Ling 3.1 Flash

BuildFastWithAi's review describes Ling 3.1 Flash as InclusionAI's new hybrid reasoning model released September 29, 2026, a 560B-total-parameter MoE with about 25B active per token, positioned around coding, multi-step analysis, and tool-using agents. The review explicitly distinguishes the announced 1M-token design target from the Vercel AI Gateway-served configuration, which currently exposes 262,144 tokens of context and up to 32,768 output tokens, a difference the author flags as material for production deployment planning. The review consolidates InclusionAI's launch benchmark numbers, including 52.5% on AutomationBench, 68.7% on SkillsBench, 87.9% on CyberGym, 57.9% on Finance Agent v2, 85.5% on DRACO, 40.4% on Terminal-Bench 4, 55.9% on SWE Atlas Codebase QnA, and 65.3% on HealthBench Professional, while noting promotional Vercel access is mentioned through October 13, 2026. It frames the model as a substantial scale jump from Ling 3.0 Flash and warns that the Vercel-served context is materially smaller than InclusionAI's announced design ceiling.

ArtificialWatch's wire reports that Ant Group's Ling team announced Ling-3.1-flash on September 29, 2026, describing a Mixture-of-Experts model with roughly 560B total parameters and about 25B active per token, targeting a 1M-token context window, with weights "planned to open-source soon." The article also notes that inclusionai/ling-3.1-flash is already listed on the Vercel AI Gateway and on nano-gpt, and quotes Ant's own benchmark figures including 1,673 Elo on GDPVal-AA v2.1, 75.16 on FrontierSWE, and 65.35 on HealthBench Professional. The same writeup flags a discrepancy: models.dev lists Ling 3.1 Flash with a 262,144-token context and 32,768-token max output on Vercel and nano-gpt, well below the 1M Ant describes, with the Vercel route currently free and nano-gpt charging $0.075/$0.22 per million tokens in/out. No Hugging Face repository under inclusionAI exists yet, and the model is not yet on OpenRouter, so the "open-source release" framing is premature. Ant has not published its own token pricing.

ZICQ's article covers the Ling-3.1-flash release by Ling Labs, citing approximately 560 billion total parameters and 25 billion active parameters per token in a MoE architecture designed for long-context workloads. It lists creator-published benchmark scores of 1,673 Elo on GDPVal-AA v2.1, 75.16 on FrontierSWE, and 65.35 on HealthBench Professional, framing the model as broadly applicable across programming, healthcare, and document-processing domains. A 1-million-token context window is highlighted as the headline capability. The piece states Ling Labs plans to open-source the model after a two-week free usage period, presenting this as a forthcoming contribution to the open-source AI community, though the article does not address Vercel AI Gateway specifically or any current served deployment. Multi-domain strength, long-sequence handling, and the promised open-source release are positioned as the primary draws for developers evaluating cross-industry applications.

The Ling 3.1 Flash tracker on BenchLM records the model as carrying a 262,000-token context window and no published first-party API price, with the field-median price shown as $1 input per million tokens. Eight published benchmark rows are available, led by an Agentic category at 81.0 (verified, 22% weight) across five benchmarks, while Coding (40% weight) and Knowledge (12%) rows are verified but still score-pending. Reasoning, Multimodal, Multilingual, Instruction-Following, and Math categories are listed as Not measured with zero benchmarks. BenchLM flags Ling 3.1 Flash as unranked, citing insufficient eligible comparative evidence and no first-party hosted token rate, and describes it as a "Radar7 confirmed" entry from the Sep 25 – Oct 1, 2026 window. The page is updated as of October 1, 2026 and tracks 637 models with 508 benchmarks overall, providing only a supplementary third-party benchmark view rather than any Vercel-specific configuration detail. Decision-snapshot framing is explicitly advisory, not an absolute quality ranking.

InclusionAI, part of Ant Group, released Ling-3.1-flash as a mixture-of-experts language model with 560 billion total parameters and roughly 25 billion activated per token, positioning it for agent tasks, search, and office applications. The model is designed for a one-million-token context window, but the initial two-week free trial is capped at 256,000 tokens. InclusionAI stated plans to raise the context cap to its full one-million-token design and to release the model as open source once the trial period ends. The announcement, summarized by AI Weekly from a TechNode report, did not include benchmark numbers or name a spokesperson, leaving the open-source timeline and expanded context window as unconfirmed future commitments rather than completed events.