← All models

Ling-3.0-flash

Language Available Full comparison ↗

ant-group/ling-3-0-flash · by Ant Group · mixture-of-experts

Pricing — 1 offering(s)

OpenRouter

Input tokens

  • $0.00 / 1M tokens (input) Historical 2026-07-24 → 2026-08-09
  • $0.021 / 1M tokens (input) Current 2026-08-10 → present

Output tokens

  • $0.00 / 1M tokens (output) Historical 2026-07-24 → 2026-08-09

    Same launch-promotional caveat as the input tier.

  • $0.063 / 1M tokens (output) Current 2026-08-10 → present

Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.

See how Ling-3.0-flash fits into a cost-aware routing setup

See how →

Capability profile

How Ling-3.0-flash rates across core capability dimensions, with the task-level evidence behind each rating.

coding Moderate

inclusionAI's own reporting: SWE-Bench Pro ~56.6, SWE-Bench Multilingual ~72.4. Ant's launch claim is that Ling-3.0-flash "matches or beats" its own 1T-parameter flagship (Ling-2.6-1T) on most benchmarks shown — a vendor claim, no independent reproduction.

reasoning Moderate

Hybrid-reasoning (toggleable). Positioned for production agents; the dedicated Ring family targets heavier reasoning. No independent reasoning-benchmark data found.

tool use Moderate

"Built for production-scale agents" per Ant's own announcement; no agentic-benchmark score published.

instruction following Moderate

Ant reports parity with its 1T flagship on instruction-following tasks (vendor claim).

context window Moderate

262,144 tokens native; 1M is a stated goal, not shipped.

multilingual Moderate

SWE-Bench Multilingual result suggests reasonable cross-language coding; general multilingual not separately benchmarked.

speed Strong

~5.1B active parameters plus hybrid-linear attention — very cheap and fast to serve for its capacity; the model's headline pitch.

cost efficiency Strong

Ratings are estimated — limited independent data is available for this model.

Operator guidance

Route here when cost and serving efficiency are the overriding constraints and the task is within reach of a ~5B-active model — Ant's efficiency play. For heavier reasoning, use the sibling Ring models or a larger flagship. All capability signal is vendor-reported (Ant's own announcement / inclusionAI benchmarks); there is no first-party Ant pricing console and no independent benchmark reproduction, so treat the "matches our 1T flagship" claim as directional.

Use cases

Limitations

Citations