inclusionAI: Ling 3.0 Tiny

OpenRouter's public catalogue now includes the route inclusionai/ling-3.0-tiny, with the listing's release date given as 6 August 2026 and an AIMI endpoint monitor first observing the route the same day. The page records the model as a 7.9B-total mixture-of-experts design with 1.3B active parameters, a 262,144-token context window, and a 32,768-token maximum output. Metadata flags on the route mark reasoning, token streaming, and function calling with tools as supported, while several other capabilities remain unconfirmed. The article cautions that provider metadata and route pricing can shift before production use, so the confirmed details are the architecture, context and output limits, and the August 2026 launch timing.

InclusionAI's Ling 3.0 Tiny is a mixture-of-experts model with 1.3B active parameters out of 7.9B total, built for responsive agents, instruction following, and multi-turn conversations. Its catalogue entry lists a 262,144-token context window and an August 6, 2026 release date. The model offers switchable thinking and instant modes, letting callers trade reasoning depth against response latency. The Ling 3.0 Tiny listing also surfaces the wider inclusionAI text lineup, including Ling 3.1 Flash, Ling 3.0 Flash, and Ling 3.0 Flash VL as separate text models. Activity panels for the Tiny variant show no token volume data yet, so real-world usage telemetry for this exact variant is not yet established.

AI Indigo records the open-source release of InclusionAI Ling 3.0 Tiny on November 5, 2026, noting the mixture-of-experts model was listed on OpenRouter that day with a 262,144-token context window. The same entry points to open weights published on Hugging Face, giving the model a dated public availability milestone alongside its MIT-licensed weight distribution. The event listing frames the release within a broader InclusionAI cadence, referencing Ling 3.1 Flash and the Ling-3.1-flash release as earlier entries from the same vendor. Those sibling releases are noted only as timeline context; the substantive claims here remain the November 5 listing date, the 262K context, and the open-weight release.

Artificial Analysis lists Ling 3.0 Tiny from InclusionAI as an open-weights reasoning model with a 262k token context window, 7.9B total and 1.3B active parameters, MIT license, and text-in/text-out modalities, linking to Hugging Face weights. The listing notes that a non-reasoning variant may also exist for this model. Measured serving data on that page is sparse: the Intelligence Index, cost per task, and verbosity fields are marked unavailable because no publicly available API was used, while latency figures show 34.83 seconds to first answer token and 85 tokens per second output speed. Displayed pricing of $0.00 per 1M input and output tokens reflects the absence of public API access rather than an actual rate.

InclusionAI (Ant Group) published Ling-3.0-tiny as a separate model in its Ling 3.0 line, with its own architecture and model card rather than a quantization or re-upload of Ling-3.0-flash. The official card, hosted on the inclusionAI Hugging Face org with the repository created on 2026-08-10, describes it as a hybrid-reasoning mixture-of-experts model built for local and resource-constrained deployment. The card reports 7.9B total parameters with 1.3B activated per token, open weights under a permissive MIT-style license, and capability benchmarks including the Artificial Analysis Intelligence Index v4.1.1, Agentic Index, and agentic, coding, long-context, math and instruction-following results. It records no safety evaluation, and archived revisions show the activated parameter count being corrected between 1.4B and 1.3B entries.

Independent benchmark data for InclusionAI's Ling 3.0 Tiny, attributed to Artificial Analysis, puts it in tier Efficient with a global rank of 249 out of 508 LLMs and an Intelligence Index of 11.1, plus an output speed of 63 tok/s. The page notes Ling 3.0 Flash released two days before this model. Per-skill results are weak in several areas: Agentic 0.1, Terminal-Bench 4.0 with 0% completion, AutomationBench 0%, GDPval 3%, GDP.pdf 1%, and Humanity's Last Exam 9%. Stronger showings appear in GPQA Diamond at 73%, long-context reasoning at 60%, SciCode at 24%, and a Briefcase Elo of 476.

Ling 3.0 Tiny is a 7.9B mixture-of-experts model from InclusionAI that activates only 1.3B parameters per token, giving it a low active-compute ratio aimed at responsive agents, instruction following, and multi-turn conversation workloads. It exposes switchable Thinking and Instant modes so callers can pick deeper reasoning or lower latency per request. The model also adds native function calling and prompt caching for tool-using and long-context work, with a 256K context window and up to 32K maximum output, all in text input and text output. Reported evaluation figures include 11.1% on the Intelligence Index, 26.5% Coding Index, 4.4% Agentic Index, 73.4% GPQA Graduate Science, and 24.2% SciCode.

Ant Group's InclusionAI released Ling-3.0-tiny on Hugging Face as a 7.9B-total, 1.2–1.3B-active mixture-of-experts model using a KDA linear-attention plus MLA hybrid attention stack, the smallest member of the Ling-3.0 family. BF16, FP8, and INT4 checkpoints ship simultaneously under the MIT license, with same-day support from SGLang, vLLM, and Ollama. Unlike Ling-3.0 Flash, which uses 35 KDA and 7 Gated MLA layers with a 5:1 ratio and 512 routed experts, tiny uses a 3:1 ratio with 3 KDA plus 1 MLA layer per four-layer block and 128 routed experts, 8 active plus one shared. MLA caches a 576-dim compressed latent directly, cutting KV memory about 10x.