The number that matters in the Qwen3.8-27B drop isn’t a benchmark — it’s the memory footprint. This checkpoint fits on a single workstation GPU, and that changes the deployment math before any score does.
- 🎯 27.78B params, Apache-2.0, multimodal across text/image/video, 262K context — and the FP8 weights are already quantized to land near 27GB, or 14–16GB at 4-bit.
- ⚡ The agentic-coding and computer-use gains are the headline: Terminal-Bench 2.1 climbing 63.4 → 73.0 and OSWorld-Verified 63.9 → 84.3 over the previous 27B.
- 📊 It reportedly edges Opus 4.6 Max on CoWorkBench (70.7 vs 68.2) — real if it holds, but that’s Qwen’s own scoreboard.
- ⚠️ Every launch score is vendor-reported, and several benchmarks are in-house or “corrected.” Treat them as hypotheses to run through your own eval harness, not as results.
- 🔍 For production the question isn’t peak score, it’s whether a local 27B clears your task’s bar at a fraction of the per-token cost and with none of the data leaving your VPC.
The wider reaction splits along the line you’d expect. Pat McGuinness calls it near-SOTA reasoning brought home to local use — the release Llama 4 should have been — and OfficeChai runs with the same competitive framing, that a locally-deployable model is beating larger rivals. Kingy AI is the useful skeptic: the new local model to beat, sure, but not an honest one-for-one swap for frontier APIs, and every number traces back to Qwen with no independent verification. The HN discussion carries the same tension — the excitement is about deployment economics and open weights, and the caution is about believing the scoreboard before you’ve run it against your own tasks.