Qwen3.5-35B-A3B
Qwen · released Feb 24, 2026 · Qwen/Qwen3.5-35B-A3B
- Type
- Open weightsApache License 2.0
- Params
- 36B
- Context
- 262K
3B active per word · about 197K words of context
Our take
Written Aug 2, 2026Qwen3.5-35B-A3B is a downloadable model from Alibaba's Qwen line that activates only 3 billion of its 36 billion parameters for each word it processes, making it unusually efficient to run locally while keeping a permissive Apache licence. It handles text, images and video in a single request of up to 262,144 tokens, and scores well on coding tasks though creative writing is its weakest area.
Pick this for maximum parameter efficiency on local hardware — only 3 billion parameters activate per token from 36 billion total. Use it for budget hosted inference with strong throughput, or long-context workloads under a permissive licence. Skip it if web-development coding is your main need, or if you need peak creative-writing quality.
The case for it
- Extreme active-parameter efficiency: only 3 billion active per token from 36 billion total, an 8:1 compression ratio.
- Coding performance 39 points above its overall chat ranking on the independent leaderboard we track.
- Best throughput-to-price ratio on the budget tier: 180 tokens per second at the cheapest tracked offer, nearly double the speed of another host at the same price.
- Truly permissive Apache 2.0 licence allows commercial use, fine-tuning and redistribution.
The case against it
- Web-development coding is notably weaker than general coding, 183.9 points lower on the relevant leaderboard category.
- Premium providers charge steep markups for marginal throughput gains — only 2 tokens per second faster for an input price 79% higher.
- Creative writing is the weakest arena category, 52 points below its overall text ranking.
How good is it?
IntelligencePuzzles, maths, exam questions
Arena Text (overall)79th of 143 · 1395.5
CodingWriting and fixing code on its own
Arena Coding82nd of 143 · 1434.9
AgenticPlanning, calling tools, staying on task
Nobody we watch has scored Qwen3.5-35B-A3B for this. We would take the rating from Arena Agent (IPS).
WritingWe do not rate this
Two boards come close and neither tests writing: Arena Creative Writing asks people which of two replies they prefer, and LiveBench Language tests whether a model understood a passage. So we show where Qwen3.5-35B-A3B placed and give it no mark out of five.
These tests check whether a model follows instructions — a precondition for all the work above, but not a measure of how well that work is done, which is why they get no rating.
Every published score for this model7 scoresEvery figure we hold, from 7 boards, with who ran it and a link to the source — including the boards no rating above is built on.
Can you run it yourself?
- 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%
GeForce RTX 4090 · 24 GB
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
GeForce RTX 5090 · 32 GB
Room to spare. 6.1 GB spare means a 10% error in the size would not change the answer.
Apple M3 Pro (18-core GPU) · 36 GB
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.
Memory use by level
Against a 24 GB card.
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
Cheapest of 13 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
| Provider | In / out per 1M tokens | Context | Throughput | Trains on prompts | Logs prompts | Zero retention |
|---|---|---|---|---|---|---|
| OpenRouter | $0.14 / $1.00 | 262K | not measured | Unknown | Unknown | Unknown |
| DeepInfrafp8 | $0.14 / $1.00 | 262K | not measured | Unknown | Unknown | Unknown |
| DeepInfrafp8 | $0.14 / $1.00 | 262K | 56 tok/s | No | No | Confirmed |
| Parasailfp8 | $0.15 / $1.00 | 262K | 82 tok/s | No | No | Confirmed |
| AkashMLfp8 | $0.14 / $1.00 | 262K | 128 tok/s | No | No | Confirmed |
| Venice AI | $0.31 / $1.25 | 256K | 131 tok/s | No | No | Confirmed |
| CoreWeavefp8 | $0.25 / $1.25 | 262K | 213 tok/s | No | No | Confirmed |
| Alibaba Cloud | $0.16 / $1.30 | 262K | 146 tok/s | No | Yesunknown period | Unknown |
| Alibaba Cloudfp8 | $0.16 / $1.30 | 262K | 51 tok/s | No | Yesunknown period | Unknown |
| NextBitfp8 | $0.22 / $1.40 | 262K | 67 tok/s | No | No | Confirmed |
| AtlasCloudfp8 | $0.23 / $1.80 | 262K | 107 tok/s | No | Yesunknown period | Unknown |
| SiliconFlowfp8 | $0.24 / $1.80 | 262K | 28 tok/s | No | No | Confirmed |
| Novita AI | $0.25 / $2.00 | 262K | not measured | Unknown | Unknown | Unknown |
DeepInfra
Novita AIAcross the 13 listings we hold: 10 say they do not train on prompts, 0 say they do and 3 do not say. 7 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
| Provider | Tool calling | JSON output | Strict schema |
|---|---|---|---|
| OpenRouter | ✓ | ✓ | ✓ |
| DeepInfrafp8 | |||
| DeepInfrafp8 | ✗ | ✓ | ✓ |
| Parasailfp8 | ✓ | ✓ | ✓ |
| AkashMLfp8 | ✓ | ✓ | ✓ |
| Venice AI | ✓ | ✓ | ✓ |
| CoreWeavefp8 | ✓ | ✓ | ✓ |
| Alibaba Cloud | ✓ | ✓ | ✓ |
| Alibaba Cloudfp8 | ✓ | ✓ | ✓ |
| NextBitfp8 | ✓ | ✓ | ✓ |
| AtlasCloudfp8 | ✓ | ✓ | ✓ |
| SiliconFlowfp8 | ✗ | ✓ | ✓ |
| Novita AI |
Tool calling: 9 of 13 listings say yes, 2 say no, 2 publish no parameter list. JSON output: 11 of 13 listings say yes, 2 publish no parameter list. Strict schema: 11 of 13 listings say yes, 2 publish no parameter list.
Models people weigh against Qwen3.5-35B-A3B
When we formed this view
Dates behind this page
Prices last checked 34h 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.
- 2 of 13 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 13 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
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/Qwen3.5-35B-A3B
- Architecture
- Mixture of experts
- Modality record
- text+image+video->text
- Catalogue slug
- qwen-qwen3-5-35b-a3b