Ant Ling
32 Intel
Ant Ling AGI is the account for Ant Group's AGI initiative, publishing the Ling and Ring families of open-source large language models built on a mixture-of-experts architecture. The Ling series are general-purpose models and the Ring series are reasoning-focused, with releases such as Ring-1T reaching trillion-parameter scale and performing strongly on IMO-style reasoning benchmarks. The group also publishes the multimodal Ming model family for image, text, audio, and video.
Is this you? Sign in with X to claim this profile.
Intel
Ant Group's Ling team open-sourced Ling-3.0-flash-Fin, a finance-enhanced model for real-world workflows, and FinFIRST, an expert-built benchmark for financial search agents. Only the announcement was reviewed, so sizes, licenses and results are unverified.
Ant Group ships FP8, FP4 and INT4 quantized checkpoints of Ling-3.0-flash-Fin on Hugging Face and ModelScope, letting teams pick a precision tier that fits their memory and latency budget for financial-workflow deployments.
Ant Ling post-trained Ling-3.0-tiny with GSPO on one DGX Spark using the AReno framework. Over 400 steps on tic-tac-toe, mean reward rose from about -0.5 to 0.4 and responses shortened to about 850 tokens. The authors frame it as task adaptation, not a general capability gain.
Ant Ling open-sourced the Ming-Image-0.1-Design family: two 6B models (base and Layer) and two agent skills for UI design and image-to-editable-PPT conversion. The base model is claimed to rank first among open-weight models on Artificial Analysis's UI/UX Design leaderboard.
Ant's Ling team launched Ring-2.6-1T, a trillion-parameter thinking model with adjustable effort levels. It reports PinchBench 87.60, Tau2-Bench Telecom 95.32 and AIME 26 95.83. It was free on OpenRouter for a week, with open weights promised. Figures are vendor-reported.
Ant Ling's successor to IcePop swaps the fixed-ratio token mask for an adaptive binary-KL trust region matched to each token's inherent noise, which the team credits for stable trillion-scale long-horizon agentic RL and a pure-RL SWE-bench Verified score above 76.
Ant Group introduces Ling-3.1-flash, a mixture-of-experts model with about 560B total and 25B active parameters and up to 1M tokens of context. It self-reports 75.16 on FrontierSWE and 1,673 Elo on GDPVal-AA v2.1, and says open-sourcing is planned.
Ant Ling open-sourced six base checkpoints for Ling-3.0-tiny (7.9B total, 1.3B active) and Ling-3.0-flash (124B total, 5.1B active), covering pretrained, mid-trained and WSM-merged stages with no post-training. The WSM paper on decay-free learning-rate schedules accompanies them. Performance claims
Ant Ling released Ling-3.0-flash, a 124B hybrid-reasoning MoE with 5.1B active parameters. It stacks KDA and MLA attention 5:1, activates 1/64 of experts, and supports 256K context scaling to 1M. The lab claims parity with its 1T flagship and shows Blender, Office and multi-agent demos.
Ant Ling released Ling-3.0-flash-Sante, a mixture-of-experts model fine-tuned for medicine and health. It targets clinical reasoning, deep research and evidence retrieval, with claimed open-source-leading results on medical benchmarks. Scores and weights are not in the announcement.
Ant Ling launched Ling-3.0-tiny, a hybrid-reasoning MoE with 7.9B total and 1.3B active parameters, on OpenRouter and Vercel AI Gateway. Demos cover local Obsidian retrieval, offline translation, and tool use for UI control. Open weights were promised, and benchmarks were not included.
Ant Ling open-sourced a DSpark speculative decoding draft model for Ling-3.0-flash, reporting 1,120 tok/s at batch 1 on four Blackwell GPUs. The thread explains the draft architecture, acceptance-aware training, and a FlashInfer host-sync fix that removed idle GPU time. Figures are self-reported, an
Ant Ling's Ling-3.0-flash-VL moved from free to paid on OpenRouter on September 23 at 9:00 AM PT. Limited-time discounted rates per million tokens are $0.021 input, $0.063 output and $0.0042 for cache hits.
Ling-3.0-tiny open weights are out in BF16, FP8 and INT4. Reported results include Artificial Analysis Intelligence 25, Agentic 16, GPQA Diamond 73.40 and IMO-AnswerBench 71.03. Architecture is a 3:1 KDA-MLA hybrid with 128 experts, 8 routed and 1 shared. Local throughput figures are given for DGX S
Ling-3.0-flash-VL is now free on OpenRouter for two weeks, and Ant Group has open-sourced FP4 and INT4 quantized versions for self-hosted multimodal agent deployments.
Ant Ling and the SGLang team built a Fused MoE V2 Pallas kernel for TPU v7x that overlaps token routing and HBM weight prefetch with compute. Reported results are 53% lower MoE prefill latency and up to 1.77x decode throughput against a similar H200 cluster. The linked deep-dive was not read.
Ant Ling released INT4 and MXFP4 builds of its 124B MoE Ling-3.0-flash that run end to end on one DGX Spark through an adapted SGLang path. Self-reported figures: about 80 tok/s decode, 2,500 to 3,500 tok/s prefill, 3 to 4 concurrent users. W4A16 is the stable FP4 default.
Ant Ling released open weights for Ling-3.0-flash in official BF16 and FP8-quantized versions on Hugging Face and ModelScope, with SGLang and vLLM deployment support acknowledged. The thread invites self-hosted evaluation and use in agentic workflows but gives no benchmark data.
Ant released Ling-3.0-flash-Fin, a 124B-total, 5.1B-active finance variant built with financial institutions. Demos show 20 to 55 tool calls over 10-Ks and 5,000-formula workbooks. Its AA Intelligence Index score rose from 38 to 41, and open weights were promised for the next week.
Artificial Analysis benchmarked Ling-3.0-tiny on phone-ready quantized builds, scoring 59 at 16K context with 5.7 seconds on an iPhone 17 Pro and top rank at 64K with 66. A rare case of intelligence and real-device latency measured on the same build.
Ant Ling describes a Weight Cache Daemon for SGLang: one process per GPU and TP rank keeps sharded, FP8-quantized weights resident and shares them with restarting engines through CUDA IPC. Ling-2.6-1T startup drops from 8.8 to about 0.53 minutes, with config checks and disk fallback.
Ant Ling argues FP4 training quality depends on grid geometry. E2M1's asymmetric rounding bins shrink magnitudes toward zero and the error accumulates across layers. Its UFP4 recipe uses a uniform grid and reportedly tracks BF16 more closely from 1.5B dense up to 124B MoE. Thread summary only.
Ant Ling proposes PowLU, a SwiGLU variant whose growth decays from quadratic toward linear. It matches SwiGLU scaling-law exponents from 26M to 368M parameters and held a smooth FP8 loss curve where SwiGLU and SwiGLU-Clip spiked near step 77k. Based on the thread, not the full paper.
Ant Ling's Ling and Ring 2.6 report covers a 7:1 Lightning Attention to MLA hybrid for 256K context, KPop adaptive binary-KL for stable MoE RL (SWE-bench Verified 70.8% to 76.28%), and token-efficiency methods. Base weights for the flash and 1T models are released. Summary based on the thread only.