The Open-Weights Model That Doesn't Go Quiet

Qwen3.8-27B went open-weights in mid-August, and no host we could reach was serving it. The vendor that announced it hadn’t shipped, and none of the usual clouds carried it, so we ran the weights ourselves for as long as it took to measure, then took it down.

It beat both open-weights models already on our LLM board on every axis we measured. Both of those, gemma-4-31b and gpt-oss-120b, are served by Cerebras, and they are the two fastest rows on the board.

It finishes the job

93% of multi-step tool tasks completed end to end, against 80% and 73%. 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.

The half that impressed us more is restraint. Two of the six scenarios are traps where the correct move is to call nothing and ask a question, because acting means inventing an order ID. It held back on nine of ten runs. 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 fifty 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 three-family judge panel, and every one came back a live 3-of-3 pass, not an abstention.

Asked for a discount code it didn’t have, it said it didn’t have one and couldn’t invent one. gpt-oss-120b answers 16% of those probes with something it made up.

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 91% of its tool calls, gpt-oss-120b on 100% — 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 speaks on 85% of its tool calls, and its overall dead-air rate is 7% against 38% and 55%.

None of this cost it capability, which is the part worth noticing — the quiet model is usually the one that traded bridging text for something. This one leads on task completion too.

What we could not measure

Its first-token latency, and that gap is real rather than an oversight. Latency belongs to the serving stack far more than to the weights: hardware, precision, kernels, batching. gemma-4-31b’s 192ms comes from the wafer-scale Cerebras silicon it runs on, which is not a number the weights alone can produce anywhere else.

The behaviour numbers don’t have that problem. The board carries the same gemma weights served two different ways, and the two land within a few points of each other — which is the control that lets these axes be read across deployments.

Two notes on how these were run, since both differ from the board’s documented default: five runs per scenario rather than three, and every model here served at FP8, the precision its own vendor publishes. Thirty runs ranks a field but won’t split two models a few points apart, so treat the 93-against-80 gap as real and anything tighter as noise.

When a host picks it up, the latency cell fills in and the row ranks. Until then, three of the four things that decide whether a voice agent works are measured, and it wins all three.