Opus 4.6-Level Coding at 27B: A First Look at Qwen 3.8 27B

Lukas Vogel

Lukas Vogel

Applied Research Editor

Published: August 17, 2026
Qwen 3.8 27B Release Coverage editorial cover

TLDRA signal-based analysis of Qwen 3.8 27B and the latest release discussion.

Qwen 3.8 27B is a 27-billion-parameter open-weight multimodal dense model developed by Alibaba's Qwen team that runs locally on 17GB of RAM and posts frontier-level coding scores. It was released on August 14, 2026 under the Apache 2.0 license, with a native 262K-token context window extensible to 1M tokens via YaRN. Its own model card claims 61.7 on SWE-bench Pro, above the 53.4 reported for Claude Opus 4.6 Max, and it became the #1 trending model on Hugging Face within about two days of launch.

Updated 2026-09-20: A Qwen-derived 1.76-bit quantization is now being tested on 12GB GPUs (see the Update below).

Key Takeaways

  • Qwen 3.8 27B is open-weight under Apache 2.0, with weights published on Hugging Face and free to self-host and fine-tune.
  • It is a native vision-language model, understanding images and video through an integrated vision encoder, not a bolt-on adapter.
  • It runs on 17GB of RAM/VRAM via Unsloth's Q4_K_M GGUF, fitting a single 24GB consumer GPU.
  • Native context is 262,144 tokens, extensible to 1,000,000 via YaRN scaling.
  • Self-reported coding scores rival closed frontier models — 61.7 SWE-bench Pro, 90.3 LiveCodeBench — though independent verification was still pending at publication.
  • It defaults to "xhigh" reasoning effort, which reviewers found chronically over-thinks simple prompts.

What Is Qwen 3.8 27B?

Qwen 3.8 27B is the compact, self-hostable member of Alibaba's Qwen 3.8 model family. It is a causal language model with an integrated vision encoder, built on the architectural foundation of Qwen 3.5. Alibaba positions it as "the most capable generation in the Qwen open-model family to date," succeeding the widely adopted Qwen 3.6 27B.

The release status is confirmed. Alibaba's Qwen and Alibaba Cloud accounts published the weights on Hugging Face on August 14, 2026, following a pre-release countdown page that had drawn over 5,600 people waiting for the drop. The model card describes it as "a native multimodal dense model" that "outperforms Qwen3.7-Plus overall and shines in real-world coding & office workflows."

It sits apart from its sibling, Qwen 3.8-Max — a 2.4-trillion-parameter API-only MoE flagship. Qwen 3.8 27B is the dense model you actually run yourself. As one third-party comparison put it, "Qwen 3.8 is not one model": the Max is for API consumers, the 27B is for people who want weights on hardware they control.

Qwen 3.8 27B at a Glance

AttributeDetail
DeveloperAlibaba (Qwen team)
TypeDense causal LM with vision encoder
Parameters27B (5120 hidden dim, 64 layers)
ModalityText, image, video (native multimodal)
Context window262,144 tokens native, extensible to 1,000,000 via YaRN
ArchitectureHybrid: Gated DeltaNet + Gated Attention (16 of 64 layers full attention)
LicenseApache 2.0
PricingFree to self-host; hosted API pricing not yet confirmed
AvailabilityHugging Face, Ollama, LM Studio, vLLM, SGLang, Unsloth
ReleasedAugust 14, 2026

How Qwen 3.8 27B Works / What Makes It Different

The architecture is where the density story lives. Only 16 of the 64 layers run full attention; the other 48 run linear attention with a constant recurrent state, using a full_attention_interval of 4. The layout alternates blocks of Gated DeltaNet feed-forward stages with periodic Gated Attention layers. This hybrid design is what lets a 27B model hold a 262K context on a single Blackwell GPU.

Three coined feature anchors define its behavior. Flexible Thinking Control exposes a reasoning_effort dial with xhigh (default), medium, and low settings, plus an enable_thinking toggle for instant responses. Preserve Thinking keeps or drops reasoning blocks from historical turns to optimize KV cache use in agent workflows. MTP (Multi-Token Prediction) ships a built-in draft head for faster speculative decoding.

