Skip to content
local-ai

Ling Flash 2.0

inclusionAI · Text generation · MoE 100B (A6B) · 131k context · Released 16 September 2025

Commercial use permitted Open weights

Ling Flash 2.0 is a sparse mixture-of-experts with a 1/32 activation ratio, which is what the developer credits for its speed: only about 6 billion of its roughly 100 billion parameters compute on each token. It is text-only, with a 32k native context the developer says extends towards 128k with YaRN. For local use the catch is familiar for a model of this shape: the small active count makes it quick, but it is the full 100 billion parameters that have to sit in memory, which puts it on a 128GB unified-memory machine or a multi-GPU server rather than a consumer card.

Strengths

  • Highly sparse, so fast for its total size
  • Sparse: only about 6 billion of ~100 billion parameters compute per token
  • A permissive licence with no revenue threshold or territory restriction
  • A capable general model at the quick end of the large tier

Weaknesses

  • Server-scale footprint: ~58GB at 4-bit and ~206GB at BF16, set by the total not the active count
  • Text-only, with no image or audio input
  • A 32k native context, shorter than several rivals before YaRN extension
  • Benchmark claims await independent verification

Hardware requirements

Quantisation Approx. VRAM Notes
Q4_K_M (approx) ~58GB Roughly 58GB at 4-bit, which needs a 128GB unified-memory machine or multiple GPUs. Driven by the 100B total, not the active count. Approximate.
Q8_0 (approx) ~106GB Roughly 106GB. Multi-GPU or server territory. Approximate
BF16 (as released) ~206GB Full precision, roughly 206GB. Server territory. Approximate

What you'd need to run this

Roughly what a machine to run this would need, at up to three levels of quality. Memory is the deciding factor.

Minimum to run it

Q4_K_M (approx) · ~58GB needed

96GB of unified memory

AMD Ryzen AI Max+ 395 (Strix Halo)

or an 80GB-class data-centre card, which is usually rented by the hour, NVIDIA A100 80GB .

Unified memory is shared with the model, so it is already counted above.

around £1,100–£2,000

See the Quiet large-model box build What else 96GB runs →

For good quality

Q8_0 (approx) · ~106GB needed

128GB of unified memory

Mac Studio M4 Max

or an 80GB-class data-centre card, which is usually rented by the hour, NVIDIA H200 .

Unified memory is shared with the model, so it is already counted above.

around £2,000–£4,000

See the Apple Silicon build for large models build What else 128GB runs →

Best quality

BF16 (as released) · ~206GB needed

A multi-GPU server, roughly 4× 80GB-class GPUs

Server memory in the hundreds of gigabytes or more, plus fast interconnect.

tens of thousands of pounds, or rented by the hour

Usually rented — see the cost calculator

Licence

MIT — read the licence

Ling Flash 2.0: common questions

What hardware do I need to run Ling Flash 2.0?
At its most compressed (Q4_K_M (approx)) it needs roughly 58GB of VRAM. VRAM figures are approximate and depend on context length and settings.
Is Ling Flash 2.0 free for commercial use?
Yes. Ling Flash 2.0 is licensed under MIT, which permits commercial use with no meaningful conditions.
Can I run Ling Flash 2.0 on Apple Silicon?
It can run on Apple Silicon through general runtimes, but it is not specifically optimised for it.
What is Ling Flash 2.0's context window?
Ling Flash 2.0 has a context window of 131,072 tokens, about 131k.

Availability

Recommended for

  • A fast large-tier general model on a 128GB unified-memory machine or server
  • High-volume inference where sparsity keeps a large model quick
  • Permissively licensed deployment kept in-house

Related models

Run it with

Glossary

Catalogue entry last verified 2 October 2026. Specifications change; verify anything you are about to spend money on.