8 B parameters · dense
Llama 3.1 8B Instruct
56 sourced configurations run it. Every configuration with a supported backend path has the memory to hold it.
32 catalogued models
Which machine runs which model.
A model gets a page when the catalog actually discriminates between machines for it. Decode ranges are estimates at a 1,024-token position, batch 1, single stream.
Generated pages
Browse 5 models
Each page ranks every sourced configuration this model runs on.
1.2 B parameters · dense
Llama 3.2 1B Instruct
56 sourced configurations run it. Every configuration with a supported backend path has the memory to hold it.
3.2 B parameters · dense
Llama 3.2 3B Instruct
56 sourced configurations run it. Every configuration with a supported backend path has the memory to hold it.
70.6 B parameters · dense
Llama 3.3 70B Instruct
40 sourced configurations run it. 16 do not have the memory to hold it.
30.5 B parameters · mixture of experts
Qwen3 30B-A3B
56 sourced configurations run it. Every configuration with a supported backend path has the memory to hold it.
Catalogued, not estimated
27 models with no page
These stay selectable in the calculator. No fit or speed result is invented for them.
| Model | Parameters | Attention | Why no result is published |
|---|---|---|---|
| DeepSeek-R1 deepseek · released 2025-01 · model ID deepseek-r1 | 671 B total 37 B active per token. | MLA | Backend target vllm-cuda has not declared a sourced MLA execution path. The candidate formulas differ by roughly an order of magnitude, so no fit verdict and no token rate are published for this combination (§2.3.2). |
| DeepSeek-V3 deepseek · released 2024-12 · model ID deepseek-v3 | 671 B total 37 B active per token. | MLA | Backend target vllm-cuda has not declared a sourced MLA execution path. The candidate formulas differ by roughly an order of magnitude, so no fit verdict and no token rate are published for this combination (§2.3.2). |
| DeepSeek-V4-Flash 284B-A13B deepseek · released 2026-04 · model ID deepseek-v4-flash | 284 B total 13 B active per token. | DeepSeek-V4 sparse/latent attention | The model uses DeepSeek-V4 sparse/latent attention. Its KV storage, decode traffic, and prefill cost do not have a reviewed formula, so no fit verdict or speed is published. |
| DeepSeek-V4-Pro 1.6T-A49B deepseek · released 2026-04 · model ID deepseek-v4-pro | 1,600 B total 49 B active per token. | DeepSeek-V4 sparse/latent attention | The model uses DeepSeek-V4 sparse/latent attention. Its KV storage, decode traffic, and prefill cost do not have a reviewed formula, so no fit verdict or speed is published. |
| Gemma 4 12B Unified Instruct gemma · released 2026-06 · model ID gemma-4-12b | 12 B total 12 B active per token. | hybrid sliding-window/global attention; global layers cache unified K/V (no v_proj) and use p-RoPE | The model uses hybrid sliding-window/global attention; global layers cache unified K/V (no v_proj) and use p-RoPE. Its KV storage, decode traffic, and prefill cost do not have a reviewed formula, so no fit verdict or speed is published. |
| Gemma 4 26B-A4B Instruct gemma · released 2026-03 · model ID gemma-4-26b-a4b | 25.8 B total 3.8 B active per token. | hybrid sliding-window/global attention; global layers cache unified K/V (no v_proj) and use p-RoPE | The model uses hybrid sliding-window/global attention; global layers cache unified K/V (no v_proj) and use p-RoPE. Its KV storage, decode traffic, and prefill cost do not have a reviewed formula, so no fit verdict or speed is published. |
| Gemma 4 31B Instruct gemma · released 2026-03 · model ID gemma-4-31b | 31.3 B total 31.3 B active per token. | hybrid sliding-window/global attention; global layers cache unified K/V (no v_proj) and use p-RoPE | The model uses hybrid sliding-window/global attention; global layers cache unified K/V (no v_proj) and use p-RoPE. Its KV storage, decode traffic, and prefill cost do not have a reviewed formula, so no fit verdict or speed is published. |
| Gemma 4 E2B Instruct gemma · released 2026-03 · model ID gemma-4-e2b | 5.1 B total 5.1 B active per token. | hybrid sliding-window/global attention with cross-layer KV sharing (20 of 35 layers) and per-layer embeddings | The model uses hybrid sliding-window/global attention with cross-layer KV sharing (20 of 35 layers) and per-layer embeddings. Its KV storage, decode traffic, and prefill cost do not have a reviewed formula, so no fit verdict or speed is published. |
| Gemma 4 E4B Instruct gemma · released 2026-03 · model ID gemma-4-e4b | 8 B total 8 B active per token. | hybrid sliding-window/global attention with cross-layer KV sharing (18 of 42 layers) and per-layer embeddings | The model uses hybrid sliding-window/global attention with cross-layer KV sharing (18 of 42 layers) and per-layer embeddings. Its KV storage, decode traffic, and prefill cost do not have a reviewed formula, so no fit verdict or speed is published. |
| GLM-5.1 744B-A40B glm · released 2026-04 · model ID glm-5.1 | 744 B total 40 B active per token. | MLA with DeepSeek Sparse Attention | The model uses MLA with DeepSeek Sparse Attention. Its KV storage, decode traffic, and prefill cost do not have a reviewed formula, so no fit verdict or speed is published. |
| GLM-5.2 744B-A40B glm · released 2026-06 · model ID glm-5.2 | 744 B total 40 B active per token. | MLA with IndexShare DeepSeek Sparse Attention | The model uses MLA with IndexShare DeepSeek Sparse Attention. Its KV storage, decode traffic, and prefill cost do not have a reviewed formula, so no fit verdict or speed is published. |
