Qwen3.8-27B: Frontier Agentic Scores on a Single Consumer GPU

Alibaba released Qwen3.8-27B on August 14, 2026 — a 27-billion-parameter dense vision-language model under Apache 2.0, published on Hugging Face and ModelScope. It is the small sibling of the 2.4-trillion-parameter Qwen3.8-Max whose weights went up the day before, and on several agentic benchmarks it does not behave like a small model at all: 61.7 on SWE-bench Pro against Opus 4.6 Max’s 53.4, and 84.3 on OSWorld-Verified against 72.7. The 4-bit quantisation is roughly 17 GB, which puts those numbers on a single 24 GB consumer card.

Advanced

The Architecture Is the Story

The model card gives the layer layout explicitly, and it explains most of what follows:

16 × (3 × (Gated DeltaNet → FFN) → 1 × (Gated Attention → FFN))

That is 64 layers, of which only 16 use full softmax attention. The other 48 use Gated DeltaNet, a linear-attention variant that carries a fixed-size recurrent state instead of a key-value cache that grows with sequence length. The full-attention layers use 24 query heads against 4 key-value heads at head dimension 256; the DeltaNet layers use 48 value heads and 16 QK heads at head dimension 128. Hidden dimension is 5,120, the FFN intermediate is 17,408, and the padded vocabulary is 248,320. The model was trained with multi-token prediction across multiple steps.

The consequence for deployment is that context length stops being the thing that blows up your VRAM budget. Only a quarter of the layers accumulate a KV cache at all, so the cache footprint at long context is roughly a quarter of what a conventional dense 27B would need. Native context is 262,144 tokens, extensible to one million via YaRN. Community measurements circulating since release put the cache near 0.5 GB at 8K tokens and about 2 GB at 32K — figures worth verifying against your own setup rather than taking as published specifications.

Benchmarks

Qwen’s own table compares Qwen3.8-27B against its predecessor Qwen3.6-27B, against the larger Qwen3.7-Plus, and against Opus 4.6 Max. The generational jumps are the largest the 27B line has seen:

• Terminal Bench 2.1 — 73.0, up from 63.4. Opus 4.6 Max leads at 78.2.
• SWE-bench Pro — 61.7, up from 53.5, ahead of Opus 4.6 Max’s 53.4.
• DeepSWE 1.1 — 42.2, up from 13.3. A 3.2× jump on the same parameter count.
• OSWorld-Verified (computer use) — 84.3, up from 63.9.
• AndroidWorld — 81.9, up from 70.3, against 62.0 for Opus 4.6 Max.
• SWE-MM (multimodal software engineering) — 38.6, up from 25.7.
• LiveCodeBench v6 — 90.3, up from 83.9.

The pattern holds on the reasoning benchmarks but with a smaller margin, and the ceiling shows: GPQA Diamond is 89.2 against Opus 4.6 Max’s 91.3, and Humanity’s Last Exam is 30.8 against 40.0. Two caveats on reading the table. QwenSWEBench, where the model posts 79.0 against the predecessor’s 49.3, is Qwen’s own benchmark, and self-authored evaluations reward the authoring lab’s training distribution. And several of the multimodal scores — MathVision at 94.6, BabyVision at 85.6, CharXiv at 90.2 — are the with-code-interpreter figures; the tool-free numbers are 90.0, 65.7 and 83.7 respectively. The BabyVision gap in particular is 20 points of tool use, not 20 points of model.

Running It

Unsloth’s deployment table puts 4-bit at 17–19 GB, 6-bit at 24 GB, 8-bit at 31 GB and BF16 at 56 GB, with 2-bit down at 11–13 GB for anyone willing to trade accuracy for a 12 GB card. Community GGUF builds appeared within a day — bartowski’s Q4_K_M is about 17 GB — and llama.cpp, LM Studio, Ollama and Jan all list it. On the server side the weights ship as BF16 safetensors compatible with vLLM, SGLang and Transformers.

The German technology publication heise online tested the model on an RTX Pro 6000 Blackwell under llama.cpp and reported that NVFP4 used roughly half the VRAM of Q8_0 at comparable accuracy. Asked to build a complete REST API for an inventory system without clarifying questions, the model produced code that compiled on the first attempt; heise called the result “indistinguishable from Claude’s programming skills,” while noting a tendency to overthink simple problems. Thinking mode is on by default and can be disabled per request, with a reasoning_effort parameter accepting low, medium or xhigh.

What This Means

Three weeks ago we covered Kimi K3 shipping open weights at 2.8 trillion parameters and 1.56 TB of download — a genuine milestone that essentially nobody outside a datacentre can run. Qwen3.8-27B is the opposite trade. It gives up the frontier on hard reasoning, and the HLE gap is not close. But for agentic coding and computer use — the workloads where an open model actually displaces an API subscription — it lands within a few points of models that cost per token and run on someone else’s hardware.

The architectural choice is what makes that possible. Restricting full attention to 16 of 64 layers is a bet that most of a long context does not need quadratic attention, and it buys back exactly the resource that has kept long-context work off consumer GPUs. Whether the linear-attention layers cost something the benchmarks do not measure — retrieval precision deep in a 262K window, say — is the question the next few weeks of community testing should answer.

Related Coverage

• Qwen3.8-2.4T-A95B: Alibaba Open-Weights Its Max-Tier Flagship — the 27B’s larger sibling, released one day earlier
• Qwen3.8-Max Draws Level With the Frontier on Agentic Benchmarks — independent evaluation of the Max tier’s agentic performance
• Kimi K3 Open Weights Ship: 2.8T Parameters, 1.4 TB to Run — the other end of the open-weights size spectrum

This post was drafted with AI assistance and reviewed by RITS staff.

Sources

• Qwen3.8-27B model card — Hugging Face (architecture, benchmark tables, licence)
• Qwen3.8 — How to Run Locally — Unsloth documentation (quantisation and memory table, sampling settings)
• bartowski/Qwen3.8-27B-GGUF — community GGUF quantisations
• Trying out local AI: This is what Qwen3.8-27B can do — heise online, independent hands-on test
• Qwen3.6-27B vs Qwen3.8-27B comparison — llm-stats.com