InclusionAI's Ling-3.1-flash Boosts Agent Scores on a 560B Open Model
InclusionAI's upgraded 560B hybrid MoE jumps to 44 on Artificial Analysis Index with sharper reasoning and agentic workflow gains.
Takeaways
- InclusionAI released Ling-3.1-flash-1019, an upgraded 560B/25B-active hybrid MoE with 1M context.
- Scores 44 on Artificial Analysis Intelligence Index, up from 41 for Ling-3.1-flash.
- AA-Briefcase jumps from ~1400 to 1524 Elo; AutomationBench-AA from 62% to 68%.
- Scientific benchmarks improve: CritPt 18% to 26%, SciCode 54% to 57%.
- Architecture: 63 Kimi Delta Attention layers plus 9 MLA layers, 512 experts, 8 active.
- Base model also released under MIT for further post-training; SGLang and vLLM supported day-0.
InclusionAI has released an updated 1019 checkpoint of its flagship open-weight model, plus a base checkpoint for custom fine-tuning. The release improves agentic, automation, and scientific benchmark results through post-training changes while retaining the architecture of the original release.
Ling-3.1-flash is a 560-billion-parameter mixture-of-experts model with 25 billion active parameters per token and a one-million-token context window. Activating a subset of parameters reduces computation compared with a dense 560B model, although deployments must still store the full set of weights.
Agent tasks drive the score increase
AA-Briefcase and AutomationBench-AA measure agentic and automation performance, while CritPt and SciCode cover scientific reasoning and programming. The broader Artificial Analysis Intelligence Index combines results across several evaluations.
| Benchmark | Previous | 1019 | Change |
|---|---|---|---|
| AA-Briefcase | ≈1,400 Elo | 1,524 Elo | ≈+124 Elo |
| AutomationBench-AA | 62% | 68% | +6 percentage points |
| CritPt | 18% | 26% | +8 percentage points |
| SciCode | 54% | 57% | +3 percentage points |
| Intelligence Index v4.3.2 | 41 | 44 | +3 points |
The largest absolute accuracy gain appears on CritPt, followed by AutomationBench-AA. That pattern is consistent with post-training aimed at multi-step reasoning and tool-mediated workflows, although benchmark scores alone cannot identify the training changes that produced the gains.