RTX 4090
Used-card condition, power supply and case clearance need separate checks.
MODEL / INCLUSIONAI
Ling-3.0-flash
Model card checked 2026-09-26 · Memory fit estimated; speed unknown unless a separately linked measurement matches your configuration.
Also searched as: ling-3.0-flash
Model card states 124B total and 5.1B active; file metadata may count additional tensors. Q4 budget estimated as 124 × 0.5 × 1.2 = 74.4 GB before KV cache. Verify the supported runtime recipe and quantized artifact. Read the card's own launch matrix before planning hardware: its low-latency recipe (built-in MTP, 256K YaRN context) is quoted for 4× 141 GB-class H20-3e nodes or 4-GPU Blackwell nodes, and on 80 GB cards (H100 / H800) it needs --tp 8, so the published path is a multi-GPU deployment rather than one desktop. SGLang is the recipe the card leads with (image lmsysorg/sglang:dev-Ling-3.0-flash, cookbook at docs.sglang.io) and vLLM needs a source build with VLLM_USE_PRECOMPILED=1. Speed is unknown.
Primary sourceCheck this model against your hardwarePick a device or GPU to see memory fit, runtime path and the deployment template.
Total parameters124B
Estimated memory~74.4 GBQ4 estimated budget
ContextCheck selected runtime and checkpointText
LicenseMITMultilingual
01 / BEST FOR
- source-checked local inference experiments
02 / ESTIMATED FIT
03 / DEPLOY WITH
These are runtime examples, not validated installation recipes. Fit estimates assume quantized files; official repositories may hold full-precision weights. MLX, GGUF and TensorRT require matching conversions or engines.
SGLangAdvanced
python3 -m sglang.launch_server --model-path inclusionAI/Ling-3.0-flash --port 30000vLLMAdvanced
vllm serve inclusionAI/Ling-3.0-flash --dtype autoWatch out
Model card states 124B total and 5.1B active; file metadata may count additional tensors. Q4 budget estimated as 124 × 0.5 × 1.2 = 74.4 GB before KV cache. Verify the supported runtime recipe and quantized artifact. Read the card's own launch matrix before planning hardware: its low-latency recipe (built-in MTP, 256K YaRN context) is quoted for 4× 141 GB-class H20-3e nodes or 4-GPU Blackwell nodes, and on 80 GB cards (H100 / H800) it needs --tp 8, so the published path is a multi-GPU deployment rather than one desktop. SGLang is the recipe the card leads with (image lmsysorg/sglang:dev-Ling-3.0-flash, cookbook at docs.sglang.io) and vLLM needs a source build with VLLM_USE_PRECOMPILED=1. Speed is unknown.
Your purchase shortlist
Filter by a runtime path and an installed-capacity floor you already established for your workload. This does not calculate required VRAM. Inputs stay in this browser.
RTX 5090
Check Blackwell kernels for the exact quantization; more VRAM does not certify faster decode.
Radeon AI PRO R9700
Verify OS, driver and ROCm/Vulkan support. Experimental forks are not stock-runtime support.
Framework Desktop Ryzen AI Max+ 395 · 128GB
BIOS GPU allocation and system use reduce available memory; 128 GB RAM is not 128 GB dedicated VRAM.
Mac Studio M5 Ultra · 96GB
m5 ultra 96g and m5 ultra 96gb refer to this capacity. Memory is shared with macOS; verify shipping date and MLX conversion.
DGX Spark
ARM64 packages and GB10 kernels must support the chosen model. Capacity is not a latency result.
Checked model specification
| Issuer / checkpoint | Values | Source / verified |
|---|---|---|
| inclusionAI / inclusionAI/Ling-3.0-flash | 124B total; Check selected runtime and checkpoint; MIT | Publisher model card / 2026-09-26 |
Applicability and failure modes
Inference experiments only. The Q4 memory budget is estimated, not a tested maximum context. Unsupported kernels, missing vision projectors, incorrect tool parsers and longer-context KV allocation can fail even when weights fit. No successful deployment, quality or speed is guaranteed.
Compare hardware purchase paths