Models / Qwen/ Qwen3.6 35B A3B

Qwen3.6 35B A3B

Name: Qwen3.6 35B A3B
Author: Qwen

Qwen · released Apr 15, 2026 · Qwen/Qwen3.6-35B-A3B

InputOutput

Type: Open weightsApache License 2.0
Params: 36B
Context: 262K

3B active per word · about 197K words of context

Our take

Written Aug 3, 2026

Qwen 3.6 is a downloadable mixture-of-experts model with only 3 billion active parameters for each token it processes, drawn from 36 billion total. It handles text, images and video across a 262,144-token working window, though no independent quality scores are available yet.

Who should pick it

Choose this for edge or local deployment where low active-parameter overhead matters, or for long-context tasks that need a quarter-million tokens of working memory. It suits budget-conscious hosted use with several providers to pick from. Skip it if you need verified quality benchmarks to compare against peers, or if you want predictable output pricing and throughput across hosts.

The case for it

Only 3 billion parameters are active per token from 36 billion total — a 12:1 ratio that keeps inference lean.
Apache 2.0 licence allows commercial use, fine-tuning and redistribution without restriction.
262,144-token working window is large for its active-parameter class.
Ten hosted offers with strong peak throughput: CoreWeave reaches 166 tokens per second, Venice 138, Atlas Cloud 120.

The case against it

No benchmark scores in our data — no Elo, MMLU or other verified quality measures to judge it by.
Output pricing varies sharply across providers, with a 1.56× spread between cheapest and most expensive for the identical model.
Throughput is highly inconsistent: 27 tokens per second at one provider versus 166 at another, and three providers do not disclose speed at all.

How good is it?

Nothing measured, on any board we watch.

We hold no score for this model.

We look for every model we track on every board we watch, and none of them has turned up Qwen3.6 35B A3B — so there is no intelligence, coding, agentic or writing score to show you, not a low one, none.

The boards we watch →

Can you run it yourself?

Fits fully in memory on 24 of the 71 devices we track, at Q4.

Fits in memory: weights load entirely on the card
Spills to system RAM: some weights offload; much slower
Too large: will not load even with offload
est: size is calculated; the verdict could change by 10%

A card many people ownSpills to system RAMest

GeForce RTX 4090 · 24 GB

Weights at Q4_K_M22.7 / 24 GBest

Usable contextNot calculated for spilled setups

Decode speedNot estimated for spilled setups

Loads, but do not expect an assistant. Some of the weights sit in ordinary system memory, which is far slower than the card.

Comfortable fit

One step upFits in memory

GeForce RTX 5090 · 32 GB

Weights at Q4_K_M22.7 / 32 GBest

Spare memory6.1 GB spare

Usable context66K of 262K

Decode speed670 tok/sest

Room to spare. 6.1 GB spare means a 10% error in the size would not change the answer.

On a MacFits in memoryest

Apple M3 Pro (18-core GPU) · 36 GB

Weights at Q4_K_M22.7 / 36 GBest

Spare memory2.3 GB spare

Usable context16K of 262K

Decode speed49 tok/sest

Borderline fit on an estimated size. It leaves 2.3 GB spare on a size we calculated rather than measured, and a 10% error either way would change the answer.

Your hardware

Checking your profile…

Memory use by level

Against a 24 GB card.

Q4_K_M

recommended

22.7 GBest

Spills to system RAMest

Q5_K_M

26.6 GBest

Spills to system RAM

Q8_0

39.8 GBest

Too large

All 3 sizes here are calculated, not measured. We hold no measured file for this model, so each size comes from the parameter count and every verdict above inherits that uncertainty.

Check against your own machine → · All 71 devices, with every size →

Or rent it from someone else

11 providers, from 12 listings.

Cheapest published offer

OpenRouter

Cheapest of 12 live listings. Picked at the widest standard context we hold, within one quantisation slice, so the numbers beside it are a price one host actually charges.

per 1M tokens: $0.14 in / $1.00 out
Context served: 262K
Throughput: Not measured

Current provider offers with price, context and prompt-privacy answers
Provider	In / out per 1M tokens	Context	Throughput	Trains on prompts	Logs prompts	Zero retention
Venice AIfp8	$0.098 / $0.95	256K	223 tok/s	No	No	Confirmed
DeepInfrafp8	$0.10 / $0.95	262K	not measured	Unknown	Unknown	Unknown
DeepInfrafp8	$0.10 / $0.95	262K	28 tok/s	No	No	Confirmed
OpenRouter	$0.14 / $1.00	262K	not measured	Unknown	Unknown	Unknown
Parasailfp8	$0.15 / $1.00	262K	43 tok/s	No	No	Confirmed
AkashMLfp8	$0.14 / $1.00	262K	28 tok/s	No	No	Confirmed
AtlasCloudfp8	$0.19 / $1.11	262K	98 tok/s	No	Yesunknown period	Unknown
CoreWeavefp8	$0.25 / $1.25	262K	190 tok/s	No	No	Confirmed
Phala	$0.20 / $1.27	262K	29 tok/s	No	No	Confirmed
Novita AI	$0.25 / $1.49	262K	not measured	Unknown	Unknown	Unknown
SiliconFlowfp8	$0.20 / $1.60	262K	42 tok/s	No	No	Confirmed
Io Netfp8	$0.29 / $1.89	262K	153 tok/s	No	No	Confirmed

