> ## Documentation Index > Fetch the complete documentation index at: https://docs.apiyi.com/llms.txt > Use this file to discover all available pages before exploring further. # Qwen3.6 Open-Weight Duo Live: APIYI-Hosted, No GPU Rental > Qwen3.6 series adds two open-weight models — qwen3.6-27b and qwen3.6-35b-a3b — hosted on APIYI's official relay. 27B's coding rivals 397B-class models; 35B-A3B shares lineage with closed-source Flash. Pay-as-you-go Chat flat pricing (27b $0.42/$2.52, 35b-a3b $0.26/$1.56). Customers skip the GPU rental, deploy, and ops. ## Highlights * **Qwen3.6 expands to 5 models**: in addition to closed-source Max-Preview / Flash / Plus, the series now includes `qwen3.6-27b` (27B dense) and `qwen3.6-35b-a3b` (35B MoE / 3B active) — both **open-weight** releases * **Hosted on APIYI's official relay**: open-weight ≠ free to run. Weights are public on Hugging Face (`Qwen/Qwen3.6-27B`, `Qwen/Qwen3.6-35B-A3B`), but local inference still needs GPUs, VRAM, an inference framework, and ops. **APIYI's hosted relay handles all of that** * **Skip GPU rental**: pay per token. Build with the API; later, decide whether to switch to self-hosting — a smooth path * **Flat pricing — no tiers**: the open-weight tiers use single-rate pay-as-you-go Chat: `qwen3.6-27b` at \$0.42 in / \$2.52 out, `qwen3.6-35b-a3b` at \$0.26 in / \$1.56 out per 1M tokens * **OpenAI-compatible**: shares the `/v1/chat/completions` endpoint with the closed-source tiers — only the `model` field differs. One SDK, five models * **\~85% of list with the recharge bonus**: list price matches Alibaba Cloud's official rate; APIYI's recharge bonus brings effective price down to roughly 85% of list Hosted from Qwen's open-weight checkpoints: `Qwen/Qwen3.6-27B` (Hugging Face) and `Qwen/Qwen3.6-35B-A3B` (Hugging Face). This launch is the **APIYI hosted relay** distribution — weights match the official release; calls follow OpenAI Chat Completions semantics. Source: Qwen team's Hugging Face model cards. Data retrieval: 2026-04-28 (UTC+8). ## Background Alongside the closed-source Qwen3.6 production tiers (Max / Plus / Flash), the Qwen team also released the 27B dense and 35B-A3B MoE checkpoints as open weights on Hugging Face — giving customers who need auditable weights, controllable licenses, or future self-hosting an off-ramp. But "open-weight" doesn't mean "easy to run." A 27B dense model needs at least an A100 40G; vLLM / TensorRT-LLM for inference; monitoring; failover; an upgrade pipeline. The 35B-A3B activates only 3B params, but VRAM still scales with the 35B total. **For most customers, "we want to use open Qwen3.6" and "we can run an open Qwen3.6 cluster" sit on opposite sides of an SRE team.** This launch closes that gap: **open weights + APIYI hosted relay = call the API directly, without owning the GPU**. Build the PoC on the API, quantify token cost, then decide later whether self-hosting makes sense. ## Deep dive ### Core features **27B dense · open weights** Qwen team open weights (Hugging Face `Qwen/Qwen3.6-27B`); coding ability rivals 397B-class models, single-GPU friendly. **35B-A3B · open MoE** Qwen team open weights (Hugging Face `Qwen/Qwen3.6-35B-A3B`); same lineage as closed-source Flash, 3B active for extremely low compute cost. **Weights public, compute on us** Just call the API — no GPU rental, deploy, monitoring, or version-upgrade pipeline. Billed per token, scales on demand. **Simple, predictable budget** Open-weight tiers bill at a single flat rate — no need to estimate P95 input tokens. Monthly cost is a straight linear function of token volume. ### Performance & positioning | Dimension | qwen3.6-27b | qwen3.6-35b-a3b | | -------------------------- | ------------------------------------------------------------- | ----------------------------------------------------------------- | | Architecture | 27B dense | MoE 35B total / 3B active | | License | Qwen team open-weight | Qwen team open-weight | | Hugging Face repo | `Qwen/Qwen3.6-27B` | `Qwen/Qwen3.6-35B-A3B` | | Coding capability | Rivals 397B-class (small-size coding leader) | Same lineage as closed-source Flash | | Modalities | Text | Text | | Recommended scenarios | Cost-sensitive coding aid, compliance audit, local-deploy PoC | High-volume low-cost dialog, transition phase before self-hosting | | Single-GPU self-host floor | A100 40G+ | VRAM scales by 35B total params; compute by 3B active | **Why a small-size coding model matters** — for IDE inline completion, PR review, code search, and other **low-latency / high-frequency** workflows, 27B beats the closed-source flagship: faster responses, lower cost, and code understanding that approaches the larger models. ### Specs * **Model IDs**: `qwen3.6-27b` / `qwen3.6-35b-a3b` * **Endpoint**: `POST /v1/chat/completions` (OpenAI Chat Completions compatible) * **Input modalities**: Text * **Streaming**: ✅ Supported * **Function calling / tool use**: ✅ Supported * **Billing**: Pay-as-you-go - Chat (flat, **no tiers**) * **Channel**: APIYI hosted relay * **Context / max output**: matches the official weight cards on Hugging Face ## Use cases ### Recommended scenarios Use `qwen3.6-27b` for IDE inline completion, PR Bots, commit summaries — 27B-scale latency is fast, unit price is far below the Max-Preview flagship. Use `qwen3.6-35b-a3b` for customer support, bulk translation, content moderation — 3B active means extremely low compute cost, just \$0.26/\$1.56 per 1M tokens. For customers who need auditable / filing-ready weights: pull the public Hugging Face checkpoints for internal compliance review while running production traffic on APIYI's hosted relay. Before evaluating whether open Qwen3.6 fits your stack, run the PoC on APIYI to validate quality and quantify token cost — then decide whether to stand up your own GPU cluster. ### Code example ```python theme={null} import openai client = openai.OpenAI( api_key="sk-your-api-key", base_url="https://api.apiyi.com/v1" ) # 27B coding powerhouse: low-latency code completion resp = client.chat.completions.create( model="qwen3.6-27b", messages=[ {"role": "system", "content": "You are an inline code completion assistant. Return only the code diff, no commentary."}, {"role": "user", "content": "Complete this function: def merge_intervals(intervals): ..."} ], stream=True ) for chunk in resp: if chunk.choices[0].delta.content: print(chunk.choices[0].delta.content, end="", flush=True) # 35B-A3B speed MoE: high-concurrency dialog resp = client.chat.completions.create( model="qwen3.6-35b-a3b", messages=[ {"role": "user", "content": "Explain quantum entanglement in one sentence."} ] ) print(resp.choices[0].message.content) ``` ### Best practices * **Try the small models first**: for cost-sensitive workloads, start with 27B / 35B-A3B. If quality holds, stop here. Only escalate to closed-source Plus / Max-Preview when needed — avoid paying flagship rates by default. * **Predictable budgeting with flat pricing**: the open-weight tiers don't have tiers. Estimate monthly cost as a straight linear function: `input_tokens × input_rate + output_tokens × output_rate`. No need to model P95 token-count distribution. * **Smooth path to self-hosting**: APIYI's hosted SDK and self-hosted vLLM both speak OpenAI-compatible. When the time comes, point `base_url` to your internal gateway — application code stays untouched. ## Pricing & availability ### Listed prices | Model | Billing | Input | Output | | ----------------- | ------------------------------------- | -------------------- | -------------------- | | `qwen3.6-27b` | Pay-as-you-go - Chat (flat, no tiers) | \$0.4200 / 1M tokens | \$2.5200 / 1M tokens | | `qwen3.6-35b-a3b` | Pay-as-you-go - Chat (flat, no tiers) | \$0.2600 / 1M tokens | \$1.5600 / 1M tokens | **Flat vs. tiered**: closed-source Max-Preview / Flash / Plus use tiered pricing keyed on input-token count per request; open-weight 27b / 35b-a3b use a **single flat rate** — easier to budget. List prices match Alibaba Cloud's official rate; APIYI's recharge bonus brings the effective unit price to roughly **85% of list**. ### Stack with APIYI's recharge promo Recharge promotion details: [/en/faq/recharge-promotions](/en/faq/recharge-promotions) Effective unit prices after the recharge bonus (\~15% off): | Model | Effective input | Effective output | | ----------------- | --------------- | ---------------- | | `qwen3.6-27b` | ≈ \$0.357 / 1M | ≈ \$2.142 / 1M | | `qwen3.6-35b-a3b` | ≈ \$0.221 / 1M | ≈ \$1.326 / 1M | ## Wrap-up The Qwen3.6 open-weight duo stitches together two things that have historically been mutually exclusive: **controllable open weights** and **no GPU rental / no ops burden**. * **27B dense** — coding powerhouse: low latency, high frequency, controlled cost; ideal for IDE-embedded scenarios * **35B-A3B open MoE** — same lineage as closed-source Flash, 3B active means extreme speed; ideal for high-concurrency dispatch **Strategy**: start cost-sensitive workloads on the open-weight tiers (27b / 35b-a3b); validate quality on the API, then decide whether to upgrade to closed-source production. With the recharge bonus, the open 35B-A3B input lands around \$0.22 / 1M tokens — currently the cheapest open-source option on the Aliyun official-relay channel. Sources: Qwen team Hugging Face model repositories (`Qwen/Qwen3.6-27B`, `Qwen/Qwen3.6-35B-A3B`). This launch is the APIYI hosted relay distribution; weights match the official release. Article data retrieval date: 2026-04-28 (UTC+8). ## Related reading * [Qwen3.6 text model series overview](/en/api-capabilities/qwen-3-6/overview) - All 5 models, pricing, and routing strategy * [Qwen3.6 Max-Preview & Flash live](/en/news/qwen-3-6-max-flash-launch) - Closed-source production tier deep dive * [Qwen3.6-Plus launch — Alibaba's strongest coding agent](/en/news/qwen-3-6-plus-launch) - Plus deep dive