Its KDA layers are unverified on Metal: the one Mac run (19.5 tok/s on a 512 GB M3 Ultra) used a private MLX conversion in an oMLX release candidate, and no public MLX build exists. The LLMCheck index estimates Ling-3.1-flash at ~18 tok/s on an M3 Ultra at Q4_K_M, a figure from its published bandwidth model, not a measured run. It needs about 360 GB of unified memory.
Want to run Ling-3.1-flash faster — or step up to bigger models? Find your Mac →This is an LLMCheck index estimate from its published two-term bandwidth model, not a measured run; the index labels it so everywhere it appears. Own this Mac? Submit a real run →
| Metric | Value |
|---|---|
| Tokens per second | estimated~18 tok/s |
| Provenance | estimate · how estimates are made |
| Quantization | Q4_K_M |
| Engine | llama.cpp |
| Figure dated | 2026-10 |
| Minimum RAM | 360 GB |
| Parameters | 560B MoE (25B active) |
| Chip bandwidth | 819 GB/s |
Each row as the index holds it, with its source. “Mac RAM” is the memory of the Mac a run used, not the model’s requirement (360 GB). The headline figure follows the reference rule: the lower-median measured plain-decode run at 5 bits or fewer (a run whose quant was not stated counts only when no stated-quant run qualifies), else the estimate; a run marked “not used to anchor estimates” is listed here but never becomes the headline.
| Speed | Quant | Engine | Mac RAM | Context | Source | Note |
|---|---|---|---|---|---|---|
| community19.5 tok/s | MLX mixed 4/8-bit | MLX | 512 GB | 4k | omlx.ai | 80-core GPU. oMLX v0.7.1rc1, MTP/DFlash/expert offload off; Ling-3.1-flash-MLX-mixed-4_8bit, a conversion with no public repo (the page labels it 8-bit; bpw not published), running its KDA layers on a release candidate, so it does not anchor; 312.3 GB peak. The first Mac measurement of this model.; not used to anchor estimates |
| estimated18 tok/s | Q4_K_M | llama.cpp | 512 GB | — | LLMCheck estimator | — |
Ollama — the index has not verified an Ollama tag for Ling-3.1-flash, so it doesn’t print one (a guessed tag fails with “pull model manifest: file does not exist”). Search the Ollama library →
MLX — pick a build from Hugging Face’s MLX listings, then run it with its repo id:
Downloads: Hugging Face repo → · model card →
Prefer a GUI? LM Studio downloads GGUF and MLX builds in one click. Step by step: install Ollama on a Mac.
| Chip | Speed | Quant | Engine | Bandwidth |
|---|---|---|---|---|
| M5 Ultra | estimated26 tok/s | Q4_K_M | llama.cpp | 1228 GB/s |
Ant Group's MIT-licensed 560B hybrid-linear MoE (13 Oct 2026): 25B active (KDA + MLA layers, small KV cache), 262K context; its SWE-bench Pro 68.7 is on Ant's own LingCode scaffold. Upstream llama.cpp runs it (bailingmoe3), but Q4_K_M is 358 GB, so it needs a 512 GB Mac Studio; a 2-bit 219 GB build might squeeze onto 256 GB with the GPU limit raised, untested. The only Mac measurement, 19.5 tok/s on a 512 GB M3 Ultra, used a private MLX mixed 4/8-bit conversion in an oMLX release candidate; no public MLX build exists, and its KDA layers are otherwise unverified on Metal.
MIT licence · 560B MoE · 262K context · LLMCheck capability 47/50 · see its leaderboard rank →
Ling-3.1-flash needs about 360 GB of unified memory. These current Apple Silicon Macs have the headroom to run it comfortably:
As an Amazon Associate, LLMCheck earns from qualifying purchases. These affiliate links cost you nothing extra and help keep the LLMCheck index free.
See how Ling-3.1-flash stacks up against other models on your specific Mac hardware.
Open Compare Tool Full Leaderboard