full price: 1.25 GB for every 8K tokens, computed from the published attention shape. With runtime overhead the engine puts the total at 7.3 GB at 8K and 11.8 GB at 32K, where the cache is as large as the weights. That is EXCELLENT at 8K on every 12 GB card, from the RTX 3060 12GB to the RTX 4070 Ti, with 3.3 GB to spare, and GOOD on the RTX 3080 and the RTX 2080 Ti. At 32K the 12 GB cards are 1.2 GB short and have to offload; the 16 GB cards carry it with 2.3 GB to spare; the 24 GB cards, the RTX 5090, the DGX Spark and every Mac from 24 GB up are comfortable at both lengths. Of the 32 machines we track, 28 run it comfortably at 8K and 19 at 32K. The model accepts up to 131,072 tokens of context, which would take 29.8 GB. Renting instead of buying: an RTX 4090 cost $0.46 per hour at the median of verified Vast.ai offers on 2 October at 18:00 UTC.

Watch out

• Not an 8 GB card model at this quantization: 7.3 GB at 8K against 7.0 GB usable means the RTX 3070, RTX 3060 Ti, RTX 4060 and RTX 5060 have to offload part of the weights into system RAM. It runs, but not in the way a fit verdict means it.
• Long context is where the memory goes: 32K takes every 12 GB card and the Mac mini M6 16GB out of the comfortable range, 64K needs 17.8 GB and 128K needs 29.8 GB.
• The rise starts from a trough. The Heat Score stood at 51 on 7 September, peaked at 65 on the 11th and slid to 35 by the 26th; this week brought it back to 57, not to new ground.
• The numbers behind the move are small: a trending score that went from 1 to 8, and 92 likes. All three Granite 4.2 models rose in the same days; our readings do not say why, and the rule does not ask.
• One measured GGUF, Q4_K_M from IBM; other quantizations change every number above. No benchmarks and no speed figures: whether it is good at your task is yours to test.
• The rental price is one marketplace at one hour; it moves during the day.