Five gigabytes again, context at a fifth of the price

What it is for

Ornith 1.5 9B is a 9.7-billion-parameter model from Ornith AI with open weights, curated here for general and coding work. The repository dates from 18 August 2026, Ornith AI's own organisation publishes the GGUF build we measure, and Ollama carries the family in its library as ornith. On 9 October it stood at 466,339 downloads over the last 30 days, up 1.8 % week on week, with 369 likes on Hugging Face and a trending score of 16. It accepts up to 262,144 tokens of context, and only 8 of its 32 layers keep a full attention cache; the other 24 hold a state that does not grow with context.

The move is a rebound, like last week's. The Heat Score had slid from 64 on 21 September to 40 on the 29th and stood at 45 on the evening of 3 October; then it went 47, 55, 60, 60, 64 and 67 on the following evenings. The readings behind it are mixed: the trending score went 15, 14, 18, 19, 20, 18 and 16, and downloads over 30 days fell from 455,080 to 449,256 by 7 October before rising to 466,339 in two evenings. The family moved together again: Ornith 1.5 35B-A3B rose 13 to 60, but its 20.2 GB file leaves no comfortable room on a 24 GB card, and Ornith 1.5 397B rose 8 to 23, far below the 50 a model needs. The runners-up were GLM-5.2 with a rise of 20 to 55, and Qwen3.8 Flash Next and Llama 3.1 8B with 15 each; none of the three gets a comfortable fit verdict on a card of up to 24 GB, Llama 3.1 8B because we hold no attention shape for it and the engine does not guess. Gemma 4 12B rose 13 to 55 but was picked in week 37 and is still on cooldown. Qwen3 14B, up 8 to 65, was the only other model that passed every test. Last week's pick, Granite 4.2 8B, sits at 41, sixteen points below where it was picked.

The weights are again the small part, and this time the context is cheap too. The Q4_K_M file is 5.4 GB, and because only 8 of 32 layers cache the whole context, every 8K tokens cost 0.25 GB, computed from the published attention shape; last week's Granite 4.2 8B charged 1.25 GB for the same 8K with all 40 layers caching. With runtime overhead the engine puts the total at 6.7 GB at 8K, 8.2 GB at 32K, 10.2 GB at 64K and 14.2 GB at 128K. That is EXCELLENT at 8K on every 12 GB card, from the RTX 3060 12GB to the RTX 4070 Ti, with 3.9 GB to spare, and still GOOD at 32K; the RTX 3080 is GOOD and the RTX 2080 Ti EXCELLENT at 8K. The 16 GB cards are comfortable up to 64K, the 24 GB cards up to 128K with 6.9 GB to spare, and the full 262,144 tokens take 22.2 GB, GOOD on the RTX 5090, TIGHT on the 36 GB Mac Studio and comfortable on the 64 GB Macs and the DGX Spark. Of the 32 machines we track, 28 run it comfortably at 8K and 27 at 32K. Renting instead of buying: an RTX 4090 cost $0.49 per hour at the median of 64 verified Vast.ai offers on 9 October at 20:00 UTC and an RTX 3090 $0.27; RunPod listed the same cards at $0.34 and $0.22 at that hour.

Watch out

• Not an 8 GB card model without care: 6.7 GB at 8K against 7.0 GB usable leaves 0.3 GB, which the engine calls TIGHT on the RTX 3070, RTX 3060 Ti, RTX 4060 and RTX 5060, and 32K already spills 1.1 GB into system RAM.
• The cheap context has a ceiling: 128K takes 14.2 GB, 0.1 GB more than a 16 GB card has usable, and the full 256K takes 22.2 GB, 1.1 GB more than an RTX 4090. A 24 GB card stops at 128K.
• The rise starts from a trough. The Heat Score stood at 70 on 15 September and at 40 on 29 September; this week brought it to 67, three points short of its own high, not to new ground.
• The readings behind the move are mixed: a trending score that peaked at 20 on 7 October and fell back to 16, and downloads that only turned up in the last two evenings. The 35B sibling rose in the same days; our readings do not say why, and the rule does not ask.
• One measured GGUF, Q4_K_M from Ornith AI's own organisation; other quantizations change every number above. The context figures assume an f16 cache and count the 24 constant-state layers inside the runtime overhead. No benchmarks and no speed figures: whether it is good at your task is yours to test.
• The rental prices are one marketplace median and one list price at one hour; they move during the day.