Qwen3.6 Text Model Series

Alibaba’s Qwen3.6 family on APIYI: Max-Preview coding flagship + Flash speed-first multimodal + Plus balanced workhorse + 27B / 35B-A3B open-weight variants (hosted by APIYI — no GPU rental needed). All routed via Aliyun’s official relay, OpenAI Chat-compatible. List prices match Alibaba’s official rate; APIYI’s recharge bonus brings effective unit prices to ~85% of list.

Qwen3.6 is Alibaba Tongyi Qianwen’s next-generation model family, released in Q2 2026 across three closed-source production tiers — Max (flagship), Plus (balanced), Flash (speed-first) — plus two open-weight variants: 27B and 35B-A3B. APIYI routes all five models through Aliyun official relay / APIYI hosted relay, OpenAI Chat Completions compatible. Closed-source tiers match the official portal’s auth and rate-limit policies; the open-weight tiers are hosted by APIYI’s official relay so customers don’t have to rent GPUs or stand up local inference.

🚀 Highlights: Max-Preview claims #1 on six coding benchmarks including SWE-bench Pro and Terminal-Bench 2.0; Flash is a 35B-A3B MoE with native 256K (expandable to 1M) multimodal context; Plus is a 72B/18B-active workhorse with a 1M context window. The open-weight qwen3.6-27b (27B dense) and qwen3.6-35b-a3b (35B MoE / 3B active) are hosted on APIYI’s official relay — no GPU rental needed, billed per token. Built for coding agents, long-context RAG, multimodal dispatch, and compliance-sensitive workloads needing auditable weights.

Closed-source production tiers (Aliyun official relay)

qwen3.6-max-preview

Coding flagship

#1 on 6 coding benchmarks; AIME 2025 93%, GPQA 86%, LiveCodeBench 79%.

qwen3.6-flash

Speed-first multimodal

35B-A3B MoE, native text / image / video input, 256K base context expandable to 1M.

qwen3.6-plus

Balanced workhorse

72B total / 18B active, 1M context, Terminal-Bench 61.6 beats Claude Opus 4.5.

Open-weight tiers (hosted by APIYI · no GPU rental)

qwen3.6-27b

27B dense · coding powerhouse

Qwen team’s open-weight release (Hugging Face Qwen/Qwen3.6-27B). Coding ability rivals 397B-class models. Hosted by APIYI’s official relay — no local GPUs required.

qwen3.6-35b-a3b

35B-A3B open-weight MoE

Qwen team’s open-weight release (Hugging Face Qwen/Qwen3.6-35B-A3B); same lineage as closed-source Flash, different distribution tier. Only 3B active params — extremely low compute cost.

Why APIYI’s Qwen3.6 via Aliyun official relay?

Calibrated against Alibaba Cloud Bailian’s official channel, with deep optimization for enterprise production across stability, cost, and integration ergonomics:

Aliyun official relay

Routed via Alibaba Cloud Bailian’s official channel. Auth and rate-limit policies match the official portal — low domestic latency, enterprise-grade SLA.

No concurrency cap · scale freely

No hard RPM / TPM ceilings (subject to upstream supply). Enterprise customers can scale on demand; tickets and dedicated channels available for high-concurrency coordination.

List-price match + ~15% off via recharge

List price matches Alibaba Cloud’s official rate. Stack with recharge bonuses for an effective unit price around 85% of list.

Global zero-friction access

No overseas server or proxy required. Domestic data centers, residential broadband, and overseas nodes can all connect directly to api.apiyi.com — no overseas migration needed.

Full OpenAI-compatible ecosystem

OpenAI Chat Completions compatible. Switch seamlessly across GPT / Claude / DeepSeek / GLM and more via APIYI’s unified model catalog.

Professional service · enterprise support

Deep expertise in model selection and Agent workflows; full PoC → canary → production support for enterprise customers.

