InclusionAI (Ant Group)
Top signals
- #1ModelsinclusionAI/Qwen3.6-35B-A3B-singprobe7.0
- #2WritingLing: A MoE LLM Provided and Open-sourced by inclusionAI7.0
- #3WritingMing-Omni-TTS: Simple and Efficient Unified Generation of Speech, Music, and Sound with Precise Control7.0
- #4WritingMing-UniAudio: Speech LLM for Joint Understanding, Generation and Editing with Unified Representation7.0
- #5WritingMing-UniVision: Joint Image Understanding and Generation via a Unified Continuous Tokenizer7.0
Agent answer
InclusionAI (Ant Group) has 305 loaded public signals: 0 hiring, 3 forks, 206 releases or model cards, 21 talking, and 75 repos. Latest signal: inclusionAI/AReno v0.10.0. Data-business radar is currently scoped to frontier labs, so this category does not expose radar lanes. The standing analysis was generated with deepseek-v4-pro and 94 evidence refs.
has loaded 305 public signals
has hiring signal count 0
has fork signal count 3
has release signal count 206
Thesis
InclusionAI (Ant Group's AGI arm) is running a two-track play: a frontier open-weights text line (Ling) now reaching a 560B/25B-active hybrid-linear MoE flagship P18W4, layered with fast multi-modal and vertical specialization — healthcare (Sante), finance (Fin), ASR/audio/TTS/OCR/image, and typed decision models P2E7P7P9P11P26P23. At the same time it is publishing the training, serving, and sandbox infrastructure that produces those models — AReno (fine-tune/RL), TorchTurbo (post-training engine), vllm-omni (omni serving), and AKernel (sandbox/cluster orchestration) P1P17P10P6. The consistent MIT/Apache licensing, Hugging Face + ModelScope + API-gateway distribution, and EU AI Act training-content summaries signal a lab optimizing for open-weight adoption and regulatory legibility rather than secrecy P2P18W1W5W6.
Signal desks
Hiring: No cited evidence in this pack — no open roles, team descriptions, locations, or job postings appear in the supplied evidence.
Forks: No fork metadata is present in this pack (repos are flagged "Fork: no" throughout P5P8P17). The closest upstream-adaptation signals are: vllm-omni is rebased onto vLLM 0.30.0 P10; TorchTurbo is built on PyTorch FSDP2 and Hugging Face Transformers P17; and multiple derivatives sit on Qwen backbones (LemonDecider from Qwen3.5-4B P23, PolyOCR-Venus on a Qwen3.5 VLM P26).
Releases: High-frequency cadence across two axes — training/serving infra (AReno v0.9.0→v0.10.0 P1P21, AKernel v0.2.0 rc1→rc3 P6P15P16, Choruz nightly builds P12P13P14) and models (Ling-3.1-flash P18E4, Ling-3.0-flash-Sante P2E10, UBTree speculative-decoding drafts P3P4, Ming ASR/audio/TTS P7P9P11, LemonDecider P23).
Talking: Third-party coverage frames the lab by model taxonomy (Ling/Ring/Ming W5), open-weight 560B MoE launch W4, context-aware ASR W2, speculative-decoding speedups W3, and EU AI Act training-content transparency W6.
Shipping
- Ling-3.1-flash: 560B total / 25B active hybrid-linear MoE, 1M-token context, MIT FP8 weights on Hugging Face P18W4; 3,667 downloads / 148 likes shortly after release E4.
- Ling-3.0-flash-Sante: healthcare-specialized 124B/5.1B-active model on Ling-3.0-flash, 256K context, flagged as "a research model, not a medical device" P2; 146 downloads E10.
- UBTree drafts: speculative-decoding draft models for Ling-3.0-tiny (183M) and Ling-3.0-flash (482M), with SGLang integration still pending P3P4W3.
- Ming modality line: Ming-3.0-asr-tiny context-aware ASR across 15 dialect sets P7P8; Ming-3.0-audio-flash (126B/6.5B-active MoE) P9; Ming-3.0-tts (0.5B dense + 7.9B/1.3B MoE on Ling V3) P11.
