Inclusionai Releases Ling-3.0-flash Open-Weight Mixture-of-Experts Model
Inclusionai released Ling-3.0-flash, an open-weight 124B Mixture-of-Experts model that activates 5.1B parameters per token. Delivering 406.5 tokens per second at $0.03 per million tokens, it offers high throughput for code generation and self-hosting.
Impact: Medium
Why it matters
Download weights freely to run locally on Ollama or vLLM, or use hosted APIs at 91% lower cost than competing open-weight models.
TL;DR
- 01Ling-3.0-flash activates only 5.1B of its 124B parameters per token, enabling 406.5 tokens/s generation.
- 02Delivers 85.5% on GPQA and 55.4% on Terminal-Bench for $0.03 per 1M blended tokens or free self-hosting.
- 03Compatible with local inference engines including Ollama, vLLM, and llama.cpp.
Key facts
- Total Parameters
- 124B (MoE)
- Activated Parameters
- 5.1B per token
- Blended Token Price
- $0.03 per 1M tokens (3:1 ratio)
- Hosted API Rates
- $0.02 input / $0.06 output per 1M tokens
- Inference Speed
- 406.5 tokens/s (1.84s median TTFT)
- Key Benchmarks
- GPQA 85.5%, Terminal-Bench 55.4%, SciCode 41.1%
Sparse MoE Architecture and Inference Speeds
Ling-3.0-flash employs a 124B-parameter sparse Mixture-of-Experts (MoE) structure, activating approximately 5.1B parameters per token. Independent testing by Artificial Analysis measured output generation at 406.5 tokens per second with a median time-to-first-token delay of 1.84 seconds.
Benchmark Results and Code Generation
On the Artificial Analysis intelligence scale, Ling-3.0-flash scored 37.8, placing higher than 68% of tracked models. Specific task scores include 85.5% on GPQA, 55.4% on Terminal-Bench, 41.1% on SciCode, 67% on Long Context Reasoning, 27.2% on τ-Bench Banking, and 23.7% on Humanity's Last Exam. The model is specifically targeted at code generation and CLI execution.
Deployment Modes and API Economics
The model weights are open and free to download for self-hosting with ollama, vllm, or llama.cpp on local hardware without token charges. Managed cloud endpoints list at $0.02 per million input tokens and $0.06 per million output tokens ($0.03 per 1M tokens blended at a 3:1 ratio), ranking it cheaper than 91% of comparable open-weight hosted models.
Try it in 2 minutes
ollama run ling-3.0-flashbash
✓ When to use
- High-throughput code generation, terminal agent loops, and automated test writing.
- Self-hosted inference where server memory limits full dense models.
What to do today
- Pull Ling-3.0-flash weights into Ollama or vLLM to benchmark local generation speed.
- Route high-volume coding or agentic CLI tasks to the API endpoint to reduce inference spend.
Sources