How to choose among the five

Max-Preview · coding & complex reasoning

Scenarios: Coding Agent driver, real-world software-engineering tasks (SWE-Verified class), Cursor / Claude Code workflow primary model.

Benchmarks: SWE-bench Pro 58.4 (beats GLM-5.1’s 56.6), AIME 2025 93%, GPQA 86%, LiveCodeBench 79%, Terminal-Bench 2.0 #1.

Note: Marked Preview — weights still iterating. Run a small canary before flipping main traffic.

Flash · high-volume multimodal long-context

Scenarios: Image / video understanding, long-document summarization, bulk translation, post-RAG full-document synthesis.

Architecture: 35B total / 3B active MoE (35B-A3B), native 256K context expandable to 1M tokens.

Multimodal: Native text / image / video input. Unit price ~1/8 that of Max.

Plus · the balanced workhorse

Scenarios: Daily dialog, customer support, content generation, enterprise knowledge-base Q&A, mid-complexity reasoning.

Architecture: 72B total / 18B active MoE — inference speed roughly 3× Claude Opus 4.6.

Benchmarks: Terminal-Bench 2.0 at 61.6 beats Claude Opus 4.5 (59.3); SWE-bench Verified 78.8.

qwen3.6-27b · open-weight coding powerhouse

Scenarios: cost-sensitive coding assistance; API-validation phase before committing to local deploy; customers with compliance requirements for auditable open-source licenses.

Notes: 27B dense, open weights, coding ability rivals 397B-class models. Hosted by APIYI — no local GPU.

qwen3.6-35b-a3b · open-weight speed MoE

Scenarios: high-frequency low-cost workflows; transition phase before moving to self-hosted inference; projects requiring downloadable weights for compliance.

Notes: Same lineage as closed-source Flash (35B total / 3B active). Open-weight version hosted by APIYI — skip GPU rental, deployment, and ops.

Recommended routing

Strategy: Flash by default + Plus on escalation + Max-Preview as ceiling; downgrade to open-weight 27b / 35b-a3b for ultra-cost-sensitive workloads.

Route routine dialog and multimodal batches to Flash; escalate to Plus for stronger reasoning; reserve Max-Preview for coding agents, complex reasoning, and multi-step planning; drop to the open-weight tiers when cost matters most or auditable weights are required.

Pricing

All five models bill under pay-as-you-go - Chat. Closed-source tiers (Max-Preview / Flash / Plus) use tiered pricing keyed on the total input token count of a single request. Open-weight tiers (27b / 35b-a3b) bill at a single flat rate — no tiers. List prices match Alibaba Cloud’s official rate; APIYI’s recharge bonus brings the effective unit price to roughly 85% of list.

qwen3.6-max-preview

Single-request input tokens	Input price	Output price
0 – 128K	$1.2800 / 1M tokens	$7.6800 / 1M tokens
128K – 256K	$2.1200 / 1M tokens	$12.7200 / 1M tokens

qwen3.6-flash

Single-request input tokens	Input price	Output price
0 – 256K	$0.1700 / 1M tokens	$1.0200 / 1M tokens
256K – 1000K	$0.6800 / 1M tokens	$4.0800 / 1M tokens

qwen3.6-plus

Single-request input tokens	Input price	Output price
0 – 256K	$0.3000 / 1M tokens	$1.8000 / 1M tokens
256K – 1000K	$1.2000 / 1M tokens	$7.2000 / 1M tokens

qwen3.6-27b (open-weight · APIYI hosted)

Billing	Input price	Output price
Flat (no tiers)	$0.4200 / 1M tokens	$2.5200 / 1M tokens

qwen3.6-35b-a3b (open-weight · APIYI hosted)

Billing	Input price	Output price
Flat (no tiers)	$0.2600 / 1M tokens	$1.5600 / 1M tokens

Pricing notes:

