AI Gear Watch

What can it run

Popular open-weight models, and the cheapest box we track that has room for each one. Memory capacity decides whether a model loads at all, so this is the first question worth answering.

The requirement is the weights at that quantization plus a 20 percent allowance for the KV cache, activations, and the runtime. It is arithmetic on the format, not a measurement. Long contexts and large batches push the real number up, sometimes a lot. Every parameter count links to the model card it came from.

Search

Family

Quantization

Hardware budget

9 of 9 models

Qwen3-8B

8.2B parameters

Model card: "Number of Parameters: 8.2B". Context 32,768 natively, 131,072 with YaRN.

Quantization ?NeedsCheapest that fitsOptions
Q4 (4-bit)4.9 GBAMD Radeon RX 7900 XTX ($930)15

Parameter count from Qwen3-8B model card

gpt-oss-20b

21B parameters · 3.6B active per token

OpenAI: "21B parameters with 3.6B active parameters"; card says it can run within 16GB of memory.

Quantization ?NeedsCheapest that fitsOptions
Q4 (4-bit)12.6 GBAMD Radeon RX 7900 XTX ($930)15

Parameter count from gpt-oss-120b model card

Mistral Small 3.2 24B

24B parameters

Model card lists 24B params. Mistral does not state a context window on the card.

Quantization ?NeedsCheapest that fitsOptions
Q4 (4-bit)14.4 GBAMD Radeon RX 7900 XTX ($930)15

Parameter count from Mistral Small 3.2 24B Instruct model card

Gemma 3 27B

27B parameters

Model card lists 27B params, with a 128K input context window.

Quantization ?NeedsCheapest that fitsOptions
Q4 (4-bit)16.2 GBAMD Radeon RX 7900 XTX ($930)15

Parameter count from Gemma 3 27B IT model card

Llama 3.3 70B Instruct

70B parameters

Card table lists 70B; the computed safetensors count reads 71B. We use the nominal 70B.

Quantization ?NeedsCheapest that fitsOptions
Q4 (4-bit)42.0 GBNVIDIA Jetson AGX Orin 64GB Developer Kit ($3,249)11

Parameter count from Llama-3.3-70B-Instruct model card

gpt-oss-120b

117B parameters · 5.1B active per token

OpenAI: "117B parameters with 5.1B active parameters"; designed to fit a single 80GB GPU using MXFP4.

Quantization ?NeedsCheapest that fitsOptions
Q4 (4-bit)70.2 GBGMKtec EVO-X2 (Ryzen AI Max+ 395, 128GB) ($3,500)8

Parameter count from gpt-oss-120b model card

Qwen3-235B-A22B

235B parameters · 22B active per token

Mixture of experts: "235B in total and 22B activated". All weights must still be resident.

Quantization ?NeedsCheapest that fitsOptions
Q4 (4-bit)141.0 GBApple Mac Studio (M3 Ultra, 512GB) (price unconfirmed)1

Parameter count from Qwen3-235B-A22B model card

DeepSeek-V3

671B parameters · 37B active per token

"671B total parameters" with "37B activated for each token". The HF upload totals 685B including the MTP module.

Quantization ?NeedsCheapest that fitsOptions
Q4 (4-bit)402.6 GBApple Mac Studio (M3 Ultra, 512GB) (price unconfirmed)1

Parameter count from DeepSeek-V3 model card

Kimi K3

2800B parameters · 104B active per token

Mixture of experts: card table reads Total Parameters 2.8T, Activated Parameters 104B, 896 experts with 16 selected per token, 1,048,576 context. All weights must be resident, so nothing we track fits it at any quantization.

Quantization ?NeedsCheapest that fitsOptions
Q4 (4-bit)1680.0 GBNothing we track fits it0

Parameter count from Kimi-K3 model card

Mixture-of-experts models activate only a fraction of their weights per token, but every weight still has to be resident in memory, so sizing uses the total. See what can you actually run at home for why capacity and bandwidth answer different questions, and our methodology for how these figures are sourced.