Qwen3.8-27B Outperforms Opus 4.6 Max at 27B Parameters
On August 14, 2026, Alibaba released Qwen3.8-27B, a 27-billion-parameter open-weight model under Apache 2.0 (Hugging Face model card). The checkpoint landed at 15:00 UTC per ModelScope metadata (Kingy AI). It scores 61.7 on SWE-bench Pro against 53.4 for Anthropic’s Opus 4.6 Max, and 84.3 on OSWorld-Verified computer use against 72.7 (Local AI Zone). Those numbers are last-generation frontier quality at a size that fits on one consumer GPU.
Four days earlier, Meta shipped Muse Glimmer as the best open 30B model. Qwen3.8-27B leads it on every overlapping benchmark: Terminal-Bench 2.1 at 73.0 against 51.7, SWE-bench Pro at 61.7 against 51.2, and OSWorld-Verified at 84.3 against 65.9 (Local AI Zone).
What shipped
Section titled “What shipped”The official repository is Qwen/Qwen3.8-27B on Hugging Face (Hugging Face).
| Spec | Value |
|---|---|
| Parameters | 27.78B dense (27,781,427,952) |
| License | Apache 2.0 |
| Architecture | Hybrid decoder, 64 layers |
| Attention mix | 48 Gated DeltaNet layers, 16 full-attention layers (3:1) |
| Hidden / FFN size | 5,120 / 17,408 |
| Native context | 262,144 tokens |
| Extended context | ~1M tokens via YaRN |
| Inputs | Text, image, video |
| Output | Text |
| Decoding | Multi-token prediction (MTP) head |
Sources: Kingy AI, NVIDIA NeMo docs, Local AI Zone.
Why the architecture matters
Section titled “Why the architecture matters”Most dense models use full attention on every layer. Qwen3.8-27B repeats one pattern 16 times: three Gated DeltaNet linear-attention layers, each followed by an FFN, then one grouped-query full-attention layer and another FFN (Kingy AI). NVIDIA’s NeMo docs confirm the hybrid design carries over from the Qwen3.5 generation (NVIDIA NeMo).
Linear attention is where the win comes from. It does not grow quadratically with sequence length, so a 262K-token context window stays affordable on a single GPU. The periodic full-attention layers keep retrieval precision where it matters (Kingy AI). A multi-token prediction head acts as a built-in speculative decoder, which speeds generation (Local AI Zone).
Two developer controls define the release:
reasoning_effort— set thinking to low, medium, or extra high per request (RepoChad)preserve_thinking— keep the reasoning trace across turns for multi-step agents (RepoChad)
Thinking is on by default, and the recommended sampling runs at temperature 1.0 with top-p 0.95 (GGUF community card).
The benchmarks
Section titled “The benchmarks”The model card reports large jumps over the Qwen3.6-27B predecessor (Kingy AI):
| Benchmark | Qwen3.8-27B | Qwen3.6-27B | Delta |
|---|---|---|---|
| Terminal-Bench 2.1 | 73.0 | 63.4 | +9.6 |
| DeepSWE 1.1 | 42.2 | 13.3 | +28.9 |
| OSWorld-Verified | 84.3 | 63.9 | +20.4 |
| SWE-MM | 38.6 | 25.7 | +12.9 |
Head-to-head against closed frontier and open competitors (Local AI Zone):
| Benchmark | Qwen3.8-27B | Opus 4.6 Max | Muse Glimmer 30B |
|---|---|---|---|
| SWE-bench Pro | 61.7 | 53.4 | 51.2 |
| OSWorld-Verified | 84.3 | 72.7 | 65.9 |
| Terminal-Bench 2.1 | 73.0 | — | 51.7 |
Local AI Zone counts 15 wins in 19 direct comparisons against Claude Opus 4.6 (Local AI Zone).
One honesty note. Qwen compared against Opus 4.6, not the current Opus 5, which sits at 79.2 on the same SWE-bench Pro leaderboard (DIY Smart Code). This is not a frontier killer. It is an open 27B closing on last-generation frontier quality (DIY Smart Code).
What it means for your agent stack
Section titled “What it means for your agent stack”The practical case is simple. You can serve this model from vLLM, SGLang, Ollama, or Unsloth, and quantized builds run on one consumer GPU (DIY Smart Code). The Q4_K_M GGUF lands around 16.8 GB, sized for a 24 GB RTX 3090 (GGUF community card). Local AI Zone puts the deployment floor at 24 GB VRAM (Local AI Zone).
That changes the cost math for agent workloads. A single agent task can make dozens of API calls, and every token is metered. A local 27B turns bounded agent work into a fixed hardware cost instead of a per-token bill (DIY Smart Code).
Rules for running it
Section titled “Rules for running it”- Pull the FP8 or GGUF quant for a single-GPU box. BF16 needs more memory than most consumer cards carry (Kingy AI).
- Keep static YaRN off below 32K-token prompts. Open-source YaRN is static and costs precision at short lengths (DIY Smart Code).
- Set
reasoning_effortper task, not globally. Low effort on hard steps causes failed retries, which cost more than the thinking you saved (DIY Smart Code). - Route by difficulty. Use the local model for bounded coding, private multimodal work, and cost-sensitive routing. Escalate the hardest failures to a frontier endpoint (Kingy AI).
The bottom line
Section titled “The bottom line”The week’s review calls Qwen3.8-27B state of the art for its size, with performance comparable to Opus 4.6 Max (AI Week in Review). For DevOps teams, the question is no longer whether a local agent can do the work. It is how much of the pipeline you want to keep on hardware you control.