Ling 3.0 Tiny

Ling-3.0-tiny is the smallest model in InclusionAI's Ling-3.0 series, a hybrid reasoning mixture-of-experts with 7.9B total parameters and 1.3B activated per token. Weights are published on Hugging Face and ModelScope under the MIT license in BF16, FP8 and INT4, and the model card reports a 25 on the Artificial Analysis Intelligence Index v4.1.1 and 16 on the Agentic Index, with output speed above 160 tokens per second in Artificial Analysis testing. The architecture stacks three Kimi Delta Attention layers with one Multi-Head Latent Attention layer per four-layer block, plus sparse MoE feed-forward layers with 128 routed experts where 8 routed and 1 shared expert run per token. Thinking is on by default and toggled per request via enable_thinking, and the InclusionAI SGLang recipe serves the model on a single GPU with YaRN scaling to a 262,144-token context, with vLLM and Ollama instructions also provided.

InclusionAI's Ling 3.0 Tiny is a mixture-of-experts model with 1.3B active parameters out of 7.9B total, released on Aug 6, 2026, and offered with a 262K context window. The model is built for responsive agent workloads, instruction following, and multi-turn conversations, and it is available on OpenRouter under the identifier inclusionai/ling-3.0-tiny. For developers, Ling 3.0 Tiny supports tool calling, structured outputs, and reasoning, with switchable thinking and instant modes that let callers trade depth against responsiveness. The listing also flags a provider-observed tool call error rate of about 5.13% on its single host, so teams should validate structured output reliability for agent pipelines before committing to it.

BenchLM's catalog page for Ling 3.0 Tiny records the API model ID as inclusionAI/Ling-3.0-tiny and a 262K context window attributed to the InclusionAI model card. The page notes that InclusionAI publishes the 7.9B-total, 1.3B-active hybrid-linear MoE under MIT on Hugging Face in BF16, FP8 and INT4, validated on DGX Spark, Apple Silicon and Ollama, with thinking mode configurable per request. The profile also tracks Ling 3.0 Tiny as not publicly ranked, with no weighted text-model benchmark result published yet, and price not listed. API capabilities such as tool calling, structured outputs, batch support and prompt caching are marked as not tracked or not documented, so the page presents metadata rather than a verified performance claim.

This catalog entry describes Ling 3.0 Tiny as a mixture-of-experts model from InclusionAI with 1.3B active parameters out of 7.9B total, intended for responsive agents, instruction following, and multi-turn conversations. It enumerates the request parameters the model accepts, including presence penalty, logprobs, temperature, response format, top logprobs, include reasoning, max tokens, tools, tool choice, stop, structured outputs, seed, and frequency penalty, alongside supported reasoning, tools, and structured output features. Aggregate evaluation results on the page rate the model highly reliable at roughly 91% success across evaluations, with reasoning accuracy of 90.9% on logic and multi-step problem solving, 79.7% on coding knowledge, and 57.0% on instruction following, while generation latency is comparatively slow at the 7th percentile. The context window is listed at 262K tokens and per-token pricing at $0.015 input, $0.045 output, and $0.003 cache read, though the parameter count is marked as rumoured on that page.

Coverage of the Ling-3.0-tiny launch describes the model as a lightweight variant of the open-weighted Ling-3.0-flash, with roughly 8B total parameters and about 1.3B active, positioned between 4B-class and 8–12B open models such as Qwen and Gemma in downstream performance. The write-up frames the release as a way to bring near-flagship language model behavior to laptops and edge devices without depending on a cloud service. The same report adds local deployment measurements not present on the model card: about 100–105 tokens per second with FP8 quantization on an NVIDIA DGX Spark and 86–90 tokens per second on an Apple M4 Pro MacBook, with peak memory of 8.34 GiB at an 8K context length. Those hardware throughput and memory figures come from a secondary, Reddit-originated summary and are not tied to a named primary benchmark in the excerpt.

Ling 3.0 Tiny is described as the most compact member of the Ling 3.0 family from InclusionAI, Ant Group's AI initiative, using a sparse mixture-of-experts design with 7.9B total parameters of which only 1.3B are activated per token. Its attention layer combines Kimi Delta Attention and Multi-Head Latent Attention in a 3:1 alternating stack, with a sparse MoE feed-forward network of 128 routed experts where each token activates eight routed experts plus one shared expert. A central feature is dual operating modes: Thinking mode emits internal step-by-step reasoning with the chain of thought output separately, while Instant mode answers directly at lower latency, and callers toggle between them per request through an enable-thinking configuration. The model ships with a 262,144-token context window, native function calling, and prompt caching, and has been validated on consumer hardware including Apple Silicon machines and NVIDIA DGX Spark.

InclusionAI (Ant Group) published a first-party model card for Ling-3.0-tiny on the official inclusionAI HuggingFace organization, with the repository created on 2026-08-10. The card describes the model as a hybrid-reasoning mixture-of-experts architecture with 7.9B total parameters and 1.3B activated per token, built specifically for local and resource-constrained deployment, and it is a separate model in the Ling 3.0 line rather than a quantization or re-upload of Ling-3.0-flash. The card records capability benchmarks including Artificial Analysis Intelligence Index v4.1.1, an Agentic Index, and agentic, coding, long-context, math, and instruction-following results, but reports no safety evaluation. Its changelog shows the activated parameter count was revised from 1.4B (1.14B non-embedding) to 1.3B per token on 2026-08-25 across the introduction, model overview, MoE description, and evaluation sections, while total parameters stayed at 7.9B, so the earlier 1.4B figure is superseded.