"What is Qwen3's role in this course?"	"The hybrid thinking/non-thinking pipeline reference — the other canonical reasoning model alongside DeepSeek-R1. It fuses a thinking (reasoning) mode and a non-thinking (fast) mode into one set of weights, controlled by a thinking budget. It is the production workhorse most deployed teams run."	c3::ftdd08::recall
"What production problem does Qwen3's hybrid design solve?"	"A dedicated reasoner (like R1) always deliberates, which is a latency tax on simple queries. A fast assistant cannot deliberate on hard ones. The naive fix is two models + a router (2x weights, router misroutes). Qwen3 fuses both modes into one model, switched by a thinking budget — one deployment, no router."	c3::ftdd08::recall
"What is thinking-mode fusion (Stage 3)?"	"The novel stage: merge a thinking (long-CoT) model and a non-thinking (fast, direct) model into a SINGLE set of weights. Construct a dataset with both thinking and non-thinking responses, fine-tune one model on the combined data. A control token or budget selects which mode the model expresses. This is what R1 does NOT do."	c3::ftdd08::recall
"What is the key engineering insight that makes fusion possible?"	"The thinking pathways and the direct-response pathways are NOT mutually exclusive in the weights — they can coexist, and a control token or budget parameter selects which the model expresses. After fusion, the model has BOTH capabilities in one set of parameters."	c3::ftdd08::analysis
"Name the four stages of the Qwen3 post-training pipeline in order."	"(1) Long-CoT cold start — seed reasoning format (echoes R1 Stage 1). (2) Reasoning RL (GRPO-style) — verifiable rewards, scales reasoning (echoes R1 Stage 2). (3) Thinking-mode fusion — merge thinking + non-thinking (the NOVEL stage). (4) General RL — full distribution, keeps non-thinking strong."	c3::ftdd08::recall
"How do Qwen3's first two stages relate to R1's pipeline?"	"They are structurally the same. Stage 1 (long-CoT cold start) anchors the reasoning format just like R1's cold-start SFT. Stage 2 (GRPO-style RL with verifiable rewards) scales reasoning just like R1's reasoning RL. The DIVERGENCE is Stage 3: Qwen3 fuses modes, R1 ships a pure reasoner."	c3::ftdd08::analysis
"What is the thinking budget, and what does it control?"	"A single inference-time parameter controlling how many tokens the model may spend in its <think> deliberation block. Low budget = short/empty think, fast direct answer (chatbot speed). High budget = long self-correcting chain, verified answer (reasoner quality). Same weights serve both."	c3::ftdd08::recall
"Why does the thinking budget matter operationally?"	"Without it, a hybrid model still needs a per-query router to decide thinking vs non-thinking — and the router is itself a failure mode. The budget collapses routing into a single continuous parameter. Set a default for your workload (low for chatbot, high for math-solver); the model spends within it. No classifier needed."	c3::ftdd08::application
"For a latency-sensitive chatbot serving mostly simple queries, what thinking budget should you set by default, and why?"	"LOW. Simple queries ('capital of France', 'summarize this email') need fast direct answers, not long deliberation. A high default would impose a latency tax with no quality gain. The model adapts within the budget. Raise the budget per query only where the question warrants deliberation."	c3::ftdd08::application
"For a batch math-solver processing competition problems, what thinking budget should you set, and why?"	"HIGH. Competition math rewards deliberation — self-correction, verification, multi-step reasoning. A low budget would cut off the chain that produces correct answers. Latency is less critical in batch. The model needs the token headroom to reason. Set the default high."	c3::ftdd08::application
"Why is the thinking budget NOT 'more is always better'?"	"It is a latency-quality DIAL. A model that always thinks at max budget is as slow as a dedicated reasoner, wasting tokens on trivial queries with no quality gain. The discipline is choosing the budget that matches your latency SLA and accuracy target, and varying it per query where it matters."	c3::ftdd08::analysis
"What pretraining scale underpins Qwen3, and why does it matter for the hybrid thesis?"	"36T+ pretraining tokens. Family from 0.6B dense to 235B-A22B MoE. It matters because the NON-THINKING mode inherits this strength. A fused model is only a good hybrid if non-thinking is genuinely capable (not a degraded reasoner). The pretraining scale guarantees the non-thinking mode stays strong."	c3::ftdd08::analysis
"Why is treating fusion as 'two models glued together' an anti-pattern?"	"Fusion is ONE set of weights with two coexisting modes, selected by a control signal. Building a router to decide thinking vs non-thinking RE-CREATES the very problem fusion was designed to eliminate (two models + a router). Trust the single model; use the budget to control mode."	c3::ftdd08::analysis
"Why must you evaluate a hybrid model in BOTH modes?"	"Stage 4 (general RL) exists precisely to keep the non-thinking mode strong. If you evaluate only thinking mode, you will miss degradation in non-thinking and misjudge the model for production. Evaluate both modes against their respective benchmarks — reasoning benchmarks for thinking, general-assistant benchmarks for non-thinking."	c3::ftdd08::application
"How do R1 and Qwen3 differ as reasoning references?"	"R1 = distillation reference: proved reasoning emerges from RL and transfers via SFT-only. Study when building a student by distilling from a teacher. Qwen3 = hybrid reference: proved you can fuse modes and control with a budget. Study when deploying one model that serves chatbot + reasoner. Complementary, not competing."	c3::ftdd08::analysis
"What does 'adaptive compute' mean in the Qwen3 context?"	"Spending inference compute (tokens/time) proportional to query difficulty. The thinking budget is the mechanism: simple queries get few/no thinking tokens (fast), hard queries get many (deliberate). One model adapts its compute to the question, rather than a fixed-cost response for all queries."	c3::ftdd08::recall
"What is the 235B-A22B MoE, and what does A22B signify?"	"Qwen3's flagship model: 235B total parameters, Mixture-of-Experts architecture, with 22B parameters ACTIVE per token (A22B = Active 22B). This gives high capability (235B of stored knowledge) at lower per-token compute than a dense 235B (only 22B active per forward pass)."	c3::ftdd08::recall
"Where is Qwen3 documented, and when was it released?"	"The Qwen3 Technical Report, arXiv:2505.09388, released May 2025. Documents the four-stage pipeline, thinking-mode fusion, thinking budget, 36T+ token pretraining, and the model family (0.6B–235B-A22B MoE)."	c3::ftdd08::recall
"A team deployed Qwen3 for a general chatbot but set the thinking budget to max. Users complain it's slow on simple questions. What is wrong and how do you fix it?"	"Max budget means every query — including trivial ones — runs a long thinking chain, imposing latency tax for no quality gain. Fix: set a LOW default budget for the chatbot workload. The model will answer simple queries directly and fast. Raise the budget per query only when a question warrants deliberation."	c3::ftdd08::application
"How do the strongest teams use R1 and Qwen3 together?"	"They know both and borrow from both. Qwen3's own distillation work drew on R1's recipe (teacher traces + SFT). The field's thinking-model conventions now borrow from Qwen3's budget mechanism. The two references are complementary halves of the open reasoning story: R1 for the science of emergence/transfer, Qwen3 for the engineering of fusion/adaptive compute."	c3::ftdd08::analysis