• Closed-source tiers (tiered pricing): the tier is set by the total input tokens of a single request. All tokens in that request (input + output) bill at that tier’s rate. No cross-tier proration — e.g., a Flash request with 300K input tokens lands in 256K – 1000K and the entire request bills at $0.68 / $4.08, not split as “first 256K cheap, remaining 44K at the higher tier.”
• Open-weight tiers (flat pricing): qwen3.6-27b and qwen3.6-35b-a3b are hosted by APIYI’s official relay — no tiers. Customers don’t have to rent GPUs or run local inference; settle directly by actual token consumption.
• List prices match Alibaba Cloud Bailian. With recharge bonuses, the effective unit price lands around 85% of list.
• Cache-hit pricing is not currently disclosed separately; falls back to the base tier.

Specs

Closed-source production tiers

Dimension	qwen3.6-max-preview	qwen3.6-flash	qwen3.6-plus
Model ID	qwen3.6-max-preview	qwen3.6-flash	qwen3.6-plus
Architecture	Dense large model	MoE 35B-A3B	MoE 72B / 18B active
Context	262K tokens	256K (expandable to 1M)	1M tokens
Input modalities	Text	Text / image / video	Text
Output format	Text	Text	Text
Streaming	✅ Supported	✅ Supported	✅ Supported
Function calling / tool use	✅ Supported	✅ Supported	✅ Supported
Chain-of-thought	✅ Auto-enabled on reasoning tasks	—	✅ Always on
Billing	Pay-as-you-go Chat (tiered)	Pay-as-you-go Chat (tiered)	Pay-as-you-go Chat (tiered)
Channel	Aliyun official relay	Aliyun official relay	Aliyun official relay

Open-weight tiers (hosted by APIYI)

Dimension	qwen3.6-27b	qwen3.6-35b-a3b
Model ID	qwen3.6-27b	qwen3.6-35b-a3b
Architecture	27B dense	MoE 35B total / 3B active
License	Qwen team open-weight (Hugging Face Qwen/Qwen3.6-27B)	Qwen team open-weight (Hugging Face Qwen/Qwen3.6-35B-A3B)
Context	Matches official weight card	Matches official weight card
Input modalities	Text	Text
Streaming	✅ Supported	✅ Supported
Function calling / tool use	✅ Supported	✅ Supported
Billing	Pay-as-you-go Chat (flat, no tiers)	Pay-as-you-go Chat (flat, no tiers)
Channel	APIYI hosted relay	APIYI hosted relay

Why hosted open-weights: open-weight checkpoints are publicly downloadable, but running them needs GPUs, VRAM, and ops. APIYI hosts these open weights on its official relay, so you call the API directly — keeping the “auditable weights, controllable license” upside while skipping rental, deploy, and ops costs.

Endpoints

Endpoint	Method	Content-Type	Purpose
/v1/chat/completions	POST	application/json	Dialog / reasoning / tool use (shared by all five models, only the model field differs)

Domain: api.apiyi.com is the primary gateway. Alternative gateways like b.apiyi.com / vip.apiyi.com produce identical responses. Set base_url to https://api.apiyi.com/v1 to use the OpenAI / OpenAI-compatible SDK directly.

Code examples

Python (OpenAI SDK compatible)

from openai import OpenAI

client = OpenAI(
    api_key="sk-your-api-key",
    base_url="https://api.apiyi.com/v1"
)

# Max-Preview: coding Agent driver
resp = client.chat.completions.create(
    model="qwen3.6-max-preview",
    messages=[
        {"role": "system", "content": "You are a senior Python engineer. Return changes as a unified diff."},
        {"role": "user", "content": "Add type hints and fix any latent bugs in this snippet ..."}
    ]
)
print(resp.choices[0].message.content)

# Flash: image + text multimodal input
resp = client.chat.completions.create(
    model="qwen3.6-flash",
    messages=[
        {"role": "user", "content": [
            {"type": "text", "text": "Describe the key information in this image."},
            {"type": "image_url", "image_url": {"url": "https://your-image-url.png"}}
        ]}
    ]
)
print(resp.choices[0].message.content)