- AReno v0.10.0: MLX QLoRA on Apple Silicon, GGUF/Ollama export, native CUDA LoRA/QLoRA for Qwen3.5, Bailing V3 MTP loss + speculative decoding, and JEV decision-head training P1.
- AKernel 0.2.0 RCs: move the Python SDK onto the ADX Environment backend and add HostPath bind mounts for trusted clients P6P15P16.
Research themes
- Hybrid-linear MoE: Kimi Delta Attention (KDA) + gated MLA + sparse MoE, scaling 124B/5.1B-active to 560B/25B-active P2P18W5.
- Speculative decoding: UBTree (unigram proposer + learned bigram selector on a DFlash backbone, tree-native training) P3P4W3; Bailing V3 MTP training loss in AReno P1.
- Typed decision engines: JEV joint decision-head + backbone-LoRA training P1; LemonDecider single-pass choice/noul/score readout with explicit abstention P23.
- Multimodality: Phi-4 Multimodal training/inference flow in AReno P1; audio/ASR/TTS P7P9P11; image models (Ming-Image, LLaDA-Image) E1E8.
- Post-training & serving infra: TorchTurbo supernode-native FSDP2/EP/SP engine P17; vllm-omni unified text/audio/image/video/action serving P10.
- Domain/embodied eval: healthcare (Sante) P2, finance (Ling-3.0-flash-Fin) E7, OCR (PolyOCR 2B/9B) P26, bimanual manipulation benchmark (BiDex-100) P19.
Hiring & scaling
No hiring evidence is cited in this pack — no roles, locations, or team descriptions appear. Scaling signals are indirect product/infra artifacts rather than hiring data: TorchTurbo targets "SuperNode" training clusters P17, AKernel builds sandbox/cluster orchestration with Kubernetes HostPath and runc/gVisor/Firecracker runtimes P6P27, and vllm-omni targets multi-hardware omni serving (NVIDIA Blackwell, Ascend 950) P10. This proves the buildout exists but does not evidence hiring in the pack.
Category implications
- Strategy: moving from a general Ling/Ring/Ming taxonomy W5 into vertical fine-tunes (healthcare Sante P2, finance Fin E7) and typed decision models (LemonDecider P23) implies a GTM shift toward domain/enterprise use and agentic control surfaces rather than raw chat.
- Infrastructure: AKernel's ADX Environment backend + HostPath mounts + sandbox runtimes signal investment in managed, trusted compute/sandbox orchestration for agent workloads P6P15P16P27; TorchTurbo signals supernode-scale post-training compute P17; vllm-omni signals a unified multimodal serving runtime P10.
- Product: the Ming family (ASR, audio, TTS) plus PolyOCR and Ming-Image indicate a buildout of speech/audio and document/visual products alongside LLMs P7P9P11P26E1.
- Research: UBTree drafts and Bailing MTP target serving-latency/cost reduction via speculative decoding P3P4P1; JEV/LemonDecider target efficient structured decision-making P1P23; BiDex-100 extends evaluation into embodied, bimanual manipulation P19.
- GTM/compliance: MIT/Apache licensing, ModelScope/Vercel/OpenRouter distribution, and EU AI Act training-content summaries W6E29 signal a compliance-aware, open-weights distribution strategy P2P18W1W4W5.
Traction highlights
- Model downloads: Ling-3.0-tiny 26,419 E2; Ling-3.0-flash 17,049 E3; Ling-3.0-flash-VL 8,801 E6; UI-Venus-2-9B 4,002 E17; Ling-3.1-flash 3,667 E4; Ming-Image-0.1-Design 1,204 downloads / 435 likes E1.
- Repo stars: Ming-Image 222 E57; Realtime-Venus 217 E59; Ming-3.0-tts 197 E19; PolyOCR-Venus 60 E35; Ming-3.0-audio-flash 58 E30; Ming-3.0-asr 40 E26; TorchTurbo 29 E37.
- Leaderboards: Ming-3.0-tts ranks 1st among open-weights on Controlled Voice and 1st overall on Pronunciation Robustness (Artificial Analysis, Oct 2026) P11.
- Publisher scale: 222 models on Hugging Face, described as updating frequently W1.