What hardware do you need to run Olmo 3.1 32B Instruct?
Olmo 3.1 32B Instruct at Q4_K_M is 19 GB of weights, with a context ceiling of 64k tokens. The largest instruction-following model of Ai2's Olmo 3 family. Ai2 publishes the pretraining data, the post-training datasets, the training code and the intermediate checkpoints alongside the weights, all under Apache 2.0. It is dense, English-only, and built from the Olmo-3-32B base through supervised fine-tuning, preference optimisation and reinforcement learning on verifiable rewards. Its chat template takes tool definitions and tells the model its own knowledge stops at December 2024. On Ai2's own evaluation table it beats Qwen3 32B with thinking off at maths and competition problems, 93.4 against 84.3 on MATH and 67.8 against 27.9 on AIME 2024, and trails it on general knowledge, 80.9 against 85.8 on MMLU and 48.6 against 54.4 on GPQA. Those are the publisher's figures, and nobody has placed this model on the index the rest of this site is ranked by. On the cheapest machine that runs it, a Framework Desktop, 32GB, it generates 8.6 tok/s at 32k of context, worked out from memory bandwidth.
How good is Olmo 3.1 32B Instruct, really?
It has not been placed on the intelligence index yet. See the whole table.
Head to head with the models that need much the same memory at 32k of context: vs Ornith 1.5 35B-A3B, vs Olmo 3.1 32B Think.
What Olmo 3.1 32B Instruct costs either way
Nobody rents Olmo 3.1 32B Instruct by the token. The closest hosted match, Gemma 4 31B it, costs $0.1 per million input tokens and $0.34 per million output (OpenRouter, cheapest active endpoint, checked 2026-11-14). Buying a machine only beats that if you use it hard enough, for long enough, that the hardware price divides down below the rental bill. What a million tokens costs each way puts the two prices side by side.
Machines that run Olmo 3.1 32B Instruct
Every machine here runs it, but not to the same length. The Strix Halo Framework Desktop, 32GB stops at 32k; the AMD Radeon AI PRO R9700, 32GB takes it to 64k, this model's own ceiling. The weights are the same size on every machine; what differs is the memory left for the key-value cache, which grows with every token you keep.
| Machine | Price | Speed at 32k | Longest context | Pay-back | |
|---|---|---|---|---|---|
| Strix Halo Framework Desktop, 32GB | $1,269 | 8.6 tok/s estimated | 32k | Pays back in 101 years | Run the numbers |
| AMD Radeon AI PRO R9700, 32GB | $1,299card only | 25 tok/s estimated | 64k | Pays back in 90 years | Run the numbers |
| NVIDIA GeForce RTX 5090, 32GB | $1,999card only | 55 tok/s estimated | 64k | Pays back in 130 years | Run the numbers |
| Mac mini M5 Pro, 48GB | $2,299 | 10 tok/s estimated | 64k | Pays back in 168 years | Run the numbers |
| Mac Studio M5 Max, 36GB | $2,499 | 15 tok/s estimated | 64k | Pays back in 157 years | Run the numbers |
| MacBook Pro M5 Pro (16-inch), 48GB | $3,599 | 10 tok/s estimated | 64k | Pays back in 264 years | Run the numbers |
| DGX Spark GB10 Grace Blackwell, 128GB | $4,699 | 9.1 tok/s estimated | 64k | Pays back in 387 years | Run the numbers |
The table is the cheapest machine in each family, not the only one that runs it: 41 of the 56 machines on this site hold it at 32k. Within a family only memory decides, so that is every Mac Studio and Strix Halo box at any size, every NVIDIA card with 24GB or more, every Mac mini and MacBook Pro with 48GB or more, the Radeon AI PRO R9700, 32GB and the DGX Spark, 128GB. One table sets every machine against the others: price, the memory its GPU can use, what it holds and how long it takes to pay for itself.
One machine per family, cheapest first. Speeds are measured where a public benchmark exists and estimated from memory bandwidth otherwise; the calculator says which for any configuration. The longest context is the longest setting the calculator offers that the machine still holds this model at, cache included, and no machine is shown taking it past its own 64k limit. Each one opens the calculator on that machine at that length.
For scale, the calculator starts from 80 tok/s for a hosted API and times a local machine against it. Nothing above reaches it, and the quickest is 55 tok/s.
The cheapest machine that runs it is sold with 32 GB. There is a page on what 32 GB runs, machine by machine: it names the 6 machines here sold at that size, and the 40 current models the roomiest of them holds at 32k, this one among them.
It fits in less memory than that: the smallest size here that runs it at 32k is 24 GB, on the GeForce RTX 3090, 24GB at $1,499, card only, which hands a model 23 GB and is no longer sold.
Run Olmo 3.1 32B Instruct at 64k on the AMD Radeon AI PRO R9700, 32GB
How much memory does Olmo 3.1 32B Instruct need at 64k?
Every figure above is priced at 32k of context. At 64k, this model's own ceiling, Olmo 3.1 32B Instruct needs 24.6 GB rather than 22.4 GB. The weights are the same either way; the key-value cache goes from 3 GB to 5.1 GB.
28 of the 37 machines here still sold at a published price hold Olmo 3.1 32B Instruct at 32k and 27 hold it at 64k. The cheapest still sold goes from the Framework Desktop, 32GB at $1,269 to the Radeon AI PRO R9700, 32GB at $1,299, card only. The Radeon AI PRO R9700, 32GB is priced as the card alone, so add the PC around it before comparing it with a complete computer.
On the Radeon AI PRO R9700, 32GB it runs 25 tok/s estimated at 32k and 23 tok/s estimated at 64k. Every token generated reads the whole cache back, so a longer window is slower as well as bigger.
Two more machines, at a shorter window
The table above lists the machines that run it at 32k, the context the calculator starts on, where Olmo 3.1 32B Instruct needs 22 GB. Two more machines hold it at a shorter window, because the weights are the same size either way and the key-value cache is not. None of them is cheaper than the table above, which starts at $1,269.
| Machine | Price | Longest window it fits | Needs there |
|---|---|---|---|
| Mac mini M6, 32GB | $1,299 | 8k | 21 GB |
| MacBook Pro M5 (14-inch), 32GB | $2,399 | 8k | 21 GB |
One machine per family, cheapest first, from the same list as the table above. Each window is the longest setting the calculator offers that the machine still holds this model at, and it opens the calculator on that machine at that length. What it needs there is the weights plus the key-value cache at that window, measured against the memory each machine's GPU can address, which that machine's own page gives.
The specifics
- Parameters
- 32.2B
- Quantisation
- Q4_K_M
- Weights on disk
- 19 GB
- KV cache
- 3.0 GB at 32k context — 16 of the 64 layers grow with the context; the other 48 stop at a 4,096-token window, three local layers to every global one. Both kinds cache the same geometry, eight KV heads by 128 wide, so a token costs 4,096 bytes in whichever layer holds it. How weights and cache add up.
- Maximum context
- 64k tokens (which the config reaches by scaling the positions of a model pretrained at 8,192 tokens)
- Licence
- Apache 2.0
- Sources
- allenai/Olmo-3.1-32B-Instruct, allenai/Olmo-3.1-32B-Instruct (config), bartowski/allenai_Olmo-3.1-32B-Instruct-GGUF, openrouter.ai