Kimi-Linear-48B-A3B-Instruct
Where it runs 8K context · ≥ 5 t/s floor
| Setup | Best plan | Memory | Generate |
|---|---|---|---|
| 8 GB GPU + 32 GB RAM | Q4_K_Mllama.cpp CUDAGPU + CPU experts · 21 expert layers on CPU | KV unknown | ~8.7–14 t/s Estimated extrapolated |
| 12 GB GPU + 32 GB RAM | Q4_K_Mllama.cpp CUDAGPU + CPU experts · 17 expert layers on CPU | KV unknown | ~6.7–10 t/s Estimated extrapolated |
| 16 GB GPU + 64 GB RAM | Q4_K_Mllama.cpp CUDAGPU + CPU experts · 14 expert layers on CPU | KV unknown | ~14–22 t/s Estimated extrapolated |
| 24 GB GPU + 64 GB RAM | Q4_K_Mllama.cpp CUDAGPU + CPU experts · 6 expert layers on CPU | KV unknown | ~28–47 t/s Estimated extrapolated |
| 32 GB GPU + 128 GB RAM | Q4_K_Mllama.cpp CUDAFull GPU | KV unknown | ~159–459 t/s Estimated extrapolated |
| 2× 24 GB GPUs + 128 GB RAM | Q5_K_Mllama.cpp CUDALayer split across GPUs | KV unknown | ~71–205 t/s Estimated extrapolated |
| 96 GB workstation GPU + 256 GB RAM | Q8_0llama.cpp CUDAFull GPU | KV unknown | ~91–264 t/s Estimated extrapolated |
| Mac, 24 GB | Does not fit | ||
| Mac, 48 GB | Q5_K_Mllama.cpp MetalUnified memory | KV unknown | ~21–60 t/s Estimated extrapolated |
| Mac, 128 GB | MLX 8-bitMLX-LMUnified memory | KV unknown | ~46–61 t/s Estimated |
| Mac, 512 GB | MLX 8-bitMLX-LMUnified memory | KV unknown | ~42–119 t/s Estimated |
| DGX Spark 128 GB | Q8_0llama.cpp CUDAUnified memory | KV unknown | ~22–60 t/s Estimated |
| Ryzen AI Max+ 395, 128 GB | Q8_0llama.cpp VulkanUnified memory | KV unknown | ~13–38 t/s Estimated extrapolated |
| CPU only, 64 GB DDR5-5600 dual channel | Q4_K_Mllama.cpp CPUCPU only | KV unknown | ~9.5–11 t/s Estimated extrapolated |
Same plan engine as the planner, with its default reserves and RAM. How plans are chosen.
Files exact Hugging Face bytes
| Artifact | Size | Files | Revision | Source |
|---|---|---|---|---|
| Q8_0 bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF | 48.66 GiB | 2 | 228dbe476e | HF files |
| Q5_K_M bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF | 32.69 GiB | 1 | 228dbe476e | HF files |
| Q4_K_M bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF | 28.00 GiB | 1 | 228dbe476e | HF files |
| MLX 8-bit 8-bit mlx-community/Kimi-Linear-48B-A3B-Instruct-8bit | 48.61 GiB | 10 | 553a7a5673 | HF files |
| MLX 4-bit 4-bit mlx-community/Kimi-Linear-48B-A3B-Instruct-4bit | 25.74 GiB | 6 | 919175d754 | HF files |
| Native safetensors mixed/custom native; ; stored tensor dtypes=BF16/F32 · BF16/ moonshotai/Kimi-Linear-48B-A3B-Instruct | 91.50 GiB | 20 | e1df551a44 | HF files |
Not found in the representative quant search (not proven unavailable): GGUF FP16, EXL3, AWQ, GPTQ. Complete selected file sets; projectors, MTP/draft and lookup files are listed separately and budgeted only by plans that use them. File bytes are not allocated memory.
Quality
| AA evaluated mode | Index | LiveCodeBench | Terminal-Bench 2.1 | GPQA | HLE | Source |
|---|---|---|---|---|---|---|
| Kimi Linear 48B A3B Instruct AA estimate | 7.3 | 28.0% | — | 41.2% | 2.5% | source |
Arena text rating
Unknown
Publisher-reported benchmarks
None extracted from the model card.
AA and Arena describe the hosted model in the stated mode, not each quantized file; scores are never inherited by base, pruned, abliterated or quantized variants and never averaged together.
How to run it
No model-specific recipe in the research. The planner binds the backend’s documented command template (llama.cpp, Ollama, LM Studio, MLX, vLLM) to your plan: open a plan’s details.
Recipes are source examples (settings are the publisher’s or author’s, not measured optima) and were not executed by LocalFit. A downloadable file does not prove your backend build supports the architecture.
Reported speeds 0
No reported measurement yet. Submit one.