New Open Qwen3.8 Model Challenges Proprietary Flagships

Alibaba's team has released the Qwen3.8 weight series as open source, including the Qwen3.8-27B model and the monstrous MoE variant Qwen3.8-2.4T-A95B. For the first time, Qwen-Max class models are available on Hugging Face and ModelScope for self-hosting and fine-tuning.

If you've been tracking open LLM development over the past year, you've likely noticed an interesting shift. Developers used to compete on raw context window size or MMLU benchmark scores. Now the focus has moved to practical tasks: maintaining reasoning context in long conversations, responding appropriately to environment errors when writing code, and reliable tool use. The Qwen3.8 series is focused precisely on these areas.

What's New in Qwen3.8

In this update, developers ported the architectural improvements from Qwen3.5 to more powerful weights and added two useful controls for reasoning logic.

The first feature is the reasoning_effort parameter. Now you can explicitly regulate the depth of the model's thinking before generating a response. If the task is simple, the model doesn't waste extra tokens on endless reasoning chains. For complex refactoring or mathematical derivations, you can crank the parameter up to maximum.

The second feature solves a common problem with long agent dialogues — preserve_thinking. Usually, when moving to the next step, the context of previous reasoning is lost, causing the model to loop or repeat rejected hypotheses. Here, the thought process context is preserved between messages, noticeably speeding up iterative development.

Architecturally, the family relies on a combination of Gated Delta Networks and sparse Mixture-of-Experts. This reduces memory overhead and provides high generation speed even on contexts up to 262k tokens.

Benchmark Results and Metrics

Let's look at the benchmarks for the Qwen3.6 and Qwen3.5 model series that formed the basis for the current release.

In code generation and external function calling tasks, the 35B-A3B architecture performs on par with heavier monolithic models. At the same time, only about 3 billion parameters remain active per token.

Below are results for the older and younger models in the 3.5 lineup:

How to Run Locally or on a Server

The Qwen team has prepared integrations with almost all popular inference engines. You can run the model on a modest server as well as on a cluster of multiple GPUs.

Quick Start via Hugging Face Transformers

The most straightforward way to spin up an OpenAI-compatible endpoint is the built-in CLI of the transformers library:

transformers serve Qwen/Qwen3.8-27B --port 8000 --continuous-batching

After executing the command, the server listens at http://localhost:8000/v1.

Production Deployment via vLLM and SGLang

For high throughput, it's better to use specialized engines. Note the reasoning and tool-calling parser flags — they're needed for correct agent function operation.

Launching via vLLM:

vllm serve Qwen/Qwen3.8-27B \
  --port 8000 \
  --tensor-parallel-size 4 \
  --max-model-len 262144 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder

Launching via SGLang:

sglang serve \
  --model-path Qwen/Qwen3.8-27B \
  --port 8000 \
  --tp-size 4 \
  --context-length 262144 \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder

If you're working on Apple Silicon, use the mlx-lm or mlx-vlm packages. For running on consumer hardware via CPU and modest GPUs, quantized GGUF builds are already available on Hugging Face under llama.cpp.

Fine-tuning for Your Tasks

If you need to adapt the model for an internal codebase or specific API dialect, the repository recommends using proven libraries:

• Unsloth for fast and memory-efficient LoRA/QLoRA training
• LLaMA-Factory and Swift for SFT, DPO, and GRPO

Flagship MoE variants will require serious server resources, but the 27B version fine-tunes without issues on one or two cards like RTX 4090 or A100 when using quantization.

Who Will Benefit from This Project

The Qwen3.8 series addresses several practical scenarios at once:

• Local code assistants and autocomplete in a closed company environment where you can't send code to external cloud APIs
• Complex autonomous agents that execute multi-step pipelines, read test logs, and fix errors in the repository themselves
• Research pipelines with long context, documentation processing, and finding connections in large volumes of text
• Tuning compact versions for narrow domains without losing overall reasoning coherence

If you're looking for an open foundation for autonomous agents or a terminal code assistant, try Qwen3.8-27B. The model is already supported by most ecosystem tools and deploys in literally a couple of commands.