Qwen3.8-Max is Alibaba's newest flagship AI model — a 2.4 trillion parameter Mixture-of-Experts architecture with 95 billion active parameters per token, a 1 million token context window, and native multimodal input (text, image, and video). Released on August 3, 2026, it is the most capable model in the Qwen family to date and the first Max-class Qwen with open weights on the roadmap. When paired with Hermes Agent — the open-source, self-improving agent framework from Nous Research — it becomes one of the most powerful autonomous coding and task execution setups you can build in 2026.
Why Qwen3.8-Max + Hermes Agent is a combination worth attention
Think of it this way: Hermes Agent is the hands, and the model is the brain. Hermes provides the tool-use loop, persistent memory, skills system, cron scheduling, multi-platform messaging, and goal-mode autonomous execution. The model provides the reasoning. Until now, getting Fable-5-level reasoning inside an agent required an expensive API key. Qwen3.8-Max delivers what Alibaba calls "second only to Fable 5" performance at $2 per million input tokens and $6 per million output tokens — and it plugs into Hermes through standard OpenAI and Anthropic API-compatible endpoints.
The practical result: you can set a goal in Hermes Agent, walk away, and let Qwen3.8-Max code autonomously for hours. The video walkthrough that inspired this guide demonstrated the model building games, apps, animated videos, and MCP-integrated tools — all inside a Hermes Agent session.
What makes Qwen3.8-Max stand out
Architecture and scale
Qwen3.8-Max is a sparse Mixture-of-Experts (MoE) model. Of its 2.4 trillion total parameters, only 95 billion activate per token. That design keeps inference cost tractable while maintaining frontier-level reasoning. The model supports a 1 million token context window — roughly 991,000 tokens of input and up to 131,072 tokens of output per request. A reasoning-effort control offers three levels (low, medium, xhigh) that let developers trade speed against thoroughness.
| Spec | Value |
|---|---|
| Total parameters | 2.4 trillion |
| Active parameters per token | 95 billion |
| Architecture | Sparse Mixture-of-Experts |
| Context window | 1,000,000 tokens |
| Max output | 131,072 tokens |
| Max reasoning budget | ~262,000 tokens |
| Input modalities | Text, image, video |
| API pricing (input/output) | $2.00 / $6.00 per 1M tokens |
| Cached input | $0.25 per 1M tokens |
| Released | August 3, 2026 |
| Open weights | Planned week of August 10, 2026 |
Benchmark performance
Qwen3.8-Max's published benchmarks place it near the top tier, though not at the absolute summit:
| Benchmark | Qwen3.8-Max | Comparison |
|---|---|---|
| Terminal-Bench 2.1 | 86.6% | GPT-5.6 Sol leads at 88.8% |
| PaperBench | 93.0% | Strong agentic performance |
| DeepSWE 1.1 | 56.6% | Claude Fable 5 leads at 73.5% |
| Frontend Code Arena | 1,668 Elo | 37 points behind Claude Opus 5 |
The model does not beat Claude Fable 5 on every metric — Fable 5 leads on SWE-bench Pro (80.0 vs 67.7) and FrontierSWE (88.8 vs 73.5). But Qwen3.8-Max leads on agentic breadth (PaperBench at 93.0) and costs roughly 3-5x less per token than comparable Anthropic models. For agentic workflows that run for hours, that cost difference compounds quickly.
Native API compatibility
One of Qwen3.8-Max's most practical features is its dual API protocol support. It works with both OpenAI's Chat Completions format and Anthropic's API protocol. This means you can drop it into Claude Code, Codex, Qoder CLI, Qwen Code, OpenClaw, and — most importantly for this guide — Hermes Agent, without any harness modifications. It is a base URL and model ID change.
What Hermes Agent brings to the table
Hermes Agent is an open-source, MIT-licensed AI agent framework built by Nous Research. It is not a chatbot — it is a persistent, autonomous agent platform with a closed learning loop. Here are the features that matter most when paired with Qwen3.8-Max:
Skills system. Hermes can learn new procedures and save them as reusable markdown files. In the video walkthrough, the presenter taught the agent a new skill in about two minutes by giving it a guide and typing "learn." The skill was saved locally as a SKILL.md file, ready for reuse in future sessions. This means Qwen3.8-Max can get better at your specific workflows over time.
Persistent memory. Hermes stores facts, preferences, and environment details across sessions. When you switch between models — say from Qwen3.8-Max in Hermes to DeepSeek in another tool — the memory persists. This is what makes the "system-over-model" approach work: the brain changes, but the accumulated knowledge stays.
Goal mode. Hermes can run autonomously for hours toward a defined goal. Qwen3.8-Max was specifically designed for long-horizon agentic tasks — Alibaba demonstrated it running autonomously for 10 days in one development scenario. Together, this means you can set a coding target and walk away while the agent works.
