Mulai dari mana?
Mulai dari perangkat yang Anda miliki atau model yang diinginkan. Bandingkan kapasitas, kinerja, dan bukti sebelum deployment.
Select hardware to see model fit results.
Guides & methodology
Browse the complete catalogs and estimation method when you need them.
Semua perangkat21
Setiap kelas punya anggaran, jalur runtime, dan batas realistis sendiri.
Mac Studio M5 Ultra · 96 GB
The entry M5 Ultra configuration combines a 96 GB unified pool with 1.2 TB/s memory bandwidth. Apple announced it in August 2026; availability begins September 22.
Buka panduanMac mini M6 · 24 GB
A compact current-generation Mac with 24 GB unified memory. Keep macOS, runtime, and KV-cache headroom separate from model weights.
Buka panduanMacBook Pro M5 Max · 64 GB
A portable Apple-native tier for larger quantized models. Sustained inference still depends on thermals, runtime support, and context length.
Buka panduanFramework Desktop · Ryzen AI Max+ 395 · 128 GB
A 128 GB Strix Halo desktop. AMD allows a large graphics-memory allocation, but usable capacity and speed depend on BIOS, driver, and runtime support.
Buka panduanRadeon AI PRO R9700 · 32 GB
A 32 GB AMD workstation card for larger single-GPU quantized models. Check the exact operating-system and ROCm or Vulkan runtime path first.
Buka panduanNVIDIA DGX Spark · 128 GB
A compact local AI system with 128 GB coherent memory. It can hold large low-bit models, but decode speed and software support remain model-specific.
Buka panduanMac Studio M3 Ultra · 96 GB
A quiet single-box option for 70B-class Q4 models and Apple-native MLX workflows, with memory shared by macOS and applications.
Buka panduanMac Studio M3 Ultra · 256 GB
Enough capacity for many 100B-plus low-bit models, provided the architecture is supported and optimized by the selected runtime.
Buka panduanGeForce RTX 5090 · 32 GB
A fast single-GPU tier for 27B-class low-bit models, with limited remaining space for very long context or multiple concurrent users.
Buka panduanGeForce RTX 4090 · 24 GB
A 27B Q4 checkpoint can be a tight fit. Leave room for runtime overhead and KV cache instead of filling every gigabyte with weights.
Buka panduan8 GB CPU laptop
Small text models with short context. Expect patient, private inference rather than speed.
Buka panduan16 GB CPU desktop
Comfortable with 3B–8B Q4 models; 12B is possible only with reduced context and patience.
Buka panduanApple silicon · 16 GB
A polished 4B–8B local experience when the OS and apps have enough headroom.
Buka panduanNVIDIA · 8 GB
The mainstream 4B–8B tier. Some 12B INT4 builds fit tightly with modest context.
Buka panduanNVIDIA · 12 GB
Strong 8B–14B Q4 territory for a single user.
Buka panduanNVIDIA · 16 GB
14B models are comfortable; 20B-class low-bit MoE models are a tight upper edge.
Buka panduanApple silicon · 32 GB
24B–32B Q4 models become practical for one local user.
Buka panduanNVIDIA · 24 GB
The consumer sweet spot for 27B and 30B-A3B INT4/Q4 models.
Buka panduanApple silicon · 64 GB
32B models with long context or 70B Q4 with conservative headroom.
Buka panduanWorkstation · 48 GB
70B Q4 serving begins here, with limited room for cache and concurrency.
Buka panduanDatacenter · 80 GB
Large MoE models and high-throughput servers, still bounded by quantization and cache.
