Qwen3 235B A22B Instruct 2507
AlibabaReleased Jul 21, 2025Open weights · Apache 2.0qwen3-235b-a22b
- Intelligence (Epoch)
- 139 · 84th of 111
- Reasoning
- no
- Knowledge cutoff
- September 2025
- Context
- 262K tokens
- Max output
- 32k tokens
- Input
- text
- API features
- tools, structured outputs, JSON mode, seed, logprobs
- Alibaba's API
- yes
Read Oct 23, 06:00 UTC from the hosts' published offers and OpenRouter, spec sheet from Alibaba · cutoff: aiknowledgecutoff.com
The cheapest host, GMICloud, serves fp8 weights (maker: bf16). Weights and context per host
Intelligence for the price
8 models score higher for the same price or less; it ranks 9 of the 23 models at its price or below.
Epoch Capabilities Index against the lowest input + output price per M, log scale · prices Oct 23. Score: Epoch AI, CC BY 4.0, 139 with an interval of 135 to 141.
Against its rivals
The best of its price range and the top score
- Qwen3 235B A22B Instruct 2507this model$0.44139
- Claude Opus 5.5top score$24.00169
- Claude Haiku 5.5same price range$0.60159
- GPT-6 Lunasame price range$0.60157
Same level for 3.7× less: gpt-oss-20b at AkashML, $0.12 per M input + output · their Epoch intervals overlap (138 against 139).
Bars: Epoch Capabilities Index (Epoch AI, CC BY 4.0), drawn from 100, below the lowest score in the catalogue · price: lowest input + output per M.
Speed by provider
Output speed of Qwen3 235B A22B Instruct 2507 at each host, fastest first
- Parasail35 tok/s0.48 s to first token
- DeepInfra10 tok/s0.57 s to first token
- Venice9 tok/s0.77 s to first token
- Nebius6 tok/s13.6 s to first token
Speed: OpenRouter, median of the last 30 min, read Oct 23, 06:00 UTC · a snapshot of real traffic before the daily read, not a daily average · bars from zero.
Run it yourself
open weights, Apache 2.0 · fits 1 of the 6 workstations in our guideSelf-host guide ›- Parameters
- 235B
- 22B active per token (MoE)
- Memory needed
- 141 GB
- in INT4 · 282 GB in INT8
- Smallest cloud setup
- 2× H100
- 160 GB · $4.30/h, about $3,139 a month
- Worth it above
- 7B tokens
- a month, against the cheapest API at $0.44 per M
| Quantization | Memory | Smallest cloud GPU | Quality |
|---|---|---|---|
| INT4 (Q4) | 141 GB | 2× H100 · $4.30/h | smallest footprint, minor quality loss |
| INT8 (Q8) | 282 GB | 4× H100 · $8.60/h | near-lossless |
| FP16 (full) | 564 GB | 8× H100 · $17.20/h | reference quality |
At home: in INT4 it fits a Mac Studio M5 Ultra (256 GB unified), about $9,499 once, or $0.3615/h over 3 years, electricity aside; it pays for itself after about 22B tokens against the cheapest API.
Memory ≈ parameters × bytes per parameter × 1.2 · cloud price: on-demand rate, running 24/7 · builds: GGUF, AWQ, MLX, GPTQ, Ollama · break-even calculator
| Provider | Input | Output | Cache ↑ | tok/s | First token | Uptime 24h | Checked |
|---|---|---|---|---|---|---|---|
| GMICloudfp8 (bf16) · 256K ctx−75% | $0.0875 | $0.35 | $0.0175 | — | — | — | Oct 19 |
| Parasailfp8 (bf16) · 128K ctx (256K) | $0.14 | $0.80 | $0.05 | 35 | 0.48 s | 100% | Oct 23 |
| AtlasCloudprecision not read: host terms · 128K ctx (256K) | $0.20 | $0.88 | $0.20 | — | — | 98.5% | Sep 15 |
| DeepInfrafp8 (bf16) · 256K ctx · 16K out | $0.09 | $0.55 | — | 10 | 0.57 s | 98.5% | Oct 23 |
| Novita AIprecision not stated · 128K ctx (256K) · 16K out | $0.09 | $0.58 | — | — | — | 86.0% | Oct 8 |
| Alibabaprecision not read: host terms · 128K ctx (256K) · 32K out | $0.15 | $0.60 | — | — | — | 89.6% | Oct 9 |
| Nebiusfp8 (bf16) · 256K ctx | $0.20 | $0.60 | — | 6 | 13.6 s | 45.2% | Oct 22 |
| Venicefp8 (bf16) · 125K ctx (256K) · 16K out | $0.15 | $0.75 | — | 9 | 0.77 s | 96.7% | Oct 23 |
| StreamLakeprecision not read: host terms · 125K ctx (256K) · 32K out−40% | $0.21 | $0.84 | — | — | — | 99.9% | Oct 9 |
| Googleprecision not stated · 256K ctx · 16K out | $0.22 | $0.88 | — | — | — | 15.0% | Oct 22 |