# Plus: daily dialog and mid-complexity reasoning
resp = client.chat.completions.create(
    model="qwen3.6-plus",
    messages=[{"role": "user", "content": "Introduce yourself in one sentence."}]
)
print(resp.choices[0].message.content)

Node.js

import OpenAI from 'openai';

const client = new OpenAI({
  apiKey: 'sk-your-api-key',
  baseURL: 'https://api.apiyi.com/v1',
});

const resp = await client.chat.completions.create({
  model: 'qwen3.6-plus',
  messages: [{ role: 'user', content: 'Introduce yourself in one sentence.' }],
});

console.log(resp.choices[0].message.content);

cURL

curl -X POST "https://api.apiyi.com/v1/chat/completions" \
  -H "Authorization: Bearer sk-your-api-key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3.6-max-preview",
    "messages": [
      {"role": "user", "content": "Explain what an MoE architecture is."}
    ]
  }'

Best practices

1. Pick the right tier per task

   Default to Flash for routine dialog / classification / multimodal batching. Use Plus for mid-complexity reasoning and enterprise knowledge-base Q&A. Only escalate to Max-Preview for coding agents, complex planning, or competition-level math reasoning. Drop a tier whenever you can.

2. Estimate the tier boundary

   Profile your P95 input token count before launch. Max-Preview crossing 128K, or Flash / Plus crossing 256K, sees a sharp price jump. Summarize / chunk extra-long context to keep P95 within the lower tier.

3. Multimodal batching

   Flash supports a 1M context and video input, but a single ultra-long request triggers the higher tier. Slice long video into segments, then feed in chunks within 256K to control per-call cost.

4. Preview canary

   qwen3.6-max-preview is a Preview build — weights still iterating. For critical paths, run a small canary with A/B comparison before flipping main traffic.

5. Tools & streaming

   All three models support OpenAI-style tools and stream: true. You can drop them into existing OpenAI-compatible Agent frameworks (OpenClaw, LangChain, LlamaIndex, etc.) without rewriting tool-calling logic.

6. Stack with recharge bonuses

   List prices already match Alibaba’s official rate. Combined with recharge bonuses, effective unit price lands around 85% of list. Larger top-ups ($1,000+) earn higher bonus ratios — top up in fewer, larger transactions for the best margin.

Errors & retries

Status	Meaning	What to do
400	Param error / unknown model	Check model spelling, messages shape, and whether the input exceeds max context
401	Invalid token	Verify the Bearer Token
403	Content moderation block	Adjust prompts / reference inputs to avoid policy violations
429	Rate-limit / insufficient balance	Exponential backoff retry; check account balance
5xx	Gateway / backend error	Retry 1–2 times; if still failing, file a ticket
Timeout	Long-tail latency	Set client timeout to ≥ 120s (CoT or long-context calls take longer)

Client recommendations:

• Set request timeout to ≥ 120 seconds (Max-Preview reasoning and Plus long-context CoT take longer)
• Apply exponential backoff retry on 5xx and timeouts (recommend 2 attempts)
• Log the x-request-id response header for troubleshooting

FAQ

Do all five models share the same API endpoint?

Yes. All five share /v1/chat/completions (OpenAI Chat Completions compatible). Only the model field differs (qwen3.6-max-preview / qwen3.6-flash / qwen3.6-plus / qwen3.6-27b / qwen3.6-35b-a3b) — switch in the same codebase as needed.

What's the difference between open-weight (27b / 35b-a3b) and closed-source tiers?

Three main differences: (1) Downloadable weights — open-weight checkpoints are on Hugging Face for internal audit, compliance filing, or future migration to local inference; (2) Hosted compute — APIYI hosts the open weights on its official relay so you just call the API, no GPU rental / deploy / ops; (3) Simpler billing — open-weight tiers use flat pricing, no tiers, easier to budget. On capability: 35B-A3B shares lineage with closed-source Flash (different distribution tier); 27B is an independent dense model whose coding ability rivals models with much higher parameter counts.