Across the 12 listings we hold: 9 say they do not train on prompts, 0 say they do and 3 do not say. 8 appear in the zero-retention registry we check; the rest are unknown to us rather than confirmed either way.

What each host's API supports

From the parameter list each endpoint publishes. Streaming is omitted: nothing we hold reports it, for any model.

✓
Supported: ✗
Not supported
Not published: host gave no parameter list

API features per host
Provider	Tool calling	JSON output	Strict schema
Venice AIfp8	✓	✗	✗
DeepInfrafp8
DeepInfrafp8	✓	✓	✓
OpenRouter	✓	✓	✓
Parasailfp8	✓	✓	✓
AkashMLfp8	✓	✓	✓
AtlasCloudfp8	✓	✓	✗
CoreWeavefp8	✗	✓	✓
Phala	✓	✗	✓
Novita AI
SiliconFlowfp8	✗	✓	✓
Io Netfp8	✓	✗	✗

Tool calling: 8 of 12 listings say yes, 2 say no, 2 publish no parameter list. JSON output: 7 of 12 listings say yes, 3 say no, 2 publish no parameter list. Strict schema: 7 of 12 listings say yes, 3 say no, 2 publish no parameter list.

Models people weigh against Qwen3.6 35B A3B

Shortlist any of these to compare them side by side.

Weighed against itGemma 4 26B A4B Gemma 4 31B Gemini 2.5 Flash

Other sizes and versionsQwen3.5-35B-A3B

SucceedsQwen3.5-35B-A3B

Compare them field by field →

When we formed this view

And whether anything has moved since.

Dates behind this page

Aug 3, 2026Price changeQwen3.6 35B A3B cut across 2 hosts, by up to 5% at VeniceQwen3.6 35B A3B moved on 2 hosts: Venice: input −2% ($0.10 → $0.098 per 1M tokens); output −5% ($1.00 → $0.95 per 1M tokens); Io Net: input −11% ($0.32 → $0.29 per 1M tokens); output −10% ($2.09 → $1.89 per 1M tokens); cache read −32% ($0.16 → $0.11 per 1M tokens)detected

Aug 1, 2026Price changeIo Net raised Qwen3.6 35B A3B pricing by 72%input +48% ($0.22 → $0.32 per 1M tokens); output +72% ($1.22 → $2.09 per 1M tokens)detected

Jul 30, 2026Price changeDeepInfra cut Qwen3.6 35B A3B pricing by 33%input −33% ($0.15 → $0.10 per 1M tokens)detected

Jul 28, 2026Price changevenice repriced qwen/qwen3.6-35b-a3binput $0.15 → $0.1, output $1 → $1 per 1M tokensdetected

Jul 26, 2026ListedListed on LLMapfirst indexed by our pipelinedetected

Apr 15, 2026AnnouncedQwen3.6 35B A3B announced by Qwenannounced

Prices last checked 13h ago

What we do not know about this model yet

We hold no measured file for it, so all 3 sizes on this page are calculated from the parameter count.
No board we watch has turned up a score, so we hold no quality figures at all.
2 of 12 listings publish no parameter list, so what their API accepts is unknown to us.
Nothing we hold says whether an endpoint streams, so we do not show it either way.
3 of 12 listings do not say whether they train on prompts.

Licence and identifiers

What the licence allowsApache License 2.0, what it allows commercially, and the identifiers you need to pull this model — its Hugging Face repo, our slug and a machine-readable card.

Licence

Apache License 2.0

permissiveCommercial use allowed

Fully permissive: commercial use, redistribution, and derivatives allowed. Requires attribution and a copy of the license. Includes an express patent grant.

Identifiers

Hugging Face: Qwen/Qwen3.6-35B-A3B
Architecture: Mixture of experts
Modality record: text+image+video->text
Catalogue slug: qwen-qwen3-6-35b-a3b

Machine-readable model card (omc.json) →

Explore furtherWhere to run Qwen3.6 35B A3B Run it on GeForce RTX 5090 Run it on Apple M3 Pro (18-core GPU)Run it on L40S Best open-weight models Longest context windows

Something wrong on this page? Tell us