Buka panduanSemua model24
Qwen3.8 Flash Next
English- Estimasi memori
- ~108 GB
- Konteks
- 262,144 native; 1M extended
Ling-3.0-flash
English- Estimasi memori
- ~74.4 GB
- Konteks
- Check selected runtime and checkpoint
Nex-N2.5-mini
English- Estimasi memori
- ~21 GB
- Konteks
- 262,144 (config max_position_embeddings)
Qwen3.6 27B
- Estimasi memori
- ~18.5 GB
- Konteks
- 262K native
Qwen3.8 27B
- Estimasi memori
- ~19.5 GB
- Konteks
- 262K native
Qwen3.8 Flash Next REAP-288
- Estimasi memori
- ~40.57 GB
- Konteks
- 7.6K recommended on measured 48 GB Mac
Qwen3.6 35B-A3B
- Estimasi memori
- ~25 GB
- Konteks
- 262K native
DeepSeek V4 Flash
- Estimasi memori
- ~176 GB
- Konteks
- 1M native
Gemma 3 1B
- Estimasi memori
- ~1.4 GB
- Konteks
- 32K
Llama 3.2 3B
- Estimasi memori
- ~2.8 GB
- Konteks
- 128K
Qwen3 4B
- Estimasi memori
- ~3.6 GB
- Konteks
- 32K+
Gemma 3 4B
- Estimasi memori
- ~4.2 GB
- Konteks
- 128K
Qwen3 8B
- Estimasi memori
- ~6.8 GB
- Konteks
- 32K+
Gemma 3 12B
- Estimasi memori
- ~9.4 GB
- Konteks
- 128K
Qwen3 14B
- Estimasi memori
- ~11.2 GB
- Konteks
- 32K+
gpt-oss-20b
- Estimasi memori
- ~16 GB
- Konteks
- 128K
Devstral Small 2 24B
- Estimasi memori
- ~18.5 GB
- Konteks
- 256K
Gemma 3 27B
- Estimasi memori
- ~20.5 GB
- Konteks
- 128K
Qwen3 30B-A3B
- Estimasi memori
- ~22.5 GB
- Konteks
- 32K+
Qwen3 32B
- Estimasi memori
- ~24.5 GB
- Konteks
- 32K+
Llama 3.1 70B
- Estimasi memori
- ~46 GB
- Konteks
- 128K
Llama 4 Scout
- Estimasi memori
- ~78 GB
- Konteks
- 10M advertised
gpt-oss-120b
- Estimasi memori
- ~80 GB
- Konteks
- 128K
Qwen-Image 2.1
English- Estimasi memori
- ~33.1 GB
- Konteks
- Not a token context: one prompt plus up to 10 reference images
Pilih lapisan inferensi15
Ollama
One-command local chat and app integration
Buka panduanllama.cpp
GGUF models, CPU/GPU offload, embedded and unusual hardware
Buka panduanLM Studio
Discovering, downloading and testing models without a terminal
Buka panduanMLX LM
Native Apple silicon inference, experimentation and fine-tuning
Buka panduanoMLX
Apple-silicon serving with continuous batching, memory guards, and supported SSD-backed paths
Buka panduanvLLM
Linux GPU servers, concurrency and OpenAI-compatible production APIs
Buka panduanTensorRT-LLM
Maximum NVIDIA throughput after engine tuning
Buka panduanSGLang
Multi-request GPU serving with prefix caching and speculative decoding
Buka panduanLocalAI
One OpenAI-compatible endpoint for text, vision, speech and image models, GPU optional
Buka panduanLemonade
Ryzen AI, Radeon and Strix Halo PCs, including the XDNA2 NPU, behind one local API
Buka panduanOpenVINO Model Server
Serving OpenVINO IR models on Intel CPUs, Arc GPUs and NPUs with one API
Buka panduanKoboldCpp
One-file GGUF chat on Windows, macOS or Linux with no Python install
Buka panduanllamafile
Distributing a whole GGUF model as one runnable executable on Linux, macOS, Windows and BSD
Buka panduanKTransformers
Very large MoE models with hot experts on the GPU and cold experts in system RAM
Buka panduanUnsloth
Running GGUF or MLX models from a desktop app that also trains and exports them
Buka panduanCara estimasi dibuat
Kami memakai ukuran dan kuantisasi terdokumentasi, menyisakan ruang untuk sistem, dan menghitung cache sebagai tambahan. MoE memakai parameter total.
Kami memisahkan parameter total, MoE aktif, ukuran kuantisasi, dukungan runtime, dan kinerja terukur.
- Memory fit is an estimate, not a speed measurement.
- Size MoE by total parameters; shared and distributed memory need additional checks.
- Vendor claims, community measurements, runtime conditions and unknowns are disclosed separately.
Catalog data updated: 2026-10-03
Sebelum mengunduh 40 GB
Does a 24 GB GPU run a 30B model?
Often at Q4/INT4 with a conservative context. Qwen3 30B-A3B and Gemma 3 27B are representative fits, but cache and runtime overhead still matter.
Are active MoE parameters the memory requirement?
No. Active parameters affect compute per token; total parameters still need to be stored in memory or offloaded.
Which runtime should a beginner choose?
Ollama for a terminal-first setup or LM Studio for a visual desktop. llama.cpp is the portable fallback; vLLM is for higher-throughput GPU serving.
How do you report speed?
Exact device-and-model measurements are labeled measured. Relative manufacturer figures are labeled vendor claims. Every other combination stays unknown instead of receiving an invented tokens-per-second estimate.