On our own GPUs
Qwen3.6-35B-A3B API
Apache 2.0, 3B active per token, tuned specifically for agentic coding.
Context
128K
Max output
32K
Serving precision
BF16
Licence
Apache-2.0
Qwen3.6-35B-A3B is a mixture-of-experts model from Alibaba’s Qwen team, shipped in April 2026 with 35B total parameters and roughly 3B active per token under an Apache 2.0 licence. It is tuned explicitly for agentic coding and works with the open-source terminal agent OpenClaw. Sparse activation keeps unit cost very low, which makes it the right model to absorb the many routine steps in a coding agent while larger models handle the pivotal ones.
Open weights, deployed in facilities we build out and lease directly, operated by us. We resell nobody else’s account quota.
Specifications
- Model ID
- qwen3.6-35b-a3b
- Published by
- Alibaba Qwen
- How it is served
- On our own GPUs
- Licence
- Apache-2.0
- Total parameters
- 35B MoE
- Active parameters
- 3B
- Context
- 128K
- Max output
- 32K
- Serving precision
- BF16
- Input modalities
- Text
- Released
- 2026-04-01
- Price
- See pricing
What it is good at
- A 3B-active cost structure
- Inference touches roughly 3B parameters, so unit cost sits far below same-generation flagships. The read-a-file, list-a-directory, run-a-command steps that dominate agent traces are the cheapest possible work to hand it.
- Apache 2.0, the most permissive tier
- No restrictions on commercial use, modification or redistribution, and no user-count threshold. For teams that need private deployment or their own fine-tune, this is the lowest legal overhead available.
- Tuned for terminal agents
- Explicitly tuned for agentic coding and known to work with open-source terminal agents like OpenClaw. Tool-call formatting stays stable, which is the part of these pipelines most likely to break.
Where it fits
- High-frequency routine steps inside coding agents
- Cost-sensitive bulk code review and patch generation
- Local or private-environment development assistants
- The cheap first hop in a multi-model routing setup
Three steps
Change one line of base_url, put the model ID below in the `model` field, and leave the rest of your code alone.
model: "qwen3.6-35b-a3b"
curl https://api.nexaix.net/v1/chat/completions \
-H "Authorization: Bearer $NEXAIX_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen3.6-35b-a3b",
"stream": true,
"messages": [
{ "role": "user", "content": "Summarise this contract." }
]
}'import os
from openai import OpenAI
client = OpenAI(
# the only new line
base_url="https://api.nexaix.net/v1",
api_key=os.environ["NEXAIX_API_KEY"],
)
# everything below is your existing code
stream = client.chat.completions.create(
model="qwen3.6-35b-a3b",
stream=True,
messages=[
{"role": "user", "content": "Summarise this contract."}
],
)
for chunk in stream:
print(chunk.choices[0].delta.content or "", end="")import OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://api.nexaix.net/v1", // the only new line
apiKey: process.env.NEXAIX_API_KEY,
});
// everything below is your existing code
const stream = await client.chat.completions.create({
model: "qwen3.6-35b-a3b",
stream: true,
messages: [
{ role: "user", content: "Summarise this contract." }
],
});
for await (const chunk of stream) {
process.stdout.write(chunk.choices[0]?.delta?.content ?? "");
}About this model
Is 35B big enough?
It depends on how you tier your work. As a standalone general model it is clearly behind the trillion-parameter flagships, but in an agent loop most steps are simple and deterministic — routing them here can drop total cost by an order of magnitude. Escalate the pivotal reasoning steps to DeepSeek V4 Pro or Kimi K3.
What is the context window?
128K, extended via YaRN. If you need a million-token window, use GLM-5.2, DeepSeek V4 or Kimi K3.
Does it support tool calling?
Yes, using the standard OpenAI tools / tool_calls format — no change to your existing code.
Also worth a look
- On our own GPUsKimi K32.8T mixture-of-experts, a 1M-token window, and vision built into the base.kimi-k3
- On our own GPUsGLM-5.2750B sparse MoE, a 1M window, MIT licensed — tuned hard for software engineering.glm-5.2
- On our own GPUsDeepSeek V4 Pro1.6 trillion parameters, thinking and non-thinking modes, MIT licensed.deepseek-v4-pro
Run your own eval on it
Take the trial credit, run the evals you already have against this model, and then decide whether it gets your production traffic.
