InclusionAI Ling Flash 2.0 — fast inference model for general-purpose tasks.
Ling Flash 2.0 is a mixture-of-experts (MoE) language model built by InclusionAI, open-sourced on July 23, 2025. It carries 100B total parameters but activates only 6.1B per token (4.8B non-embedding), guided by the Ling Scaling Laws (arxiv 2507.17702). The model was trained on 20T+ tokens of high-quality data and refined with supervised fine-tuning plus multi-stage reinforcement learning.
Thanks to its 1/32 activation ratio, Ling Flash 2.0 delivers roughly 7× efficiency gains over a comparable dense architecture and achieves 200+ tokens/s on H20 hardware — 3× faster than 36B dense peers. With YaRN extrapolation the context window extends to 128K tokens. The model targets complex reasoning, multi-step code generation, frontend development, and regulated-domain tasks (finance and healthcare).
InclusionAI publishes Ling Flash 2.0 as an MIT-licensed open-weight Hugging Face model with local deployment guidance, 32K native context, and optional 128K YaRN extrapolation.
TheRouter exposes inclusionai/ling-flash-2 through its OpenAI-compatible chat route. Operational pricing, context length, regions, and supported parameters come from the live TheRouter model catalog rather than the Hugging Face checkpoint card.
| Type | Rate |
|---|---|
| Input | $0.1512 / 1M tokens |
| Output | $0.6156 / 1M tokens |
| Release date | 2025-07-23huggingface.co ↗ | verified |
| Architecture | MoE — 100B total / 6.1B activated (4.8B non-embedding), 1/32 activation ratiohuggingface.co ↗ | verified |
| Training data | 20T+ tokens of high-quality datahuggingface.co ↗ | verified |
| Training cutoff | Not publicly disclosed | unknown |
| License | MIThuggingface.co ↗ | verified |
| Inference throughput (H20) | 200+ tokens/s — 3× faster than 36B dense models in typical usehuggingface.co ↗ | verified |
| Context length | 32K native; 128K with YaRN extrapolationhuggingface.co ↗ | verified |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
MMLU-Pro | SOTA vs dense <40B at launch | huggingface.co ↗ | |
GPQA-Diamond | Competitive vs Qwen3-32B-Non-Thinking and GPT-OSS-120B/low | huggingface.co ↗ | |
LiveCodeBench v6 | Strong — outperforms 36B dense models | huggingface.co ↗ | |
AIME 2025 | Strong advanced math performance | huggingface.co ↗ | |
ARC-Prize | Reported competitive on logical reasoning track | huggingface.co ↗ |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "inclusionai/ling-flash-2",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"inclusionai/ling-flash-2","messages":[{"role":"user","content":"Explain the Ling 2.0 MoE architecture in one paragraph."}]}'InclusionAI released Ling Flash 2.0 on HuggingFace under an MIT license. The model activates just 6.1B of its 100B parameters per inference step, enabling 200+ tokens/s on H20 GPUs while matching or outperforming dense models up to 40B on GPQA-Diamond, LiveCodeBench v6, AIME 2025, and FinanceReasoning benchmarks.
The model uses a 1/32 activation-ratio MoE architecture designed under the Ling Scaling Laws research programme. Key architectural choices — expert granularity, shared-expert ratio, aux-loss-free sigmoid routing, MTP layers, QK-Norm, and Partial-RoPE — allow a small active slice to leverage a much larger parameter pool. According to InclusionAI, this yields roughly 7× efficiency gains over a comparable dense model at the same quality level.
32K tokens natively. With YaRN extrapolation (add a rope_scaling field to config.json) the window extends to 128K tokens. As output length grows, the speedup advantage over dense peers increases to 7× or more due to the model's small activation footprint.
| source | URL | retrieved | |
|---|---|---|---|
| Release date | huggingface.co ↗ | 2026-08-13 | verified |
| Architecture | huggingface.co ↗ | 2026-08-13 | verified |
| Training data | huggingface.co ↗ | 2026-08-13 | verified |
| Training cutoff | — | — | unknown |
| License | huggingface.co ↗ | 2026-08-13 | verified |
| Inference throughput (H20) | huggingface.co ↗ | 2026-08-13 | verified |
| Context length | huggingface.co ↗ | 2026-08-13 | verified |
| MMLU-Pro | huggingface.co ↗ | 2026-08-13 | verified |
| GPQA-Diamond | huggingface.co ↗ | 2026-08-13 | verified |
| LiveCodeBench v6 | huggingface.co ↗ | 2026-08-13 | verified |
| AIME 2025 | huggingface.co ↗ | 2026-08-13 | verified |
| ARC-Prize | huggingface.co ↗ | 2026-08-13 | verified |
| InclusionAI open-sources Ling Flash 2.0 — 100B MoE matching 40B dense quality at 3× speed | huggingface.co ↗ | 2026-08-13 | verified |
| How does Ling Flash 2.0 achieve 40B-dense quality with only 6B active parameters? | huggingface.co ↗ | 2026-08-13 | to verify |
| What is the maximum context length supported? | huggingface.co ↗ | 2026-08-13 | to verify |