Flagship Qwen3 — top-tier reasoning and coding from Alibaba.
qwen3-max is Alibaba's flagship commercial model in the Qwen3 generation. It leads the Qwen lineup on reasoning, coding, math, and complex tool use, with a 256K context window for repo-scale code, long documents, and multi-turn agent traces.
TheRouter ships this model on both the mainland (bailian-cn) and Singapore (bailian-sg) Bailian endpoints. The selector picks the cheaper region per request so you get vendor-direct latency without manual region pinning.
| Type | Rate |
|---|---|
| Input | $1.30 / 1M tokens |
| Output | $6.48 / 1M tokens |
| Stable alias | qwen3-max → qwen3-max-2026-01-23www.alibabacloud.com ↗ | verified |
| Available snapshots | qwen3-max-2026-01-23; qwen3-max-preview; qwen3-max-2025-09-23www.alibabacloud.com ↗ | verified |
| Architecture | Qwen3 Mixture-of-Experts (MoE) with over 1 trillion total parameterswww.alibabacloud.com ↗ | verified |
| Pretraining data | 36 trillion tokenswww.alibabacloud.com ↗ | verified |
| Context window | 256K tokens upstreamwww.alibabacloud.com ↗ | verified |
| Max output | 32,768 tokens in TheRouter configgithub.com ↗ | verified |
| Thinking mode | Hybrid thinking is supported upstream for qwen3-max and qwen3-max-2026-01-23; thinking is disabled by default. TheRouter's qwen/qwen3-max public config does not currently expose a reasoning parameter.www.alibabacloud.com ↗ | verified |
| API capabilities | OpenAI-compatible chat, streaming, function calling, structured output / JSON mode, and stop sequences through TheRouter; built-in Alibaba tools are documented upstream but not represented as separate TheRouter endpoints.www.alibabacloud.com ↗ | verified |
| TheRouter price bands | $1.20/M input + $6.00/M output up to 32K; $2.40/M input + $12.00/M output after 32K. Alibaba also publishes a third 128K-256K band ($3/$15) that TheRouter's current single-threshold schema cannot fully encode.www.alibabacloud.com ↗ | verified |
| License | Proprietary (closed-weight, API-only) | verified |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
SWE-bench Verified Alibaba reports Qwen3-Max-Instruct at 69.6 on SWE-bench Verified, positioning it among top coding models at launch. | 69.6score | alibabacloud.com ↗ | |
Tau2-Bench Alibaba describes Tau2-Bench as an evaluation of agent tool-calling proficiency and says Qwen3-Max-Instruct surpassed Claude Opus 4 and DeepSeek V3.1 on this benchmark. | 74.8score | alibabacloud.com ↗ | |
LMArena Text Arena The launch article says Qwen3-Max-Instruct preview ranked third on the Text Arena leaderboard and surpassed GPT-5-Chat. A direct leaderboard snapshot is not embedded in the article, so this is treated as an official claim rather than a reproduced leaderboard row. | Top-three global ranking for previewrank | alibabacloud.com ↗ | |
AIME 25 / HMMT This is explicitly for the still-training Qwen3-Max-Thinking variant augmented with a code interpreter and parallel test-time compute, not for the default non-thinking qwen3-max call path. | 100% with Qwen3-Max-Thinking + tools + scaled test-time compute% | alibabacloud.com ↗ |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "qwen/qwen3-max",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'Use Qwen3-Max through TheRouter's OpenAI-compatible Chat Completions endpoint for difficult text, coding, and agent prompts.
curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen/qwen3-max",
"messages": [{"role": "user", "content": "Summarize this vendor contract and list risky clauses."}],
"temperature": 0.2,
"top_p": 0.8,
"max_tokens": 4096
}'Alibaba Cloud's pricing table identifies qwen3-max as currently equivalent to qwen3-max-2026-01-23, with non-thinking and thinking modes available for the current snapshot and tiered billing up to 256K tokens.
The Qwen team introduced Qwen3-Max with over 1 trillion parameters, 36 trillion pretraining tokens, OpenAI-compatible API access as qwen3-max, and reported gains on coding and agent benchmarks.
For TheRouter users, qwen/qwen3-max is the public id that routes to Alibaba's qwen3-max chat/instruct API. Alibaba's launch article names the API model qwen3-max for Qwen3-Max-Instruct, while the pricing page says the stable alias now maps to qwen3-max-2026-01-23.
Upstream Alibaba documents hybrid thinking for qwen3-max and qwen3-max-2026-01-23, disabled by default. TheRouter's current public model config for qwen/qwen3-max does not expose reasoning, so treat TheRouter calls as the standard non-thinking chat path unless your account has a provider-specific override.
The upstream Alibaba model list marks qwen3-max as 256K context, and TheRouter config sets qwen/qwen3-max to 262,144 tokens with 32,768 max completion tokens. If you need 1M context, compare qwen/qwen3.7-max or qwen/qwen3.7-plus instead.
For new maximum-capability deployments, start with qwen3.7-max because Alibaba's current guide recommends it for the strongest reasoning and 1M context. Keep qwen3-max when you need the 256K class, existing snapshot behavior, or current TheRouter pricing/routing stability.
| source | URL | retrieved | |
|---|---|---|---|
| Stable alias | www.alibabacloud.com ↗ | 2026-07-30 | verified |
| Available snapshots | www.alibabacloud.com ↗ | 2026-07-30 | verified |
| Architecture | www.alibabacloud.com ↗ | 2026-07-30 | verified |
| Pretraining data | www.alibabacloud.com ↗ | 2026-07-30 | verified |
| Context window | www.alibabacloud.com ↗ | 2026-07-30 | verified |
| Max output | github.com ↗ | 2026-07-30 | verified |
| Thinking mode | www.alibabacloud.com ↗ | 2026-07-30 | verified |
| API capabilities | www.alibabacloud.com ↗ | 2026-07-30 | verified |
| TheRouter price bands | www.alibabacloud.com ↗ | 2026-07-30 | verified |
| License | — | — | verified |
| SWE-bench Verified | alibabacloud.com ↗ | 2026-07-30 | verified |
| Tau2-Bench | alibabacloud.com ↗ | 2026-07-30 | verified |
| LMArena Text Arena | alibabacloud.com ↗ | 2026-07-30 | to verify |
| AIME 25 / HMMT | alibabacloud.com ↗ | 2026-07-30 | to verify |
| qwen3-max stable alias advances to the January 2026 snapshot | alibabacloud.com ↗ | 2026-07-30 | verified |
| Qwen3-Max launches as Alibaba's trillion-parameter Qwen flagship | alibabacloud.com ↗ | 2026-07-30 | verified |
| Is qwen/qwen3-max the same as Qwen3-Max-Instruct? | www.alibabacloud.com ↗ | 2026-07-30 | to verify |
| Does qwen/qwen3-max support thinking mode on TheRouter? | www.alibabacloud.com ↗ | 2026-07-30 | to verify |
| How much context does qwen/qwen3-max support? | www.alibabacloud.com ↗ | 2026-07-30 | to verify |
| Should I choose qwen3-max or qwen3.7-max? | www.alibabacloud.com ↗ | 2026-07-30 | to verify |