Qwen3-Max brings 2.4T MoE to open weights, but the 27B is where the action is
Alibaba dropped two cards in a single breath.
One is Qwen3-Max, a 2.4T-parameter MoE flagship positioned against the closed-source ceiling of GPT-5, Gemini 2.5 Pro, and Claude 4.5 Opus. The other is Qwen3-27B, a dense model that fits on a single consumer GPU for inference and a pair of them for fine-tuning.
Both target the same two workflows: Coding (code generation, agentic programming) and Cowork (multi-agent collaboration).
With the cards on the table, the question is which one developers actually take home.
What 2.4T actually means
First, the biggest misconception: a 2.4T MoE is not 2.4T active at runtime.
The total parameter count is 2.4T, but each inference pass activates only a small slice of experts. Qwen's typical activation ratio sits under 10%. In other words, that "2.4T" number reads more like a catalog of capabilities. The model knows a lot, but answering any single question pulls from just a fraction of it.
Three things the number actually signals:
- Higher knowledge capacity ceiling. More training-time patterns compressed into the weights than dense models can hold. Tasks that reward breadth (RAG, complex instructions, long-chain reasoning) get a natural edge.
- A statement of intent against closed-source flagships. Dropping 2.4T into open weights is something few Chinese labs have attempted. Llama 4 Maverick sits at 400B activated parameters, and Llama 4 Behemoth is still in preview. Qwen just raised the open-source ceiling by a full tier.
- The deployment barrier hasn't moved. Running a full 2.4T MoE still demands racks of H100s or H200s, with VRAM measured in terabytes. Individual developers won't touch it. The ones who do will be enterprise inference providers.
For most developers, the Max release is a signal, not a tool.
The 27B is the real open-source bomb
The 27B dense model is the one developers can pick up and use today.
A few key points:
- Single RTX 4090 / 5090 handles inference; INT4 quantization fits inside 24GB of VRAM
- Dual A100 / H100 handles fine-tuning; LoRA and QLoRA both work cleanly
- Dense architecture; no expert-routing quirks, with vLLM, SGLang, and TensorRT-LLM already supporting it out of the box
What stands out more than the specs is the targeting: Coding and Cowork, both open-sourced together.
Previous open-source models tended to be "decent at general tasks, weak at agentic ones." They could write a function or two, then fall apart when plugged into tool-calling chains like Claude Code, Cline, or Cursor. Qwen shipping Coding (code generation, tool use, planning) and Cowork (multi-agent collaboration, role-playing, task decomposition) in one release is a direct message to the community: we're claiming the open-source foundation for both of these workflows.
The double-down on Coding and Cowork
Why these two? Because the money and the talent are there.
Coding agents became the most crowded track from late 2024 through 2025. Anthropic turned Claude Code into a product-grade experience. Cursor's valuation crossed into the tens of billions. Devin, Cline, and Codex CLI followed one after another. The underlying model's capability directly determines how far any of these products can go.
Cowork (collaborative agents) is the main battleground for agent frameworks. LangGraph, CrewAI, and AutoGen abstract "how do multiple agents divide work, communicate, and share state" into graphs and state machines. The base model's role-playing ability, instruction-following, and long-context stability decide how complex a task these frameworks can sustain.
Qwen open-sourcing both capabilities at once effectively cements the open-source foundation for both tracks. For Chinese teams, replicating something at the level of Cursor or CrewAI no longer requires rebuilding the model layer from scratch.
Open weights at the 2T tier: how fast Chinese teams are moving
A look at the competitive landscape:
- Meta / Llama: Llama 3 topped out at 405B; Llama 4 Maverick at 400B activated, with Behemoth still in preview
- Mistral: stuck to the small-and-mid model lane; the Mixtral MoE series never broke 200B
- DeepSeek: V3 at 671B MoE was previously the closest open-weight peer
Qwen's 2.4T MoE leapfrogs all of them. It's a double bet on engineering capability and compute reserves.
Worth noting:
- Alibaba has been steadily shipping the Qwen2.5 series through 2024 without breaking the open-source cadence
- Releasing Max and the 27B together signals a product-matrix approach: flagship plus deployable
- Coming back after a multi-month gap and immediately matching closed-source flagships is a tone-setter for the domestic large-model race
For Chinese developers, the most direct impact: the model cost of reproducing frontier agent systems just collapsed. The 27B runs, fine-tunes, and drops into any agent framework. The next wave of open-source coding tools will almost certainly be built on top of Qwen 27B.
What you can do with it right now
For anyone ready to get hands-on today, a few paths that work immediately:
1. Pull the weights and run 27B locally
# Hugging Face Qwen3-27B
pip install transformers accelerate vllm
# Serve with vLLM
vllm serve Qwen/Qwen3-27B-Instruct \
--tensor-parallel-size 2 \
--max-model-len 32768 \
--gpu-memory-utilization 0.9
VRAM requirements (FP16):
- Single-card inference: ≥ 60GB (A100 80G, H100 80G)
- INT4 quantized inference: 24GB (4090 / 5090 works)
- LoRA fine-tuning: dual 80G minimum
- QLoRA fine-tuning: single 24G works
2. Plug into a coding agent framework
Qwen's tool calling has consistently been among the most reliable in the open-source world. Drop it into a Claude Code-style CLI or wire it up to Cline or Roo Code with minimal code changes.
from openai import OpenAI
client = OpenAI(
base_url="http://localhost:8000/v1",
api_key="EMPTY"
)
response = client.chat.completions.create(
model="Qwen/Qwen3-27B-Instruct",
messages=[
{"role": "system", "content": "You are a Python engineer."},
{"role": "user", "content": "Write a FastAPI endpoint that reads a CSV and returns JSON."}
],
tools=[
{
"type": "function",
"function": {
"name": "run_shell",
"description": "Execute a shell command",
"parameters": {
"type": "object",
"properties": {"cmd": {"type": "string"}},
"required": ["cmd"]
}
}
}
]
)
3. Wire it into LangGraph / CrewAI
Point the base URL at the local vLLM server and you're done. LangGraph's ChatOpenAI compatibility layer and CrewAI's OpenAI-compatible LLM both accept the same configuration.
A few landmines to watch for:
- Licensing: Most Qwen releases ship under Apache 2.0 or Qwen's own license. Read the fine print before any commercial use.
- VRAM: 27B at full precision eats around 54GB. Quantization is effectively required. Max is essentially off-limits for local deployment.
- Tool calling: Different Qwen sub-versions handle tool calling differently. Run the official eval before committing to a deployment.
The bottom line
Qwen's release isn't just "another large model."
It fires three signals at once: open-source ceiling matched against closed-source flagships, deployable capability pushed down to consumer GPUs, and Coding and Cowork both getting open-source foundations on the same day.
For Chinese teams building coding agents, agent frameworks, and AI IDEs, this is a window where the model layer no longer holds them back. What remains is engineering, product, and distribution, and those are the harder problems.
Source: Latent Space, Qwen 3 Max: 2.4T MoE + 27B Dense Open Sourced, https://www.latent.space/p/ainews-qwen-38-max24t-and-27b-new