The xhigh default is polarizing. Simon Willison, running the 17GB Q4_K_M build in LM Studio, found it "a chronic over-thinker" — a single "draw an svg of a circle" prompt triggered layered reasoning about "concentric guide circles" and "restrained ambient motion." His pelican-on-a-bicycle SVG "took 21 minutes to generate, using 22,276 reasoning tokens." His verdict on the wait: "Absolutely not." The same prompt with reasoning off finished in about two minutes.

What You Can Do With Qwen 3.8 27B

The model is tuned for long-horizon agentic coding and office work. Community demos landed within hours of release. Unsloth showed the 4-bit GGUF "generating an interactive volcano simulation with geology, thermodynamics, fluid flow, buoyancy/drag" in one shot.

Simon Willison set it up with the Pi harness and had it "build a script for transforming its own .jsonl transcripts to Markdown — which it did." On the vision side, the model card cites 91.1% on OmniDocBench 1.5 for document intelligence and 84.3% on OSWorld-Verified for computer use, per a third-party writeup summarizing the model card.

If you are prototyping agentic coding through a hosted API instead of local weights, a comparable frontier chat model like Claude Opus 4.6 is the reference point Qwen benchmarks itself against.

How Qwen 3.8 27B Compares

Qwen's self-reported card frames the 27B against both its predecessor and closed models. These are vendor and community-relayed figures; independent benchmarks were pending.

BenchmarkQwen 3.6 27BQwen 3.8 27BOpus 4.6 Max
SWE-bench Pro53.561.753.4
DeepSWE 1.113.342.2Not yet confirmed
Terminal Bench 2.163.473.0Not yet confirmed
LiveCodeBench v6Not yet confirmed90.3Not yet confirmed

As one launch-day observer wrote: "27B models should not be this good, what the hell did Qwen do." Treat the Opus comparison as an unconfirmed vendor claim — the methodology, prompts, and hardware behind the numbers are absent from public posts so far. Our earlier look at Kimi K3 covers a very different open-frontier bet at the trillion-parameter end of the spectrum.

Availability: How to Access Qwen 3.8 27B

The weights are live on Hugging Face under Apache 2.0, in FP8, GGUF, and NVFP4 quantizations. Day-zero support shipped across the local-inference stack:

  • Ollama — available for Claude Code, OpenCode, Hermes Agent, and Pi harnesses, with an Apple Silicon MLX variant.
  • LM Studio — recommended the model by default on release day; LinearUncle reported over 10,000 downloads within 38 minutes.
  • vLLM and SGLang — Day-0 integration, with Qwen noting "One GPU, 1M context, Day-0 ready."
  • Unsloth — Dynamic GGUFs running on 17GB RAM.

A hosted version via Qwen Cloud and Model Studio is promised, with 1M context by default and built-in tools, but pricing was not published in the official channels reviewed here.

What We Don't Know Yet

Several claims remain unverified at publication:

  • Independent benchmarks. Every headline score traces to Qwen's own card or community relays. No third-party audit of the Opus 4.6 comparison exists yet.
  • Usable 1M context. Hands-on tests consistently hit 260K–262K; the 1M figure comes from vendor claims and YaRN configs, not demonstrated runs.
  • Multi-turn agent reliability. One early tester reported work truncated on 6 of 15 agent turns, flagged as still needing investigation.
  • Throughput consistency. Reported speeds range from 25.4 tok/s on an RTX 3090 to ~147 tok/s NVFP4 on an RTX 5090, across different quants and runtimes.

Frequently Asked Questions

What is Qwen 3.8 27B?

Qwen 3.8 27B is a 27-billion-parameter open-weight multimodal dense model developed by Alibaba's Qwen team. It released on August 14, 2026 under Apache 2.0 with a 262K native context window and native image and video understanding.

Is Qwen 3.8 27B open source?

Qwen 3.8 27B ships under the Apache 2.0 license with weights published on Hugging Face. It is open-weight, free to self-host, fine-tune, and use commercially.

How much does Qwen 3.8 27B cost?

Qwen 3.8 27B is free to download and run locally under Apache 2.0. Hosted API pricing through Qwen Cloud and Model Studio has not been published in the official channels covered here.

What hardware do you need to run Qwen 3.8 27B?

Qwen 3.8 27B runs on 17GB of RAM or VRAM using Unsloth's Q4_K_M quantized build. It fits on a single 24GB GPU such as an RTX 3090 or 4090, and NVFP4 quants are also available.

Qwen 3.8 27B vs Opus 4.6 — how do they compare?

