inclusionAI ships Ling-3.0-flash, a 124B MoE with 5.1B active

A model release worth reading twice. inclusionAI has posted Ling-3.0-flash on Hugging Face under an MIT license: 124B total parameters, but only about 5.1B activated per token, roughly 8% of the model firing on any given step. The team's own claim is that it matches or outperforms its predecessor Ring-2.6-1T while using roughly 12.4% of that model's total parameters.

The architecture is where the interesting choices sit. It is described as a native hybrid-linear MoE, built with a 5:1 alternating stack of Kimi Delta Attention and Multi-head Latent Attention layers, and trained through a curriculum that stretches from 8K to 32K to 256K context. On the model card's own benchmarks, SWE-Bench Pro comes in at 56.6%, SWE-Bench Multilingual at 72.4%, MathArena AIME 2026 at 93.2 and HLE at 22.7. The team also claims Time to First Token drops by 60 to 80% in long-input scenarios thanks to hierarchical caching.

Why this matters if you are not training foundation models yourself: sustained agentic tool use over long contexts has been the reason most teams stayed on hosted frontier APIs. A model that activates 5.1B per token, holds a 256K window, and ships under MIT changes the deployment math for teams running coding agents on their own hardware. The card explicitly lists compatibility with Claude Code, Kilo Code, Qwen Code, Hermes Agent and OpenClaw, and describes training against more than 10,000 interactive environments to teach agent behaviors directly.

The honest caveat is the obvious one. These are self-reported benchmarks, custom tool-call and reasoning parsers are required to deploy the model, and independent evaluations have not caught up. What the page does not give you is the training data mix, the cost of training, or how the model behaves outside the tasks it was tuned to publish. According to a Kilo announcement the model is being offered free on that platform for a limited time, which is the fastest way to A/B it against whatever coding agent you are paying for today.

If the numbers survive contact with real workloads, the leverage this gives buyers on long-running inference contracts is the part worth watching.