At 4-bit quantization, Qwen3.8-27B is 55.6GB of BF16 safetensors, and about 17 to 18GB at 4-bit
- the difference between needing a multi-GPU cluster and spinning up a single 4090 worker. That single fact is why the model hit #1 on Hacker News within hours of its weights landing on August 14.
Qwen3.8-27B, released August 14, 2026, brought Alibaba's Qwen3.8 generation to a size intended for local deployment: a 27B-parameter dense, natively multimodal model with downloadable weights under Apache 2.0. Its companion release, the 2.4-trillion-parameter Qwen3.8-Max, gets the headlines. But nobody self-hosts a 2.4T model. A 27B is a different story: it fits on a single rented GPU, it fine-tunes on realistic budgets, and its predecessor is one of the most-praised local models of the year.
What the architecture actually does differently
The core architectural change is its attention pattern. Instead of using full (quadratic) attention at every layer, the model alternates: three layers of gated Delta Net linear attention followed by one layer of full attention, repeated through all 64 layers of the network. Linear attention mechanisms like gated Delta Net scale much better with sequence length than standard full attention, which becomes expensive as context grows because every token has to attend to every other token. By handling most of the layers with a linear-attention variant and reserving full attention for a minority of layers, the architecture keeps compute and memory costs manageable even as the context window stretches into the hundreds of thousands of tokens. This is what makes the 262K native context window practical rather than theoretical.
The practical consequence for anyone running it locally is that only 16 of the 64 layers keep a per-token KV cache. That inverts the usual sizing advice. On most 27B models the context window is what breaks your budget. Here, weights dominate, and a 24GB card can carry 64K-96K of context at 4-bit without drama.
The second architectural feature that matters for agent workloads:
the multi-token prediction head ships in the checkpoint, so speculative decoding works without a separate draft model. llama.cpp runs it with --spec-type draft-mtp.
The inclusion of an MTP head appears to provide dual benefits: improved next-token prediction through auxiliary training signal, and a 15-25% inference speedup via speculative decoding.
The benchmark numbers - and what is actually verified
The launch table concentrated on coding and office agents. Qwen3.8-27B scored 61.7% on SWE-bench Pro, 42.2% on DeepSWE 1.1 and 79.0% on QwenSWEBench at release, all sizable gains over the earlier Qwen3.6-27B.
On SWE-bench Pro, a real-world software engineering benchmark, it scores 61.7% compared to 53.5% for Qwen3.6-27B. DeepSWE 1.1 sees a dramatic leap from 13.3% to 42.2%. For multi-step office workflows measured by CoWorkBench, it reaches 70.7%.
| Benchmark | Qwen3.6-27B | Qwen3.8-27B | Delta |
|---|---|---|---|
| SWE-bench Pro | 53.5% | 61.7% | +8.2 pp |
| DeepSWE 1.1 | 13.3% | 42.2% | +28.9 pp |
| Terminal-Bench 2.1 | 63.4% | 73.0% | +9.6 pp |
| OSWorld-Verified | 63.9% | 84.3% | +20.4 pp |
| CoWorkBench | - | 70.7% | - |
| GPQA Diamond | - | 89.2% | - |
Sources: Qwen model card, August 14, 2026. All scores are vendor-reported.
Now the careful version. Every launch score comes from Qwen. Several benchmarks are in-house, corrected or modified. The SWE-bench Pro comparison imports Anthropic's Opus result instead of rerunning it under Qwen's setup.
Every launch score comes from Qwen. Several benchmarks are in-house, corrected or modified.
The DeepSWE jump from 13.3% to 42.2% is striking enough that it deserves particular scrutiny. DeepSWE 1.1 is evaluated with the Claude Code harness at temp=1.0, top_p=0.95, and a 256K context window
- conditions that benefit a model with Qwen3.8's long-context architecture but that differ from how many community eval setups run. Until an independent index picks the model up and runs the same harness, the DeepSWE number is promising, not confirmed.
In a real-world architecture evaluation, Qwen3.8-Max preview scored 80/100, just behind Kimi K3 at 83. That's one of the few third-party data points available. For the 27B specifically, community replication is still forming.
