News · 2026-07-23

Ant's Ling-3.0-flash goes live free: 124 billion parameters, 5 billion doing the work

Ant released Ling-3.0-flash on July 23, and the headline specification is the sparsity: 124 billion total parameters, of which only 5.1 billion activate per token -- about 4%. It ships with a native 256,000-token context extendable to a million, OpenAI- and Anthropic-compatible APIs, and, for now, no price tag. It is live free on OpenRouter and free on Vercel's AI Gateway through August 3.

Key facts

• The number: 124B total parameters, 5.1B activated per token -- roughly a 4% activation ratio.
• Who and when: Ant, released July 23, 2026, positioned for long-horizon agent workflows and stable tool calling.
• Access: inclusionai/ling-3.0-flash:free on OpenRouter at zero cost; free on Vercel AI Gateway through August 3.
• Primary source: Ant's model documentation.

The activation ratio is the story. A mixture-of-experts model holds a large pool of specialist sub-networks and routes each token to only a few of them. The total parameter count determines roughly how much the model can know; the activated count determines roughly what it costs to run. At 4.1%, Ling-3.0-flash is on the aggressive end of that trade -- it carries the storage footprint of a very large model while doing the arithmetic of a small one.

The kitchen analogy holds up well here. A restaurant with 124 cooks does not put all 124 on every dish. It routes each order to the four or five who specialize in it. Your bill reflects the cooks who touched your plate, not the payroll.

Ant positions the model explicitly for long-horizon agent workflows and stable tool calling, which is the right target for these economics: agent loops make many calls, and cost per call is the constraint that decides whether a workflow is viable. OpenRouter's live record lists 262,144 tokens of context, zero pricing on both prompt and completion, and support for tools and tool_choice. Ant's platform documentation adds a global ceiling of two requests per second and a 90-second synchronous timeout, with streaming recommended -- a real constraint for anyone planning heavy parallel agent traffic, and the sort of detail that only shows up if you read the rate limit page.

On architecture, Ant discloses less than the launch buzz suggests. The 3.0 series moves from what it calls a migration-based hybrid-linear architecture to a native hybrid-linear design, uses lower expert-activation ratios, and adds KDA (Kimi Delta Attention) for memory updating and information retention across long sequences. Expert count, routing rule, layer mix, and training recipe are all unpublished. There is no technical report for this model.

Several claims circulating about Ling-3.0-flash do not hold up, and they are worth listing because they affect how you would actually use it.

Its "hybrid reasoning" is verified only at a high level: Ant says deep thinking can be enabled or disabled, and its model page describes switching between thinking and non-thinking modes. A specific thinking.type request field with a default of enable has been widely repeated but does not appear in Ant's OpenAI-compatible API reference. OpenRouter's metadata showing reasoning default_enabled: false is gateway configuration, not evidence of Ant's native default, and nothing supports the claim of automatic difficulty-based reasoning.

The correct OpenRouter identifier is inclusionai/ling-3.0-flash:free, with the suffix. The unsuffixed slug is not the free listing.

It should not be called open-weight. OpenRouter's record lists no associated Hugging Face repository, and no Ling-3.0-flash checkpoint or license was located. Ant's general statement that the Ling series is open-sourced does not substitute for a model-specific release. The one published technical report in this family, arXiv:2606.15079, is explicitly about Ling and Ring 2.6 -- a different generation -- so no benchmark or methodology claim about 3.0-flash can be sourced from it.

The free window also needs care. OpenRouter shows free pricing with no listed expiration. Vercel's announcement says free on AI Gateway "through August 3rd," while the same page elsewhere says "the next three weeks," which is internally inconsistent. Treat August 3 as Vercel's published cutoff for Vercel, and do not extend it to OpenRouter.

The strongest counter-argument to the enthusiasm: a low activation ratio is an inference-economics claim, not a capability claim, and with no technical report and no independent evaluation there is nothing here to assess but vendor positioning. Kilo's launch post cites early testing without methods or results. Vercel's listing verifies that the model is available and routable, not that it is good. Claimed independent benchmark runs on Reddit could not be verified.

The honest caveat, then: this is a genuinely testable free model with unusually attractive economics and a large context window, and it is worth putting real agent traffic through this week while it costs nothing. It is not yet a substantiated open-weight release, and nobody outside Ant has published a number about how well it works.

Primary source, verified: read the paper →

Ground Truth. (2026, July 23). Ant's Ling-3.0-flash goes live free: 124 billion parameters, 5 billion doing the work. Ground Truth. https://groundtruth.day/news/ling-3-0-flash-free-sparse-agent-model.html

@misc{groundtruth:ling-3-0-flash-free-sparse-agent-model,
  title  = {Ant's Ling-3.0-flash goes live free: 124 billion parameters, 5 billion doing the work},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/ling-3-0-flash-free-sparse-agent-model.html}
}

Topics: models · mixture-of-experts · agents · api · efficiency