Qwen3.8-27B Is the New Default Local Model Alibaba's Qwen team released Qwen3.8-27B, a dense, natively multimodal 27B model under Apache 2.0 with weights on Hugging Face, designed for single-GPU agentic work. The model features a 3:1 hybrid of Gated DeltaNet linear attention and full gated-attention blocks, supporting 262,144 tokens native context, and shows major agentic benchmark gains over its predecessor Qwen3.6-27B, including OSWorld-Verified at 84.3 and SWE-bench Pro at 61.7. However, vendor-run evals and full-precision measurements warrant caution until third-party tests. AI https://sourcefeed.dev/c/ai Article Qwen3.8-27B Is the New Default Local Model Alibaba's Apache 2.0 dense release targets single-GPU agentic work — and its license split signals where open weights are heading. Rachel Goldstein https://sourcefeed.dev/u/rachel goldstein Alibaba's Qwen https://qwen.ai team shipped two models this week, and the headlines mostly went to the wrong one. Qwen3.8-Max is the flagship — a 2.4-trillion-parameter mixture-of-experts monster with roughly 95B active parameters, available API-first with open weights promised later. It needs more than a terabyte of memory just to load. Nobody reading this is running it. The release that matters is Qwen3.8-27B https://huggingface.co/Qwen/Qwen3.8-27B : a dense, natively multimodal 27B model, Apache 2.0, weights on Hugging Face today. Its predecessor, Qwen3.6-27B, quietly became the default homelab and on-prem model over the past year — the best compromise between capability and consumer hardware. This release is Alibaba defending that position, and on the evidence so far, keeping it isn't in much doubt. The architecture is the interesting part The benchmark table is what everyone screenshots, but the layer diagram is where the real decision was made. Qwen3.8-27B is 64 layers arranged in a 3:1 hybrid — three Gated DeltaNet linear attention blocks for every one full gated-attention block. Only a quarter of the layers maintain a conventional KV cache. That ratio is why the context claims are credible on hardware that isn't a DGX. The card lists 262,144 tokens native, extensible to a million with YaRN. On a fully quadratic 27B dense model, a 262K context would drown a consumer GPU in KV cache; with linear attention carrying 48 of the 64 layers, long-context local inference stops being a paper exercise. Add multi-token prediction training — which feeds speculative decoding, exactly where local token throughput hurts most — and this reads like a model designed backwards from the constraint "must be usable on one GPU," rather than a big model shrunk until it fits. Thinking mode is on by default, with a reasoning effort dial low, medium, xhigh and a preserve thinking option that carries reasoning context across turns. That last one is aimed squarely at agent loops, and the post-training clearly is too: the card's biggest generational jumps are agentic — OSWorld-Verified computer use at 84.3 versus 63.9 for Qwen3.6-27B, SWE-bench Pro at 61.7 versus 53.5, and a claimed 73.0 on Terminal Bench 2.1. About those numbers The card doesn't just compare against its predecessor; it claims wins over Claude Opus 4.6 Max on SWE-bench Pro, LiveCodeBench v6, and most of the vision-agent suite. Hacker News reacted the way it should: "There has to be something wrong with those benchmarks," as one commenter put it, "I find it extremely hard to believe a 27B model can work similar or exceed opus 4.6." Three grains of salt, in descending size. First, these are vendor-run evals, harness and all — the Terminal Bench number is under the Terminus harness specifically, and agentic benchmarks are notoriously harness-sensitive. Second, the comparison column is Opus 4.6, not the current Anthropic frontier — a classic vendor-benchmark move. Third, and most practically: those scores were measured at full precision, and you will not be running full precision. A Q4 quant on a 24GB card is a different model than the one in the table. None of that makes the release hype. The generational deltas against Qwen3.6-27B are enormous and internally consistent, and Alibaba's agentic post-training pedigree is real. The honest read: this is very likely the strongest open dense model at this size, and the Opus-parity framing is marketing until third-party evals land. Both things can be true. What running it actually looks like The official FP8 build https://huggingface.co/Qwen/Qwen3.8-27B-FP8 uses fine-grained block-128 quantization with, per Qwen, near-identical metrics — but at ~1 byte per parameter you're still looking at roughly 27GB of weights, so FP8 is for 40GB+ cards or multi-GPU boxes. Serving is one line on vLLM https://docs.vllm.ai vllm serve Qwen/Qwen3.8-27B-FP8 or SGLang https://github.com/sgl-project/sglang , plus a Docker Model Runner path. On consumer hardware, community numbers put Q4 K M GGUF builds around 16GB — comfortable on a 24GB GPU with room for context — while 16GB cards land in IQ4 XS/Q3 territory at 13–15GB with modest context. Spill layers into system RAM and you're near 12 tok/s, which is where thinking-on-by-default gets expensive: an xhigh reasoning budget at local speeds means minutes of