Qwen3.6 35B A3B | EmpirioLabs AI Docs

POST /v1/chat/completions

Qwen3.6 35B A3B is a 256-expert mixture-of-experts reasoning model with 128K context, function tools, and strict structured JSON output.

At a glance

Field	Value
Model id	`qwen3-6-35b-a3b`
Model release date	2026-04-16
Input modalities	Text
Output modalities	Text
Context window	128K
Weight precision	eschamoe W2: 2-bit weights + INT8 dense + FP16 KV
Max output tokens	16,384
Features	reasoning, function_calling, json_mode, cache
Native inference	Yes
New	Yes
Structured output	JSON Schema
Supported endpoints	`POST /v1/chat/completions`, `POST /v1/responses`, `POST /v1/messages`, `POST /v1/completions`, `POST /v1beta/models/qwen3-6-35b-a3b:generateContent`
Alternate model ids	`qwen3.6-35b-a3b`, `qwen/qwen3.6-35b-a3b`, `EschaLabs/Qwen3.6-35B-A3B-Escha-W2`

Pricing

Charge	Spec	Rate
Input	per 1M prompt tokens	$0.07 (was $0.248)
Output	per 1M generated tokens	$0.42 (was $1.485)
Implicit cache read	per 1M cached input tokens	$0.035
Web Search (Linkup)	per call when invoked	$0.013

Example request

$ curl https://api.empiriolabs.ai/v1/chat/completions \
>   -H 'Authorization: Bearer $EMPIRIOLABS_API_KEY' \
>   -H 'Content-Type: application/json' \
>   -d '{"model": "qwen3-6-35b-a3b", "messages": [{"role":"user","content":"Hello"}]}'

Parameters

Parameter	Type	Required	Default	Description
`temperature`	number	no	`0.7`	Sampling temperature. 0 is deterministic and 2 is maximum randomness. · Range: 0 – 2
`top_p`	number	no	`0.95`	Nucleus sampling probability mass. Lower values make outputs more focused. · Range: 0 – 1
`max_tokens`	number	no	`4096`	Maximum output tokens. · Range: 1 – 16384
`stop`	string	no	-	Up to 4 strings where the model will stop generating further tokens.
`enable_thinking`	boolean	no	true	Enable reasoning before answering.
`reasoning_effort`	enum	no	`"medium"`	Reasoning effort level. none disables thinking. low, medium, high, and max set bounded thinking budgets sized to the selected model. · Allowed: `none`, `low`, `medium`, `high`, `max`
`top_k`	number	no	`20`	Limit sampling to the top K candidate tokens when supported. · Range: 1 – 200
`min_p`	number	no	`0`	Minimum probability threshold for token sampling. · Range: 0 – 1
`frequency_penalty`	number	no	`0`	Penalty based on how often a token has already appeared. · Range: -2 – 2
`presence_penalty`	number	no	`0`	Penalty for tokens that already appeared in the generated text. · Range: -2 – 2
`seed`	number	no	-	Optional random seed for reproducible sampling. · Range: 0 – 2147483647
`response_format`	enum	no	-	Constrain the output to valid JSON. Use JSON mode for any JSON object, or supply a JSON Schema to force an exact response shape. · Allowed: `text`, `json_object`, `json_schema`
`web_search_linkup`	boolean	no	false	Optional web search powered by Linkup. When enabled, recent web sources are retrieved using your latest user message as the query and provided to the model as additional context. Adds $0.013 per call when invoked on top of the model’s normal token cost. Disabled by default.
`disable_formatting`	boolean	no	false	When enabled, the gateway will not append the “Sources” footer to assistant responses that used Linkup web search. Useful when the model output is piped to another system that expects no decoration.

Notes

Text-only. This build does not accept image or video input, unlike the base Qwen3.6 35B A3B.

Weights

Served from the 2-bit eschamoe W2 build published by Escha Labs (eschalabs.com) as EschaLabs/Qwen3.6-35B-A3B-Escha-W2 on Hugging Face, under Apache-2.0. The experts are quantized to 2 bits, mixed per projection (gate_up_proj at 2-bit and down_proj at 3-bit), the dense layers are int8, and the KV cache is FP16. Escha Labs publishes the quality comparison against an FP8 baseline of the same model: parity or better on math, graduate science, tool use and long context, about 2 percent lower on broad knowledge, and about 7 percent lower on long-horizon code generation, which is the one clear gap. See the model card for the full benchmark table and protocol.

Behavior

Supports streaming, function tools, structured JSON output including strict schemas, and thinking mode on by default. Set enable_thinking=false for direct answers. With thinking on, the reasoning arrives in reasoning_content and the answer in content, so read both. A low max_tokens with thinking on can be spent entirely on reasoning, so allow room for the answer.

Caching

Automatic prefix cache reads are billed at the cached-input rate when reported. Explicit cache controls are not supported. Cancelling a streaming request mid-generation bills only the tokens produced up to that point.

Machine-readable schema: GET https://api.empiriolabs.ai/v1/models/qwen3-6-35b-a3b.