Qwen's self-reported benchmarks place Qwen 3.8 27B at 61.7 on SWE-bench Pro versus 53.4 for Claude Opus 4.6 Max. These are vendor numbers, and independent benchmarks were still pending at publication.

Is Qwen 3.8 27B multimodal?

Yes, Qwen 3.8 27B is a native vision-language model with an integrated vision encoder. It understands images and video, including hour-scale video, alongside text.

What is the context window of Qwen 3.8 27B?

Qwen 3.8 27B has a 262,144-token native context window. It is extensible to 1,000,000 tokens via YaRN scaling.

What to Watch Next

Three signals will define whether the hype holds. First, watch for independent benchmark runs that reproduce or dispute the Opus 4.6 comparison — Ivan Fioravanti's planned MLX evaluations are one to track. Second, watch whether the 1M-context claim survives real long-document testing beyond the 262K window hands-on users have hit. Third, watch official Qwen Cloud API pricing, which will set the reference floor for hosted access.

Update — 2026-09-01

Since publication, Unsloth reported new Qwen3.8-27B GGUFs with 10% higher accuracy on its cited benchmarks, alongside 1-bit quantizations that it says retain 77% accuracy and run on 8GB of RAM. Alibaba’s Qwen account thanked Unsloth, but these accuracy figures remain vendor-reported.

An experimental EXL3 build was also reported for RTX 3090, 4090, and 5090 systems, with 200,000-plus-token contexts on 24GB of VRAM when paired with DFlash2. Separately, a refusal-removed MLX variant brought the model to Apple Silicon in 2-, 4-, 6-, and 8-bit formats; its creators warned that the modified model can produce malware, fraud, and weapons instructions.

Update — 2026-09-16

The community-quantization story has outpaced the still-absent official launch. Unsloth reports its Qwen3.8-27B GGUF is now the most-liked GGUF of all time on Hugging Face, crossing 10M downloads and 3.7K likes in 24 days — a distribution signal that dwarfs the day-one download counts noted above, even as an official Alibaba release date, pricing, and announcement remain unpublished.

On the hosted side, Alibaba's Qwen account says the model now runs on Cerebras for high-speed inference, though quoted throughput swings widely across hands-on posts — from ~1,850–1,950 tok/s in one live "build an OS on the fly" demo up to a 2,016 tok/s figure that Patrick Moorhead pairs with rate-limit complaints on a paid plan; listed Cerebras pricing runs $0.99/M in and $1.49/M out. Treat these speed numbers as unreconciled — they conflict across accounts.

Local optimization has also matured. Community benchmarks now lean heavily on MTP speculative decoding, which one tester reports requires a recent llama.cpp build with --spec-type draft-mtp and roughly doubles decode throughput (e.g., ~77 → ~180 tok/s on an RTX 5090); a separate RTX 4090 run claims 141 tok/s single-stream on agent workloads with 95.5% GSM8K. A vendor EXL3 one-click installer targets 16–32GB Nvidia cards (RTX 5060 Ti through 5090), and a new lineage detail surfaced: the model is described as SFT + SimPO on traces from Kimi K3, Claude Opus 5, Grok 4.6, and GPT-5.6. Most of these throughput and quant claims remain single-source and independently unverified.

Update — 2026-09-20

The most notable new quantization signal is a report from @NFT_Chen, citing PrismML, of Ternary Bonsai 2 27B: a Qwen-derived 1.76-bit build said to weigh 5.9GB and retain 83.9 versus Qwen 3.8's 85.4 aggregate score, or 98.2%. The same post claims Apache 2.0 licensing and a 262K context window, but these are derivative-release claims rather than an independent rebenchmark of Qwen 3.8 27B.

Early hardware reports put the derivative below the article's usual 24GB-GPU baseline. @sudoingX reports Bonsai 2 running at 40 tokens/s on a single RTX 3060 with the full 262K context resident, alongside a 77-minute agentic build that produced a working file. For the base model, @cyrilXBT reports a quantized run on an 8GB RTX 4060 with a 64K context, 14.6GB on disk, and roughly 5 tokens/s decode after offloading; both are single-source tests and do not establish comparable quality or throughput.

Try related models on Kie.ai: Claude Opus 4.6, GPT-5.6 and GPT-6 Astra.

Lukas Vogel

About Lukas Vogel

Lukas reads the papers and model cards so you do not have to, focusing on reproducible claims.

View all posts by Lukas Vogel