Ling-3.1-flash

inclusionAI's newest Ling, and a reminder that "flash" has stopped meaning small. It is a hybrid-reasoning MoE with 560B total parameters and 25B active. Two months earlier Ling 3.0 carried the same tier name at 124B total and 5.1B active, so this is roughly 4.5× the total parameters and 5× the active. Artificial Analysis scores it 41 on Intelligence Index v4.3.2 against Ling 3.0's 20, a doubling of measured intelligence inside a single minor version number. The API lists $0.30 in / $0.90 out per million tokens. Ant positions it for agent workloads, search, office software and vertical applications.

It launched on September 30, 2026 as a two-week free API trial, served on OpenRouter by Novita, with nothing published by inclusionAI itself. The weights followed on October 13 on HuggingFace and ModelScope under the MIT license. The FP8 checkpoint uses e4m3 weights in 128×128 blocks and includes one MTP (NEXTN) layer for speculative decoding. The model card describes a hybrid-linear MoE: 63 KDA (Kimi Delta Attention) linear-attention layers and 9 gated MLA layers in a 7:1 ratio across 72 layers, the first 4 dense, then 512 routed experts plus one shared expert with 8 active per token. Hidden size is 4096, with 64 heads and a 157,184-token vocabulary. The card says it "builds upon Ling-3.0-flash with improvements in architecture", but at 4.5× the size it cannot be a simple continuation of that checkpoint, and inclusionAI does not say how the two relate. The EU AI Act training-content summary lists no model dependencies, more than 10T text tokens of public, licensed, synthetic and human-written data, and a data cutoff of about July 2026. It gives no exact token count, hardware or recipe. There is still no technical report.

Context length is where the claims need reading carefully. The card advertises a "1M-token context window", but config.json sets max_position_embeddings to 262,144, and the trial API served 256K. The SGLang cookbook explains the gap: 262,144 is the native context, and 1,048,576 is reached by a 4× YaRN RoPE override. SGLang reports verifying this on H200 and GB300 with exact retrieval of a fact planted at token 700K of a 900K prompt, and recommends enabling it only beyond 256K. The structured field below records the native figure. The card's benchmark results (Terminal-Bench 4.0, SWE-bench Pro, FrontierSWE, GDPVal-AA v2.1, HealthBench Professional, MCP-Atlas and others) are published only as images, so the AA score is still the main number that can be checked independently. The card also introduces LingCode, a long-horizon coding harness built on the open-source Pi agent, which inclusionAI is "considering open-sourcing".

A second checkpoint, Ling-3.1-flash-1019, appeared on Artificial Analysis on October 22. AA scored it 44 on v4.3.2, three points above the open release, and 29 on v5.0 (October 27; AA had not yet rescored the open release), at the same price and with a 1,048,576-token context. It was first served only through inclusionAI's API (release slug ling-3-1-flash-rc5), and its MIT weights followed on HuggingFace on October 27, two days after a Ling-3.1-flash-base checkpoint. It has its own entry as Ling-3.1-flash-1019; the top-level score here stays on the original checkpoint this file names.