Qwen3.6 35B A3B

A mixture-of-experts model from Qwen, with 35B parameters and 3B active per token.

Processed only in the UK
Zero data retention

Create Account Your Account

Pricing

Input: £0.08 / 1M tokens
Cached input: £0.02 / 1M tokens
Output: £0.75 / 1M tokens

Pricing Details

Specification

Model ID: qwen/qwen3.6-35b-a3b
Weights: Qwen/Qwen3.6-35B-A3B-FP8, Apache 2.0 licence
Context window: 262,144 tokens
Input: Text and images (JPEG, PNG or WebP)
Output: Text
Reasoning: On by default
Features: Tool calls, JSON output, logprobs, streaming
Prompt caching: Automatic per API key, in blocks of 2,096 tokens
Precision: FP8 weights and KV cache
Hardware: Mallard AI's own GPU server
Withdrawal notice: 30 days by email

Performance

Time to first token: ~0.83 s
Throughput: ~79 tokens/s

	Text	Images
Test prompt, mean	16,390 tokens, 75% cached	1,827 tokens, 1.14 images
Output	356 tokens	250 tokens
Time to first token, median	0.83 s	0.96 s
Time to first token, 90th percentile	1.39 s	1.31 s
Throughput, median	79 tokens/s	84 tokens/s

Measured through the public API on 24 October 2026.

Time to first token by prompt size

Prompt tokens	New prompt	Sent again
1,024	0.22 s	0.22 s
2,048	0.37 s	0.37 s
4,096	0.72 s	0.39 s
8,192	1.34 s	0.41 s
16,384	2.74 s	0.44 s
32,768	5.82 s	0.49 s
65,536	13.66 s	0.46 s
131,072	36.93 s	1.10 s
262,143	106.89 s	1.14 s

Medians of three runs, measured on 24 October 2026.

Example request

curl

curl https://api.mallardai.co.uk/v1/chat/completions \
  -H "Authorization: Bearer $MALLARD_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen/qwen3.6-35b-a3b",
    "messages": [{"role": "user", "content": "What is the capital of Wales?"}]
  }'

Quick Start

Request parameters

Field	Accepted
`temperature`	0 to 2
`top_p`	0 to 1
`top_k`	0 to 1000
`frequency_penalty`, `presence_penalty`	−2 to 2
`repetition_penalty`	0.01 to 2
`seed`	Any integer
`max_tokens`	Up to the context window less the prompt
`stop`	Up to 4 sequences
`tools`, `tool_choice`	OpenAI's format, with `tool_choice` auto, none, required or a named function
`response_format`	`json_schema` (strict) or `json_object`
`logprobs`, `top_logprobs`	`top_logprobs` up to 20
`reasoning`, `reasoning_effort`	Switch reasoning off with `{"enabled": false}` or `"none"`
`user`, `safety_identifier`, `session_id`, `prompt_cache_key`, `x-session-id` header	Keep each of your users' cache apart
`n`, `best_of`	1

Limits

Concurrent requests: 4 per API key (1 until your first top-up), more on request
Images: 4 per request
Linked images: 10 MB each
Request timeout: 60 minutes