If the weights are open, why use APIYI's hosted API?

Self-hosting an open large model needs at least: capable GPUs (27B requires at least one A100 40G; 35B-A3B needs more VRAM), an inference framework (vLLM / TensorRT-LLM), monitoring, failover, and an upgrade pipeline. APIYI’s hosted official relay handles all of that — billed by token, scales on demand, and shares the same OpenAI-compatible SDK as the closed-source tiers. Build with the API; later, decide whether to switch to self-hosting. The path stays smooth.

How exactly does tiered pricing work?

The tier is determined by the total input tokens of a single request. All tokens (input + output) in that request bill at the corresponding tier’s rate. Example: a Flash request with 300K input tokens lands in 256K – 1000K and bills the entire request at $0.68 / $4.08 — no split between “first 256K cheap, remaining 44K higher.”

Max-Preview is a Preview — is it production-ready?

Yes, but run a canary first. The Qwen team has stated subsequent revisions will continue to refine weights. For critical paths, A/B test against your benchmark tasks and only flip main traffic once a stable version lands.

How do I send multimodal input to Flash?

Use OpenAI’s Vision-compatible format: in messages, send content as an array where each element is {type: "text", text: ...} or {type: "image_url", image_url: {url: ...}}. For video, follow the official doc’s video_url / frame-sampling fields.

Are APIYI's prices the same as Alibaba Cloud Bailian's official site?

List price matches the official rate. The difference: APIYI stacks recharge bonuses for an effective unit price around 85% of list, plus a unified account that supports the entire OpenAI-compatible ecosystem (GPT / Claude / Gemini / DeepSeek / GLM, etc.) — no need to maintain multiple vendor accounts.

Are function calling / tool use supported?

Yes. All three models accept OpenAI-standard tools / tool_choice. You can reuse existing Agent-framework tool-calling logic. Max-Preview shines at multi-step tool calls and long-horizon planning.

Is chain-of-thought output supported?

Max-Preview auto-enables CoT on reasoning tasks; Plus has CoT always on; Flash is speed-first and does not output CoT by default. Field names follow Alibaba Cloud’s response format (reasoning_content, etc.).

What if my request exceeds 1M context?

Flash and Plus cap at 1M tokens; Max-Preview at 262K. Exceeding the cap returns 400. Apply summarization / chunking / RAG retrieval before sending — don’t try to push everything in a single call.

Can I use the official OpenAI SDK directly?

Yes. Set base_url to https://api.apiyi.com/v1 and pass any of the model IDs above as model — zero-code migration.

Are failed requests billed?

Client-side 4xx errors (param error / auth failure / content-moderation block) are not billed. Server-side 5xx errors that don’t reach inference are also not billed. Requests that successfully return tokens are billed by actual token count, even if the client cancels mid-stream.

Related docs

• Deep dive: Qwen3.6 Max-Preview & Flash launch
• Deep dive: Qwen3.6-Plus launch — Alibaba’s strongest coding agent
• Recharge promotions — drop unit price to ~85% of list
• Model catalog — all available models and groups
• API manual — general usage conventions

Wrap-up: The Qwen3.6 series covers the full demand curve from heavy coding agents to high-volume multimodal dispatch. Max-Preview pushes domestic coding to a new high; Flash collapses the unit price of multimodal long-context workloads; Plus is the dependable balanced workhorse; the open-weight 27B and 35B-A3B variants — hosted by APIYI’s official relay — close the loop on “controllable open weights without renting GPUs.” All five share an OpenAI Chat-compatible endpoint, list-price-matched with the official rate, and stack with recharge bonuses to roughly 85% of list — the best price/performance combo currently available in the Aliyun official-relay channel.