| GLM-5.3-Flash 320B-A18B glm · released 2026-08 · model ID glm-5.3-flash | 320 B total 18 B active per token. | hybrid sparse/linear attention | The model uses hybrid sparse/linear attention. Its KV storage, decode traffic, and prefill cost do not have a reviewed formula, so no fit verdict or speed is published. |
| GLM-5.3 744B-A40B glm · released 2026-08 · model ID glm-5.3 | 744 B total 40 B active per token. | MLA with IndexShare DeepSeek Sparse Attention | The model uses MLA with IndexShare DeepSeek Sparse Attention. Its KV storage, decode traffic, and prefill cost do not have a reviewed formula, so no fit verdict or speed is published. |
| GLM-5 744B-A40B glm · released 2026-02 · model ID glm-5 | 744 B total 40 B active per token. | MLA with DeepSeek Sparse Attention | The model uses MLA with DeepSeek Sparse Attention. Its KV storage, decode traffic, and prefill cost do not have a reviewed formula, so no fit verdict or speed is published. |
| Qwen3.5 0.8B qwen · released 2026-02 · model ID qwen3.5-0.8b | 0.9 B total 0.9 B active per token. | hybrid linear/full attention | The model uses hybrid linear/full attention. Its KV storage, decode traffic, and prefill cost do not have a reviewed formula, so no fit verdict or speed is published. |
| Qwen3.5 122B-A10B qwen · released 2026-02 · model ID qwen3.5-122b-a10b | 125.1 B total 10 B active per token. | hybrid linear/full attention | The model uses hybrid linear/full attention. Its KV storage, decode traffic, and prefill cost do not have a reviewed formula, so no fit verdict or speed is published. |
| Qwen3.5 27B qwen · released 2026-02 · model ID qwen3.5-27b | 27.8 B total 27.8 B active per token. | hybrid linear/full attention | The model uses hybrid linear/full attention. Its KV storage, decode traffic, and prefill cost do not have a reviewed formula, so no fit verdict or speed is published. |
| Qwen3.5 2B qwen · released 2026-02 · model ID qwen3.5-2b | 2.3 B total 2.3 B active per token. | hybrid linear/full attention | The model uses hybrid linear/full attention. Its KV storage, decode traffic, and prefill cost do not have a reviewed formula, so no fit verdict or speed is published. |
| Qwen3.5 35B-A3B qwen · released 2026-02 · model ID qwen3.5-35b-a3b | 36 B total 3 B active per token. | hybrid linear/full attention | The model uses hybrid linear/full attention. Its KV storage, decode traffic, and prefill cost do not have a reviewed formula, so no fit verdict or speed is published. |
| Qwen3.5 397B-A17B qwen · released 2026-02 · model ID qwen3.5-397b-a17b | 403.4 B total 17 B active per token. | hybrid linear/full attention | The model uses hybrid linear/full attention. Its KV storage, decode traffic, and prefill cost do not have a reviewed formula, so no fit verdict or speed is published. |
| Qwen3.5 4B qwen · released 2026-02 · model ID qwen3.5-4b | 4.7 B total 4.7 B active per token. | hybrid linear/full attention | The model uses hybrid linear/full attention. Its KV storage, decode traffic, and prefill cost do not have a reviewed formula, so no fit verdict or speed is published. |
| Qwen3.5 9B qwen · released 2026-02 · model ID qwen3.5-9b | 9.7 B total 9.7 B active per token. | hybrid linear/full attention | The model uses hybrid linear/full attention. Its KV storage, decode traffic, and prefill cost do not have a reviewed formula, so no fit verdict or speed is published. |
| Qwen3.6 27B qwen · released 2026-04 · model ID qwen3.6-27b | 27.8 B total 27.8 B active per token. | hybrid linear/full attention | The model uses hybrid linear/full attention. Its KV storage, decode traffic, and prefill cost do not have a reviewed formula, so no fit verdict or speed is published. |
| Qwen3.6 35B-A3B qwen · released 2026-04 · model ID qwen3.6-35b-a3b | 36 B total 3 B active per token. | hybrid linear/full attention | The model uses hybrid linear/full attention. Its KV storage, decode traffic, and prefill cost do not have a reviewed formula, so no fit verdict or speed is published. |
| Qwen3.8 2.4T-A95B qwen · released 2026-08 · model ID qwen3.8-2.4t-a95b | 2,446.2 B total 95 B active per token. | hybrid linear/full attention | The model uses hybrid linear/full attention. Its KV storage, decode traffic, and prefill cost do not have a reviewed formula, so no fit verdict or speed is published. |
| Qwen3.8 27B qwen · released 2026-08 · model ID qwen3.8-27b | 27.8 B total 27.8 B active per token. | hybrid linear/full attention | The model uses hybrid linear/full attention. Its KV storage, decode traffic, and prefill cost do not have a reviewed formula, so no fit verdict or speed is published. |
| Qwen3.8 Flash-Next 180B-A6B qwen · released 2026-08 · model ID qwen3.8-flash-next | 180 B total 6 B active per token. | hybrid linear/full attention with n-gram embeddings | The model uses hybrid linear/full attention with n-gram embeddings. Its KV storage, decode traffic, and prefill cost do not have a reviewed formula, so no fit verdict or speed is published. |
Still useful
What a model without a page still gives you
Weight accounting depends on sourced parameter counts alone.
Weight bytes stay publishable for every catalogued model, because they need no attention formula. Fit, decode, and prefill do need one, and the site withholds all three rather than borrowing a formula from a different architecture. Open the calculator to select any of these models against a configuration and read the reason back with the inputs it was resolved from.