inclusionAI releases Ling-3.1-flash-1019: 560B MoE, 25B active, 1M-token context
inclusionAI has released Ling-3.1-flash-1019, a post-training upgrade of Ling-3.1-flash with 560B total parameters, 25B active per token, and a 1M-token context window. The company reports a score of 44 on the Artificial Analysis Intelligence Index v4.3.2, up from 41 for its predecessor.
Ling-3.1-flash-1019 at a glance
inclusionAI has released Ling-3.1-flash-1019, an upgraded version of Ling-3.1-flash with 560B total parameters, 25B active parameters per token, and a 1M-token context window. The company says further post-training optimization improved reasoning and real-world agentic performance. Pricing is not yet disclosed.
Benchmark results (company-reported)
According to inclusionAI, Ling-3.1-flash-1019 improves on its predecessor across the following evaluations:
| Benchmark | Ling-3.1-flash | Ling-3.1-flash-1019 |
|---|---|---|
| AA-Briefcase (Elo) | approx. 1400 | 1524 |
| AutomationBench-AA | 62% | 68% |
| CritPt | 18% | 26% |
| SciCode | 54% | 57% |
| Artificial Analysis Intelligence Index v4.3.2 | 41 | 44 |
inclusionAI says the gains reflect better complex reasoning, multi-step task execution, and agentic workflows.
Architecture
The model is a hybrid-linear mixture-of-experts design. Specifications listed on the model card:
- Parameters: 560B total, 25B activated per token
- Layers: 63 KDA layers and 9 Gated MLA layers (7:1 ratio), with 4 dense layers
- Experts: 512 routed, 1 shared, 8 activated per token
- Attention heads: 64
- Hidden size: 4096
- Expert intermediate size: 1280
- Dense intermediate size: 10240
- Vocabulary size: 157184
The Hugging Face page lists the checkpoint at 572B parameters in Safetensors format, with BF16, F8_E4M3 and F32 tensor types.
Serving and deployment
inclusionAI recommends SGLang. A day-0 image (lmsysorg/sglang:dev-ling-3.1-flash) includes Ling-3.1-flash support, MTP (NEXTN) speculative decoding, and Hopper decode optimizations. The documented configuration for 8 × H200 is TP8 plus EP8. On 4 × GB300, the card says to use --tp 4 --ep 4. Two optional environment flags, SGLANG_ENABLE_FP8_BLOCK_GEMV=1 and SGLANG_ENABLE_FUSED_KDA_OUTPUT_QUANT=1, are described as giving roughly 10% faster batch-1 decode on Hopper. The card points to an SGLang cookbook with verified recipes for H200, GB300, H800, H100 and B200, covering HiCache (Mooncake L3), reasoning on and off, tool calling, and measured accuracy and speed numbers.
vLLM is also supported. The documented server command uses tensor parallel size 8, expert parallelism, prefix caching, the ling3 tool-call and reasoning parsers, and MTP speculative decoding with 3 speculative tokens. The quickstart commands on the card reference the inclusionAI/Ling-3.1-flash model path.
Recommended sampling parameters are temperature 1.0, top_p 0.95, and top_k 20, with enable_thinking set to true. The card links to Hugging Face, ModelScope and OpenRouter. As of the Hugging Face page, no Inference Provider deploys the model. The page also links a public training-content summary (PDF), which it says does not replace the technical documentation, usage terms, or license.
What this means
The update is a post-training revision, not a new architecture: the parameter counts, expert layout and 1M-token context window are described as the model's specifications, and the reported gains come from further post-training optimization. The Artificial Analysis Intelligence Index v4.3.2 score of 44 (up from 41) and the agentic results on AA-Briefcase and AutomationBench-AA are all inclusionAI's own reported figures. With 25B of 560B parameters active per token and a 7:1 mix of KDA and Gated MLA layers, the design targets long-context, agentic workloads at a lower per-token compute than the total parameter count suggests. Serving still requires multi-GPU nodes: the documented configurations use 8 × H200 or 4 × GB300.
Related Articles
Baidu releases Unlimited-OCR-2, using SR-SWA to parse documents of unbounded length in a finite context window
Baidu released Unlimited-OCR-2 on 2026-10-27. It introduces Sparse Reference Sliding Window Attention (SR-SWA), which Baidu says keeps per-step decoding latency and memory nearly constant regardless of document length. Baidu claims state-of-the-art results on MPDocBench and MarathonDocBench.
iFLYTEK SparkAudio 1.0 speech model supports 99 languages and 202 dialects, pairs 0.65B encoder with 30B-A3B MoE
iFLYTEK detailed SparkAudio 1.0 at its 2026 Global 1024 Developer Festival. The speech foundation model combines a 0.65-billion-parameter audio encoder with a 30B-A3B mixture-of-experts language model, trained on 13 million hours of audio. iFLYTEK also launched Robot Super Brain Platform 3.0 for embodied robots.
Nanbeige4.3-33B released: 33B open model with 512K context targets agentic coding and cowork tasks
Nanbeige has released Nanbeige4.3-33B, a 33B-parameter BF16 model with a 524,288-token (512K) context length. It is the flagship teacher model of the Nanbeige4.3 series and will anchor distillation into smaller models such as Nanbeige4.3-3B. Nanbeige claims state-of-the-art agentic performance among open models in the ~30B class.
KRAFTON releases Raon-1 mini preview, a 28.8B-parameter MoE model with 4.3B active and 262,144-token context
KRAFTON AI has released Raon-1 mini preview, a Mixture-of-Experts language model trained from scratch with 28.8B total and 4.3B active parameters. It is Apache 2.0 licensed, supports 262,144 tokens of context, and ships with a speculative decoding draft model that KRAFTON says reaches up to 2.18× higher output throughput on a single NVIDIA B300.
Comments
Loading...