10 of 10 providers · prices per million tokens, read Oct 23, 06:00 UTC · speed: OpenRouter, median of the last 30 min, read Oct 23, 06:00 UTC · uptime: OpenRouter, last 24 h · −N%: a discount off the host's list price · under each host: weights precision, context and max output as served, the maker's value in brackets when the host's differs
| Provider | Weights | Context | Max output | Tools | Structured outputs | Reasoning |
|---|---|---|---|---|---|---|
| GMICloud | fp8 | 256K | not read: host terms | yes | yes | no |
| DeepInfra | fp8 | 256K | 16K | yes | yes | no |
| Novita AI | not stated by the host | 128K | 16K | — | — | — |
| Alibaba | not read: host terms | 128K | 32K | yes | no | no |
| Nebius | fp8 | 256K | not stated by the host | yes | yes | no |
| Venice | fp8 | 125K | 16K | yes | yes | no |
| Parasail | fp8 | 128K | not read: host terms | yes | yes | no |
| StreamLake | not read: host terms | 125K | 32K | yes | yes | no |
| AtlasCloud | not read: host terms | 128K | not read: host terms | — | — | — |
| not stated by the host | 256K | 16K | yes | yes | no |
Weights, context and max output: read at each host's own catalogue or docs where its terms allow, otherwise as the host declares to OpenRouter; hover a cell for its source and date · tools, structured outputs, reasoning: as declared to OpenRouter · not verified by us
Qwen3 235B A22B Instruct 2507 spec sheet, field by field, with sources
Qwen3 235B A22B Instruct 2507 spec sheetmaker-declared
What Alibaba declares about the model itself — read from its own weights repository and documentation, not from any host. What each endpoint actually serves is in the pricing above: the two can differ, and both are true.
- Context window
- 262,144 tokens
- Parameters
- 235,093,634,560≈ 235.1B
- Active experts
- 8 of 128 active per token
- Weight precision
- bfloat16
- License
- Apache 2.0 — open weights, permissivecurated · read 2026-09-07Scope: the licence on the published weights, and nothing else. It is not the maker's acceptable-use policy, which is a separate document, and it is not the contract you sign with whichever host you call — hosts set their own terms, which this index does not read.
- Knowledge cutoff
- not published by the maker≈ September 2025 third-party estimate — not the maker's figurecurated
- Modalities
- text
- Max output (official)
- to verify
- Release date
- to verify≈ July 21, 2025 third-party catalogue — not the maker's figure
- Maker lifecycle status
- to verify
Also declared in the weights
- Architecture class
- Qwen3MoeForCausalLM
- Layers
- 94
- Vocabulary
- 151,936 tokens
- Weights first published
- July 21, 2025
- Weights last updated
- September 17, 2025
Qwen/Qwen3-235B-A22B-Instruct-2507 · config.json @ ac9c66c · read 2026-10-08 · 6 of 10 fields published by the maker
| Date | Provider | Input | Output | Change |
|---|---|---|---|---|
| Sep 5 | $0.25 → $0.22 | $1.00 → $0.88 | −12% |
Every published change of input or output price, newest first · change on input + output · full series: pricing-history.csv.
| Provider | Chatbot 1,000 in / 500 out | RAG / long context 8,000 in / 500 out | Batch summarize 4,000 in / 1,000 out |
|---|---|---|---|
| GMICloud | $0.26 | $0.88 | $0.70 |
| DeepInfra | $0.36 | $0.99 | $0.91 |
| Novita AI | $0.38 | $1.01 | $0.94 |
| Alibaba | $0.45 | $1.50 | $1.20 |
| Nebius | $0.50 | $1.90 | $1.40 |
| Venice | $0.53 | $1.57 | $1.35 |
| Parasail | $0.54 | $1.52 | $1.36 |
| StreamLake | $0.63 | $2.10 | $1.68 |
| AtlasCloud | $0.64 | $2.04 | $1.68 |
| $0.66 | $2.20 | $1.76 |
What each host bills for 1,000 requests of a representative job
Your own workload
Other Alibaba models · all its models, hosts and history ›
- Qwen3.8 27B16 providers · from $1.29
- Qwen3.6 35B A3B12 providers · from $1.00
- Qwen3.5 397B A17B10 providers · from $3.45
- Qwen3.6 27B9 providers · from $2.25
- Qwen3.5-35B-A3B8 providers · from $1.14
- Qwen3 30B A3B7 providers · from $0.24