Qwen 3.8
TL;DR Highlight
Qwen3.8-Max-Preview is live on Token Plan, with a 2.4T open-weight release promised soon
Who Should Read
A backend developer weighing cost and performance to decide whether to switch LLM API providers
Core Mechanics
- Qwen3.8 is going open-weight soon
- The post says it has 2.4T parameters.
- It only says 'soon', with no release date.
- Max-Preview ships first
- You can try it on Alibaba's Token Plan, Qoder, and QoderWork.
- There are separate pricing pages for international and China.
- Alibaba claims second only to Fable 5
- It says the model is comparable to leading frontier AI models.
- The post uses the phrase 'second only to Fable 5'.
- No benchmark numbers in the post
- It gives no per-model scores, pricing, or context length.
- The performance claim is the company's own.
Community Reactions
- Mixed experience with earlier Qwen models
- One commenter's tests found DeepSeek V4 Pro better, faster, and much cheaper than Qwen 3.7 Max.
- They also rated Qwen the most censored of the Chinese models.
- Qwen 3.7 Max was too wordy, per one user
- One commenter said it wrote multiple paragraphs for a single prompt.
- They said they do not trust Qwen benchmark scores.
- Opinion that big models burn many tokens
- A commenter said GLM 5.2 and Kimi K3 are token hungry and feel slow.
- Another said one feature on the $19 Kimi K3 plan used nearly their whole five-hour limit.
- Local-model users want smaller sizes
- One commenter runs 35B MoE and 27B dense locally and rarely needs Claude, especially for sensitive data.
- Another said mtplx ran Qwen 3.6 27B 2-3x faster.
- A commenter shared an OpenCode config
How to Apply
- Run your existing prompts on Qwen3.8-Max-Preview
- Take 5-10 representative prompts and run them as-is on Qwen chat or Token Plan.
- Log answer length, accuracy, and token usage side by side with your current model.
- Connect it in OpenCode with the shared JSON
- Paste the provider block from the comments into your OpenCode config.
- The limit values are unofficial, so correct them against official docs.
- Add a length limit to your system prompt
- Because of the verbosity complaints, state a concise-answer rule and a max paragraph count.
- Compare output token counts before and after on the same inputs.
- Build a swap-evaluation set before weights drop
- Pull cases from real request logs and save them as an eval set.
- When weights release, compare cost, latency, and accuracy against your current model.
Try It
# Paste into Claude or ChatGPT: plan a comparison eval before adopting Qwen3.8 Max Preview (a simple comparison prompt, not the post's method)
I am a developer building a service on LLM APIs. My current model is {{current model}}, and I want to test Qwen3.8-Max-Preview.
My use case: {{service description}}
Create a comparison evaluation plan covering:
1. Eight test case types to pull from real request logs
2. Grading criteria for each case
3. Metrics: accuracy, output token count, latency, verbosity
4. A markdown table template for recording results
Next step: fill in the table with real results and ask for a recommendation on whether to switch models.Terminology
Related Papers
The Judge Can Debias Itself: Awareness and Self-Debiasing for Reliable LLM-as-a-Judge
Warning an LLM judge about bias or letting it self-critique improved 20 of 21 model–bias combinations.
90% of US working hours are spent on tasks not covered well by AI benchmarks
Mapping 29 benchmarks to job tasks shows 90% of US working hours lack good coverage
Postgres 19 delayed after AI reviewers find more bugs than humans can fix
LLM code review surfaced so many bugs that Postgres 19 slipped to November and 8-10 features were reverted
Show HN: Gemini 4 Argon plays StarCraft
LLMs wrote StarCraft bots in one hour; GPT-6 Astra scored highest.
A Plausible Explanation Is Enough to Mislead Multimodal Judges
Adding plausible explanations to fixed answers cut MLLM judges' wrong-answer rejection by 12.7–13.8 points.
Neural networks are clean
Looking at neurons directly shows most have one main job, challenging superposition.