Beagle in action#eng-tools, day after weights dropHardware tiers that actually work
Running the model at full 16-bit precision requires roughly 56GB of GPU memory, while an FP8 version needs about 28GB. But 4-bit quantization cuts the model itself to roughly 17GB, putting it within reach of high-end consumer machines such as a powerful gaming desktop or well-equipped laptop.
12-16GB VRAM - possible with Ridge quantization (3.7 bpw). Ridge quantization drops the model from roughly 50GB at full precision to 11.7GB without flattening every layer to the same bit depth. Context budget shrinks significantly; treat this as experimentation, not production.
24GB VRAM (RTX 4090 / 5090 24GB) - the practical local floor. Quantized versions of Qwen3.8-27B can run on a single GPU, including 24GB configurations for lighter workloads, while larger GPUs provide more room for context, concurrency, and production throughput. Aim for Q4 or Q5; leave headroom for the KV cache.
32-48GB - the comfortable tier. 24GB of unified memory or VRAM is the fit-first minimum; 32-48GB is the practical sweet spot.
FP8 on 28GB+ - Unsloth documents Qwen3.8-27B NVFP4 support in vLLM 0.25.0+ and SGLang.
One thing the hardware guides understate: long context costs memory on top of the file - about 2GB at 8K and 8GB at 32K. Budget your context window before you commit to a quantization level.
What is genuinely new versus incremental
The architecture is not novel - those numbers are not new. A direct comparison of the published Qwen3.8 configuration with the Qwen3.6 configuration finds the same model type, layer counts, hidden sizes, attention layout, vision encoder, vocabulary, and context setting. The model is primarily a post-training upgrade, not a redesign.
What is genuinely new is the combination. The launch-day conclusion is unusually clear: this is the most convincing candidate yet for the best dense, locally deployable multimodal model around 30 billion parameters.
Qwen3.8 occupies the unusually valuable middle: large enough to be broadly capable, small enough for community four-bit builds to fit on a single high-end consumer GPU at moderate context.
The competitive performance of Qwen3.8-27B against larger MoE models challenges the assumption that MoE is always superior. Dense models retain advantages in memory efficiency (no router overhead, smaller total footprint), inference simplicity (no dynamic routing decisions), and quantization compatibility (MoE experts quantize poorly).
The less obvious point: this model makes the local-agent evaluation practical for teams that could not previously afford to run it. Qwen3.8-27B is important because it makes a serious agent evaluation practical for more teams. It does not make the evaluation unnecessary. You still need to run it on your own codebase and your own tasks before committing.
Qwen3.8-27B local coding agent: common questions
What hardware do I need to run Qwen3.8-27B locally?
A single GPU with 24GB VRAM is the practical minimum. At 4-bit quantization, the model weights occupy roughly 17-18GB, leaving room for moderate context. The 32-48GB range is the comfortable sweet spot. Full BF16 precision requires around 56GB, which means multi-GPU setups.
How does Qwen3.8-27B compare to Qwen3.6-27B?
The architecture is nearly identical - same layer count, attention layout, and context window. The gains come from post-training. SWE-bench Pro improved from 53.5% to 61.7%; DeepSWE 1.1 jumped from 13.3% to 42.2%. Both scores are vendor-reported and need independent replication before treating them as definitive.
Is Qwen3.8-27B actually open source?
Open weights, not open source. Its weights can be downloaded for local or hosted deployment, subject to the published license. Open weight does not automatically mean open source: training data and training code may remain private, and commercial restrictions can still apply. The license is Apache 2.0, which is permissive for most commercial use cases - verify against your specific deployment.
What is the GatedDeltaNet attention design and why does it matter?
48 of the 64 layers use Gated DeltaNet, a form of attention whose memory use stays the same no matter how long the conversation gets. Only the other 16 layers keep a regular KV cache, the part that normally grows with every token. In practice, that means long chats and big documents cost far less memory here than on a typical 27B model. For agent workloads with large context, this materially changes what fits on your hardware.
Should I switch my local coding agent to Qwen3.8-27B now?
Wait for independent benchmark results if your workload is critical. Run Qwen's launch numbers against your own eval harness on your own codebase before switching. Your own workload results beat any leaderboard. If you were already running Qwen3.6-27B and found it useful, the 3.8 upgrade is a low-risk experiment worth scheduling - just don't let the DeepSWE headline number be the deciding factor before external validation arrives.