Ant Group Opens Ling 3.1 Flash Weights: 560B Hybrid-MoE Under MIT License With 1M Context
InclusionAI, the model arm of Ant Group, published the weights of Ling 3.1 Flash on Hugging Face on October 14, 2026, under an MIT license. The model ran as a free-trial API for two weeks from September 30 before the weights were made available.
Architecture
Ling 3.1 Flash uses a hybrid-linear mixture-of-experts design. The 72 transformer layers split into 63 KDA (kernel delta attention) layers and 9 gated MLA (multi-latent attention) layers in a roughly 7:1 ratio. KDA is a linear-attention variant that runs more efficiently at inference time than standard quadratic attention. The 9 gated MLA layers are interspersed to add standard attention capability where it matters most.
The expert count is 512 routed experts plus 1 shared expert, with 8 active per token. With 25 billion parameters active per forward pass out of 560 billion total, the architecture is designed for throughput at scale. The vocabulary covers 157,184 tokens.
Context is 1 million tokens. The free-trial API was capped at 256K; the full 1M is available with the open weights.
What Makes This Distinct
The MIT license is the headline detail. Most recent large-scale open releases have used Apache 2.0 (Reflection Beam, K2 Horizon) or custom licenses with commercial restrictions (Qwen3.8-Max’s open checkpoint). MIT imposes no such conditions: commercial use, redistribution, modification, and sublicensing are all permitted with no revenue limits or attribution requirements beyond the license text.
At 560B total / 25B active, Ling 3.1 Flash is the largest model the Ling series has shipped. The predecessor Ling 3.0 Flash ran 124B total with 5.1B active. The scale increase is roughly 4.5x in total parameters and 5x in active parameters, while the active count relative to total stays similar (around 4.5%).
InclusionAI also released a training-content disclosure summary (TDS PDF) in its AI-Transparency repository, covering what data was used in pre-training. This is notably rare for commercial Chinese labs.
Deployment
Recommended serving: SGLang with MTP (multi-token prediction) speculative decoding on 8x H200 at TP8+EP8 (tensor-parallel 8 + expert-parallel 8), or 4x GB300. vLLM is also supported. The 8x H200 configuration gives developers running modern H200 cluster allocations a direct on-ramp.
API throughput is approximately 212 tokens per second at the hosted endpoint (Klu leaderboard, October). For a 560B model, that is competitive with similarly-scaled open-source options.
InclusionAI also introduced LingCode alongside the open-source release — a long-horizon coding harness built on the open-source Pi agent framework. LingCode is described as supporting “considering open-sourcing in the future.”
Benchmark Claims
Ling 3.1 Flash’s model card lists strong results on GDPVal-AA V2.1, HealthBench Professional, FrontierSWE, Terminal-Bench 4.0, SWE-bench Pro, MCP-Atlas, and SkillsBench. The scores are shown as images rather than text in the model card, which prevents direct citation. From an earlier AICoder summary of pre-release data, vendor-reported figures include SWE-bench Pro at 65.39, Terminal-Bench 4.0 at 40.40, CyberGym at 87.90, and BrowseComp at 91.67. None of these have been independently reproduced as of the open-source release date.
For context, GLM-5.3 (a 320B MoE from Z.AI) reached 95.4% on SWE-bench Verified in independent testing. Ling 3.1 Flash does not have a SWE-bench Verified score. The SWE-bench Pro figure of 65.39 is on a different and harder dataset — Claude Fable 5 scores approximately 80.0 on SWE-bench Pro for comparison.
Positioning
Ling 3.1 Flash joins a cluster of large open-source Chinese models released in the second half of 2026: Kimi K3 (2.8T, Apache 2.0), Reflection Beam (501B, Apache 2.0 pending red-team), and K2 Horizon 375B A23B (375B, Apache 2.0). The MIT license sets Ling 3.1 Flash apart on permissiveness.
At $0.45 per million input tokens on the hosted API, pricing is below Kimi K3 ($3/$15) and below Reflection Beam on self-hosted equivalents. The API endpoint on LLM Gateway was marked deactivated as of October 20 — the free-trial period ended. Weights are available for self-hosting, and additional inference providers are expected to follow the open-source release.