Mixture of Agents (MoA). Hermes 2.0's MoA feature lets you combine multiple models into a single virtual model. You could route a prompt through both Qwen3.8-Max and Claude Fable 5 in parallel, with an aggregator model fusing the best of both responses. This is a way to exceed any single model's quality ceiling.
MCP integration. Hermes connects to Model Context Protocol (MCP) servers, extending its tool surface. The video demo showed Qwen3.8-Max connecting to Hicksfield MCP to generate 3D designs from a book cover image — all from a single prompt.
How to set up Qwen3.8-Max in Hermes Agent
Step 1: Install Hermes Agent
If you have not already installed Hermes, run the install script:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
This works on Linux, macOS, WSL2, and Android (Termux). Windows native users can use the PowerShell installer from the same docs page.
Step 2: Configure the Qwen3.8-Max provider
Hermes supports any OpenAI-compatible endpoint. Qwen3.8-Max is available through:
- Alibaba Cloud Model Studio (QwenCloud) — the official hosted endpoint
- OpenRouter — if you prefer a multi-provider gateway
- A custom OpenAI-compatible endpoint — if Alibaba's own API is your choice
The simplest path is through the Hermes dashboard. Go to the Manage tab, click Update to get the latest version, then create a new profile. In the model selection, choose either the QwenCloud provider or OpenRouter, and select qwen3.8-max as the model.
If you prefer the config file, add it to ~/.hermes/config.yaml:
model:
provider: custom
default: qwen3.8-max
base_url: "https://dashscope.aliyuncs.com/compatible-mode/v1"
context_length: 1000000
You will need your API key from Alibaba Cloud or OpenRouter set as an environment variable. Verify with hermes config check before starting a session.
Step 3: Start a session and test
With the provider configured, start Hermes:
hermes
Type a prompt. If Hermes responds, you are connected. Try a coding task that uses tools — for example, "create a simple HTML game" — to confirm the agent loop is working end to end.
Step 4: Teach it a skill
One of the most powerful workflows is teaching Hermes new skills with Qwen3.8-Max as the brain. Give the agent a guide document (a tutorial, a coding standard, a deployment runbook) and say "learn this." Hermes will create a SKILL.md file from the guide, stored locally. Future sessions can load that skill automatically.
Step 5: Try goal mode
For autonomous, long-running tasks, use goal mode. Give Hermes a goal like "build a landing page for a coffee shop with responsive design" and let it run. Qwen3.8-Max's long-horizon design means it can iterate through multiple files, test its own output, and debug — all without you touching the keyboard.
Step 6: Experiment with Mixture of Agents
If you have access to both Qwen3.8-Max and another frontier model (Claude Opus, GPT-5.6, or DeepSeek V4), configure a MoA preset in Hermes. This runs both models on your prompt and aggregates the best parts into a single response. The result is often better than either model alone.
moa:
presets:
coding:
reference_models:
- provider: openrouter
model: qwen/qwen3.8-max
- provider: openrouter
model: anthropic/claude-opus-4.8
aggregator:
provider: openrouter
model: anthropic/claude-opus-4.8
reference_max_tokens: 600
What people are building with this combination
The video walkthrough that inspired this guide demonstrated several real projects:
- Coding apps and games. Qwen3.8-Max built interactive web applications inside Hermes with no handholding. The presenter tested 50 different builds, including games, tools, and websites.
- MCP-powered 3D design. Connected to Hicksfield MCP, the model took a book cover image as input and generated 3D designs.
- Animated videos with Remotion. Using a Remotion skill, the agent created fully animated promotional videos from a single prompt — complete with research, scripting, and editing.
- Autonomous coding sessions. With goal mode active, the model ran autonomously, building and iterating without intervention.
For a deeper look at the system-over-model architecture that makes these workflows possible — hot-swapping Qwen 3.8 Max, Kimi K3, and GPT-5.6 without rebuilding your setup — see our guide to building a multi-model AI coding workstation.
Qwen3.8-Max vs Claude Fable 5: the honest comparison
Alibaba's "second only to Fable 5" claim is directionally correct but needs nuance. Here is what the data actually says:
Where Qwen3.8-Max wins: Agentic breadth (PaperBench 93.0), multimodal input at trillion-parameter scale, cost ($2/$6 vs Anthropic's $15/$75 for Opus-class models), and the fact that open weights are coming. It also supports both OpenAI and Anthropic API protocols natively, which Fable 5 does not.
Where Claude Fable 5 wins: Pure software engineering (SWE-bench Pro 80.0 vs 67.7, FrontierSWE 88.8 vs 73.5), reasoning (HLE-Full 53.3 vs undisclosed), and overall coding Elo on Code Arena (~60 vs ~47-48). Fable 5 remains the undisputed coding leader.
