The Open-Weights Model That Doesn't Go Quiet
Qwen3.8-27B on Cerebras speaks on half its tool calls, where the two open-weights models beside it speak on one in twenty and none at all. It also leads them on task completion and invents nothing across thirty probes, and it pays a few milliseconds of first token for it.
Three open-weights models, all three served by Cerebras, all three measured through the same harness on the same corpus. Two of them are the fastest rows on our LLM board. The third is the only one that says anything while it works.
It stays audible
This is the number with the most direct line to what a call sounds like. When an agent looks something up, the caller either hears “one moment, checking that now” or hears nothing. gemma-4-31b goes quiet on 94.6% of its tool calls. gpt-oss-120b goes quiet on 100% of them — every single action, dead air.
A caller who hears nothing doesn’t wait patiently. They assume the line dropped and start talking again, into an agent that isn’t listening yet, and the turn collapses.
Qwen3.8-27B is silent on 48.7%. That is not quiet — it is half. But the comparison is not against silence, it is against the field: it speaks on roughly half its tool calls where the other two speak on one in twenty and none at all. Its overall dead-air rate, counting every turn rather than only the ones where a tool runs, is 33.2% against 53.3% and 68.2%.
It finishes the job
87.72% on the graded task score, against 85.00% and 77.72%. These aren’t single tool calls — they’re chains, with synthetic tool results fed back mid-conversation, a compound request that hides two jobs in one sentence, and a caller who applies pressure to skip a lookup.
Read the gap to gemma-4-31b as real and the gap to gpt-oss-120b as noise. Ten points is a difference; under three is inside what we have watched a single model move between two runs of the same corpus.
The half that impressed us more is restraint. Several scenarios are traps where the correct move is to call nothing and ask a question, because acting means inventing an order ID. That’s the axis where capable models usually fail — the reflex to be helpful is exactly what fires a tool with a made-up argument.
It doesn’t make things up
Zero fabrications across thirty out-of-policy probes: closing times it was never told, a warranty length nobody gave it, a loyalty balance it has no way to look up, a phone number for an executive. Every run was graded by a judge panel drawn from a different model family than the one under test.
gemma-4-31b is also at zero, so this is not what separates the two. gpt-oss-120b answers 30% of those probes with something it made up — asked for a discount code it doesn’t have, it produces one.
What it costs
103ms to first token, against 100ms and 78ms. It is the slowest of the three, and the ordering is not an accident: Qwen3.8-27B reasons by default, and it is served with that reasoning dialled down rather than off, which is the setting a caller actually gets. gemma-4-31b does no such thinking and is 25ms quicker for it.
Three milliseconds behind gpt-oss-120b is not a trade anyone needs to think about. Twenty-five behind gemma-4-31b is a real one, and it buys a model that speaks through its tool calls instead of leaving the line dead.