Ling-3.0-flash
Ant Group's 124B hybrid-linear MoE runs coding agents on a 5.1B activation budget
What it is
inclusionAI's 124B-total / 5.1B-active MoE mixing 35 KDA linear-attention layers with 7 Gated MLA layers (5:1), 512 routed experts with 8 active, and 256K context trained on an 8K→32K→256K schedule. Thinking mode is on by default and switchable per request; official FP8 builds ship alongside, with an inclusionAI vLLM fork and SGLang cookbook for serving.
Why it's interesting
56.6 on SWE-Bench Pro and 72.4 on SWE-Bench Multilingual from a 5.1B activation budget is serious agentic-coding efficiency — the same KDA-heavy recipe Kimi K3 bet on, at a fraction of the footprint, under an MIT tag. It extends inclusionAI's steady run of genuinely permissive releases.
Use cases
- Cheap self-hosted backends for SWE-agent-style coding pipelines
- Long-context multilingual code work within 256K tokens
- Hybrid linear-attention research on an open MIT base
Who it's for
Coding-agent builders; teams serving high-throughput inference on a budget
Setup
Advanced. Recommended 4× H20-3e (141GB-class) or a 4-GPU Blackwell node; H100/H800 works at --tp 8. FP8 official; community GGUF and NVFP4 quants bring it lower
Limitations & cautions
Serving still wants a multi-GPU node despite the small activation budget, launch benchmarks are self-reported with independent evals early, and the missing LICENSE file should be a five-minute fix we shouldn't have to keep asking for.
Editorial takeaway
Kimi K3 made KDA famous at 2.8T parameters. Ling-3.0-flash makes it affordable at 124B — and unlike K3, the license fits in three letters.