The practical takeaway: If your primary selection criterion is coding quality on hard software engineering tasks, Fable 5 is still the better model. If you need agentic autonomy at scale, multimodal input, 1M token context, and a fraction of the cost — Qwen3.8-Max is the stronger choice. And when paired with Hermes Agent's Mixture of Agents, you can combine both models and get the best of each.
For more on how these models compare in the broader landscape of AI trends, see our coverage of OpenAI Astra and Claude Fable solving decades-old math problems and our analysis of sovereign AI and enterprise data control.
The open-weight question
Qwen3.8-Max is currently API-only (as of August 4, 2026). Alibaba has announced that open weights for both Qwen3.8-Max and a companion Qwen3.8-27B checkpoint will land on Hugging Face and ModelScope the week of August 10, 2026. A 27B dense companion is the checkpoint intended for on-premise GPU hardware — the 2.4T flagship is a multi-node datacenter artifact that requires roughly 2 terabytes of storage for weights alone at 4-bit precision.
The weight license has not been published at launch. Previous Qwen open-weight releases used Apache 2.0, but teams planning on-premise deployment should wait for the official license before committing.
Pitfalls to watch for
Cost at scale. Qwen3.8-Max is cheap for a frontier model, but multi-hour autonomous runs in goal mode consume tokens continuously. Set a spending cap on your provider account before launching an unattended session.
Context length and local models. If you plan to run a smaller Qwen model locally through Ollama, be aware that Ollama defaults to a 4K token context window. Hermes requires a minimum 64K context. You will need to create a custom Modelfile with PARAMETER num_ctx 65536 and rebuild the model variant. This is a well-documented gotcha that costs hours if you do not know about it.
Active vs total parameters. The 2.4 trillion figure is the total parameter count. The 95 billion active per token is what determines single-query reasoning power. Do not compare the 2.4T headline directly to dense models — it is not an apples-to-apples comparison.
Benchmark confidence. The only independent coding data for Qwen3.8-Max comes from Code Arena (human preference, not automated correctness). No independent reasoning or agentic benchmarks have been published yet. Treat all comparisons as provisional until third-party evaluators like Artificial Analysis publish scores.
FAQ
What is Qwen3.8-Max?
Qwen3.8-Max is Alibaba's flagship AI model released on August 3, 2026. It is a 2.4 trillion parameter Mixture-of-Experts model with 95 billion active parameters per token, a 1 million token context window, and native text, image, and video input. It is the most capable model in the Qwen family to date.
How does Qwen3.8-Max compare to Claude Fable 5?
On available data, Claude Fable 5 leads on pure software engineering benchmarks (SWE-bench Pro 80.0 vs 67.7, FrontierSWE 88.8 vs 73.5) and reasoning (HLE-Full 53.3 vs undisclosed). Qwen3.8-Max leads on agentic breadth (PaperBench 93.0) and costs roughly 3-5x less per token. Fable 5 remains the stronger coding model; Qwen3.8-Max is the more cost-effective agentic model.
Can I run Qwen3.8-Max locally?
Not the 2.4T flagship — it requires multi-node datacenter hardware. The companion Qwen3.8-27B checkpoint, with open weights planned for the week of August 10, 2026, is the model intended for on-premise GPU hardware. For local use through Hermes Agent today, use the QwenCloud or OpenRouter API endpoint.
How do I plug Qwen3.8-Max into Hermes Agent?
Open the Hermes dashboard, go to the Manage tab, click Update, create a new profile, and select Qwen3.8-Max as the model from either the QwenCloud or OpenRouter provider. Alternatively, configure it manually in ~/.hermes/config.yaml with the API base URL and your API key. Run hermes config check to verify.
What is Hermes Agent's Mixture of Agents feature?
Mixture of Agents (MoA) is a virtual model provider in Hermes that fans a single prompt out to multiple models in parallel, then uses an aggregator model to combine the best parts into one response. You can mix Qwen3.8-Max with Claude Opus or GPT-5.6 in a preset and select it as your model. The aggregated output is often better than any single model alone.
Is Hermes Agent free to use?
Yes. Hermes Agent is open-source under the MIT license. You can use it commercially, modify it, and deploy it anywhere. The cost is whatever your model provider charges per token. With Qwen3.8-Max at $2/$6 per 1M tokens, a full day of autonomous coding can cost under $20.
What is goal mode in Hermes Agent?
Goal mode lets you give Hermes a target and let it run autonomously until the goal is met. The agent works through multiple steps, calls tools, tests its own output, and iterates — all without intervention. Qwen3.8-Max was designed for long-horizon agentic tasks, making it an ideal model for goal mode.
Discussion
0 comments