Status: accepted concept; refreshed for GLM 5.2; source freeze required before implementation Prerequisites: AI from First Principles v0.2 and Open Frontier Models: Shared Techniques v0.1 Candidate release: GLM 5.2, with GLM 5 retained as its architecture and post-training lineage source
Reconstruct selected sparse-model, sparse-attention, speculative-decoding, reasoning-control, and agentic-evaluation techniques from pinned GLM 5/5.2 artifacts while separating model behavior from the serving, tool, and environment scaffolding around it.
After completing the course, a learner can:
- reconcile the GLM 5 report with pinned GLM 5.2 BF16/FP8 artifacts, configuration, licence, and release claims;
- implement and inspect a toy sparse block under the shared accounting model;
- implement a toy DSA indexer and compare per-layer selection with cross-layer top-k reuse;
- model multi-token prediction acceptance length and distinguish model changes from speculative-decoding policy;
- represent reasoning mode and inference budget as explicit experimental variables;
- model tool calls, observations, invalid actions, and termination as a constrained trajectory;
- distinguish model error, tool-schema error, environment error, and evaluator error;
- design a long-horizon evaluation without crediting agent scaffolding to the model.
- Reading GLM 5.2 — GLM 5/5.1/5.2 lineage, BF16/FP8 boundaries, configuration, licences, sparse anatomy, reported long-context and agentic claims, and missing evidence.
- Sparse execution — routed experts, active parameters, capacity, utilisation, and deterministic failure analysis.
- Sparse attention and index reuse — DSA selection, indexer cost, cross-layer similarity, Full versus Shared layers, the GLM 5.2 four-layer reuse pattern, and quality/throughput trade-offs. The release calls this deployed pattern
IndexShare; its linked research paper names the general methodIndexCache. - Multi-token prediction and serving — draft steps, acceptance length, index/KV reuse, rejection sampling, latency, and correctness boundaries.
- Reasoning controls and long horizons — effort levels, budgets, stopping, compaction, outcome/process signals, and evaluation units.
- Tool-use trajectories — schemas, actions, observations, invalid states, recovery, termination, and sandbox boundaries.
- Agentic evaluation — model-versus-scaffold-versus-serving ablations, context-management disclosure, long-horizon success, partial credit, variance, and claim limits.
| Lesson | Explorable | Focused implementation |
|---|---|---|
| Release anatomy | GLM 5/5.2 evidence and parameter map | Normalise BF16/FP8 artifacts |
| Sparse execution | Router, capacity, and utilisation trace | Implement deterministic routing |
| Sparse attention | Per-layer versus shared top-k trace | Implement bounded index reuse |
| Speculative serving | Draft/accept/reject timeline | Implement lossless token acceptance |
| Reasoning controls | Effort/budget/compaction state | Enforce a bounded reasoning loop |
| Tool trajectories | Model-tool-environment state machine | Validate constrained tool calls |
| Agentic evaluation | Model/scaffold/serving ablation and failure taxonomy | Score deterministic trajectories |
Extend the foundation model with a toy sparse block, sparse indexer with cross-layer reuse, and multi-token draft path, then place it inside a bounded tool environment. Compare model-only, serving-only, scaffold-only, and combined changes on deterministic trajectories. Attribute improvements and failures to the component actually changed.
The course does not teach generic agent construction or production serving as product workflows. It uses bounded reconstructions and a constrained environment to study pinned GLM claims and evaluation methods. Shared artifact, accounting, and reproduction skills remain in the frontier core. GLM 5.2 benchmark values remain reported claims until a matching harness and repeated trials reproduce them.
- Exact GLM 5 report and IndexCache paper revisions, GLM 5 repository commit, GLM 5.2 BF16/FP8 model-card commits, configurations, licences, and evaluation protocols are frozen.
- Sparse routing, top-k index reuse, speculative acceptance, and tool trajectories are deterministic and independently tested.
- Every loop has an explicit action schema, budget, invalid-action path, and termination condition.
- Agentic results include model/serving/scaffold ablations, context-management disclosure, and failure attribution.
- The course passes locally without production weights, accounts, APIs, GPUs, or network services.