Ling-3.0-flash

AA coding index 57.0 (Artificial Analysis, frozen since 2026-09-11)

inclusionAI's open MoE (Ling 3.0; 124B total, 5.1B active, hybrid reasoning on by default). AA now scores it — coding index 50.6 — and our first-party BCB-Hard (36%) confirms it punches above its active size. Free in opencode Zen's rotating set; on OpenRouter the :free route rotated over to Ling 3.0 Tiny, so flash is paid-only there now ($0.08/$0.22 per M).

BigCodeBench-Hard pass@1 25% (37/148), via official protocol · llama.cpp Q4_K_M · one request at a time · a shared cluster GPU · mean of 3 runs, range 0.0 pp. The brutal counterpart to HumanEval — where HumanEval saturates near the top, BCB-Hard spreads the field, so this is the number that actually separates coding ability.

88.2% ± 2.1 over 1390 problems

Q4_K_M · a shared cluster GPU · KV f16 · 1 run. This is the non-agentic half of BFCL: does the model call the right function with the right arguments. It says nothing about how it behaves over a long agentic conversation, which is a separate measurement. The score is BFCL’s own: the unweighted mean of four groups (simple, multiple, parallel, parallel multiple), where simple is itself the mean of Python, Java and JavaScript. Every group counts the same whatever its size, and irrelevance detection — which we also measure, another 240 problems — is not part of it. The margin follows that same weighting.

~70 GB at Q4 — a single 80 GB card or a 2-GPU split; community GGUFs exist but are days old and unvetted. We tested it hosted, back when it rode the OpenRouter <code>:free</code> route.