Ant's Ling-3.0-flash Activates Only 5.1B Params: Architecture and Cost Analysis
jkris050 · reddit · 2026-08-05
Ant Group's lab inclusionAI open-sourced the Ling-3.0-flash model under the MIT license on Aug 4. It has 124B total parameters but only activates 5.1B (about 1/64) per token, utilizing 512 routed experts and 1 shared expert.
Architectural Highlights:
- Natively adopted hybrid linear attention from the start of pretraining rather than retrofitting.
- Features 35 KDA layers alternating 5:1 against 7 gated MLA layers.
Performance & Cost Debate:
- The lab claims it matches or beats their own 1T param flagship, Ring-2.6-1T.
- The poster notes that if a lab can cut activated params by 12x without losing performance, it redefines the "cheap tier"—it's not a degraded version, but the same capability firing fewer experts.
- Deployment Pain Points: Currently lacks GGUF and llama.cpp support, requiring 4 GPUs for custom vLLM/SGLang forks. It's cheap to run, but holding 124B weights means VRAM ownership costs remain high.
Related event: Ant Group open-sources Ling-3.0-flash: 124B params, only 5.1B active(11 posts)→
More from Infra
- IT spend near $3T a year and CIO surveys show budgets shifting to dedicated AI — BenBajarin · 2026-10-26
- Anthropic S-1 reveals 48 cents of compute cost per dollar of Q2 revenue — nmasc_ · 2026-10-26
- Persistent AI joins UR network to route private AI inference with encrypted, IP-hidden traffic — 0xSammy · 2026-10-26
- Anthropic's disclosed compute commitments: $114.5B Google, $161.2B Broadcom top the list — BenBajarin · 2026-10-26
- SpaceX strikes $40B deal with Apollo Global to lease AI chips and data center hardware — Polymarket · 2026-10-26
- SpaceX AI revenue up 830%, Claude designs 318K antivenom proteins in 64 hours — PeterDiamandis · 2026-10-26