← Back to feed

Alibaba’s Qwen team released Qwen3.8-27B on August 14, 2026, a dense 27-billion-parameter model under Apache 2.0 with native vision capabilities and vendor-reported agentic coding scores that exceed Claude Opus 4.6 Max across several benchmarks. The full weights are 55.6 GB. At Q4_K_M quantization, the model fits in 17 GB and runs on a 24GB consumer GPU.

The predecessor, Qwen3.6-27B, was already the benchmark leader in its size class. This release extends that lead and closes the gap to frontier-tier models on coding-specific tasks.

Key Benchmark Numbers

All figures are Alibaba-reported. No independent lab reproduction has been published as of August 17.

Benchmark	Qwen3.8-27B	Qwen3.6-27B	Opus 4.6 Max
Terminal-Bench 2.1	73.0%	63.4%	78.2%
SWE-bench Pro	61.7%	53.5%	53.4%
DeepSWE 1.1	42.2%	13.3%	—
OSWorld-Verified	84.3%	63.9%	72.7%
AndroidWorld	81.9%	70.3%	62.0%
GPQA Diamond	89.2%	87.8%	91.3%
HLE	30.8%	24.0%	40.0%
LiveCodeBench v6	90.3%	83.9%	88.8%

The DeepSWE jump is the most striking: from 13.3 to 42.2, a 3x gain over the previous 27B model. SWE-bench Pro puts Qwen3.8-27B above Anthropic’s former flagship on the metric most associated with real-world software engineering work.

Artificial Analysis places it at Intelligence Index 52.

What You Need to Run It

The minimum viable setup for local inference:

• Q4_K_M (16–17 GB): RTX 4090, RTX 3090 Ti, AMD 7900 XTX, Apple M-series with 24GB+ unified memory
• Q5_K_M (22 GB): high-end gaming cards; better quality-performance balance
• BF16 (55+ GB): H100 or equivalent

Community MLX builds for Apple Silicon were live within hours of release. GGUF files are available on Hugging Face via ggml-org.

Native context is 262K tokens. The 1M context figure in Alibaba’s materials refers to the hosted Qwen Cloud version using YaRN — it is not available in the open weights.

Multi-Token Prediction

Qwen3.8-27B ships with MTP (Multi-Token Prediction) support, which allows compatible inference runtimes to run a cheaper draft mechanism that guesses ahead, then verify with the main model. In llama.cpp with --spec-type draft-mtp, independent testing found approximately 72% throughput improvement over standard single-token generation on the same hardware. At 15–30 tokens/second on high-end consumer machines without MTP, the model runs slowly enough that hosted API services remain faster for most real-time use. MTP narrows that gap.

Context on the Numbers

Two caveats apply to every figure in this article.

First, every benchmark is Alibaba-reported and run with the Claude Code harness at specific parameters. SWE-bench Pro, DeepSWE, and NL2Repo were all evaluated using Alibaba’s own harness — not the standard mini-SWE-agent bash-only setup that produces the leaderboard figures used for cross-model comparison on swebench.com. The SWE-bench Pro score of 61.7% is not directly comparable to figures on that leaderboard.

Second, independent reproduction takes time. The weights are days old. Until a neutral evaluator runs the standard harness, the vendor numbers should be treated as directionally useful, not procurement-grade.

That said, one independent comparative test — 10 real-world tasks, Qwen3.8-27B vs Qwen3.6-27B, judged by GPT-5.5 — found the 3.8 won 9 of 10 and scored 8.8 vs 6.9 on average. The quality jump appears real. So does the cost: the 3.8 used roughly 3x the tokens and was considerably slower per task.

Hosting

OrcaRouter lists Qwen3.8-27B at $0.33/M input, $2.40/M output with a free rate-limited tier. Qwen Cloud’s hosted 1M-context version is marked “coming soon.” There is no SLA-backed production option from Alibaba as of this writing.

The model supports text, images, and video input in the open weights. The vision encoder adds roughly 1B parameters to the 27B count.

For the local AI community, this is the current ceiling in its size class. For anyone comparing it to frontier API models: it is free to run, it is slower, and its benchmark numbers need independent verification before they should drive decisions.