SYSTEM Cited by 3 sources

Qwen

Qwen is the open-weight LLM family from Alibaba Cloud — first released 2023, with subsequent Qwen 1.5 / Qwen 2 / Qwen 2.5 / Qwen 3 iterations spanning model sizes from 0.5B to 72B+ parameters. Variants include base, instruction-tuned, long- context, multi-modal (Qwen-VL), and code-specialised (Qwen-Coder) releases.

Qwen is one of the canonical open-weight base-model families available for fine-tuning production LLMs via LoRA adapters or full-parameter SFT — alongside Llama family as the two most-commonly-named options in Western production retrospectives.

Why it shows up in production retrospectives

• Fine-tune-friendly. Base + instruct variants cover both supervised fine-tuning and instruction-following deployment shapes.
• Range of sizes. Qwen ships from 0.5B (edge / single-GPU serving) up to 72B+ (premium quality tier) — production teams pick the sweet spot for their latency + cost budget.
• Strong English + Chinese + multilingual. Particularly strong at zero-shot Chinese; production choice for companies serving CJK markets.
• LoRA / QLoRA ecosystem. Broad support across the open- source LoRA toolchain (peft / axolotl / unsloth).
• Open-source license. Permits commercial deployment, which Llama's license restricts at scale historically (Llama 3.1 / 3.2 eased this).

Seen in

• sources/2026-02-26-instacart-our-early-journey-to-transform-discovery-recommendations-with-llms — Instacart's generative recommendations platform Phase-2 retrieval-keyword generation runs a teacher-student fine-tune where the student base model is selected via ablations across the Llama and Qwen families. Specific Qwen version + rank + data augmentation winner not disclosed. First wiki disclosure of Qwen as a production base- model choice at Instacart.
• 2026-09-04 Databricks — Achieving Extreme Efficiency through Specialized GPU Kernel Generation — Qwen 3.5 122B is the target model whose inference kernels Databricks' Proteus specialized per runtime shape (1.8–5.2× faster than vLLM on NVIDIA B200). The case study targets Qwen 3.5 122B's Gated DeltaNet (linear-attention) path — the packed-decode kernel. First wiki disclosure of a Qwen 3.5 / 122B-scale variant and of Qwen using a Gated-DeltaNet linear-attention block. (sources/2026-09-04-databricks-achieving-extreme-efficiency-through-specialized-gpu-kernel-generation)
• sources/2026-10-01-cloudflare-introducing-clef-our-open-source-decision-models-and-new-rl — Qwen as the frozen backbone of a Cloudflare decision model. Clef freezes Qwen3.8-27B and Clef-flash freezes Qwen3.5-9B, jointly optimizing a routing head + rank-256 LoRA adapters on top. During inference Clef runs Qwen for a prefill-only pass then scores schema choices in parallel (non-autoregressive). First wiki instance of Qwen used as the base for a specialized, non-autoregressive decision model rather than a text-generating fine-tune. (Qwen3.8-27B / Qwen3.5-9B are the post's version strings.)

Caveats

• Instacart's post names Qwen at the family level, not a specific Qwen version or size.
• No production accuracy / latency / cost numbers for the Qwen variant disclosed.
• Relationship to the sibling Intent Engine (which used Llama-3-8B) not discussed — different task, different base-model ablation set.

Related

• systems/llama-3-1 — sibling open-weight family, jointly ablated at Instacart.
• systems/llama-3 — used by sibling Instacart system (Intent Engine) as its fine-tuned student base.
• systems/instacart-generative-recommendations-platform — canonical consumer.
• systems/clef — Cloudflare's decision model that freezes Qwen3.8-27B / Qwen3.5-9B as its backbone + rank-256 LoRA adapters.
• concepts/lora-low-rank-adaptation — the fine-tuning mechanism named in the ablation.
• patterns/teacher-student-model-compression — the shape the Qwen student realises.
• companies/instacart
• systems/proteus — Databricks' harness that specialized Qwen 3.5 122B inference kernels.
• systems/vllm — the baseline Proteus's Qwen kernels beat.