Status: candidate sources for planning Checked: 17 August 2026
This is not yet the immutable release manifest. It identifies suitable primary artifacts for curriculum design. Before a lesson enters implementation, its source row must be expanded with:
- exact paper and report revision;
- immutable repository commit;
- exact model-card revision;
- separate code and weight licences;
- base-weight, training-code, and data availability;
- evaluation protocol and required hardware;
- claim labels used by the lesson.
The commit recorded below is the official repository head observed during planning. It is evidence of what was reviewed, not a promise to track that branch.
This register feeds several independent courses. It does not make a source eligible everywhere: each course must copy its selected rows into an immutable course-local manifest and complete the selection gates before implementation.
| Case study | Candidate teaching use | Official repository and reviewed head | Repository licence signal | Release decision |
|---|---|---|---|---|
| DeepSeek V3 | MLA, sparse MoE, FP8 training, DualPipe | deepseek-ai/DeepSeek-V3 at 9b4e9788e4a3a731f7567338ed15d3ec549ce03b |
GitHub reports MIT for the repository | Candidate; verify report, weights, and evaluation terms separately |
| DeepSeek R1 | Verifiable rewards and reasoning post-training | deepseek-ai/DeepSeek-R1 at 0cf78561f1d51c84a21b2190626b21116d5c68bb |
GitHub reports MIT for the repository | Candidate; distinguish published recipe from reproducible training stack |
| DeepSeek V4 Pro and Flash | CSA/HCA hybrid attention, mHC, sparse-expert routing bootstrap, Muon, mixed FP4/FP8, domain-expert consolidation, and million-token evaluation | Technical report arXiv:2606.19348v1; official model artifacts: Pro, Pro Base, Flash, and Flash Base at the reviewed commits |
Each reviewed artifact contains an MIT LICENSE; post-trained cards also declare MIT |
Required current endpoint for the DeepSeek lineage; freeze report, artifact, inference-code, and evaluation revisions separately before implementation |
| Kimi K2 | Sparse MoE and MuonClip | MoonshotAI/Kimi-K2 at 1b4022bbb7187cf4011a8bdf0b4cd10e2daa26c4 |
GitHub cannot assert a standard repository SPDX licence | Candidate; licence and weight terms require manual review |
| Kimi Linear | Delta/linear attention | MoonshotAI/Kimi-Linear at 8c1d85eb6b5f8fcefb15758691b0ce50b0827ce3 |
GitHub reports MIT for the repository | Candidate for the efficient-attention lesson |
| Kimi K3 | KDA, gated MLA, Stable LatentMoE, AttnRes, SiTU-GLU, native quantisation, and long-context architecture | MoonshotAI/Kimi-K3 at 3cb39dfd32e51c3328e2e4b4af21341247d06c43; technical report |
Repository and weights use the custom Kimi K3 License; review its commercial-service and attribution conditions before reuse | Required endpoint for the Kimi architecture lineage; freeze the report, code, weights, and evaluation protocols separately |
| MiniMax M1 | Hybrid attention and long-context reasoning | MiniMax-AI/MiniMax-M1 at 2abb4f45a9df4154b4bde024d51874bd127edcee |
GitHub reports Apache-2.0 for the repository | Candidate; verify model-weight terms separately |
| Qwen 3 | Dense/MoE comparison and hybrid reasoning | QwenLM/Qwen3 at 7a2f61ffc7a20d47efcd2bf97f6f2bf52729042e |
GitHub does not report a repository SPDX identifier | Candidate; select exact model size and model-card licence before implementation |
| GLM 4.5 | MoE, agentic models, and hybrid reasoning | zai-org/GLM-4.5 at 170f20b2c10659008fdbc909d478bc2a75bc3627 |
GitHub reports Apache-2.0 for the repository | Retained as historical lineage; no longer the endpoint for the first GLM course |
| GLM 5 and GLM 5.2 | Sparse MoE, DSA, cross-layer index reuse, MTP/speculative decoding, reasoning effort, long-context and agentic evaluation | zai-org/GLM-5 at 25206af860c4ac10f6411c597c574f9b1c00e53c; GLM 5 report arXiv:2602.15763v2; IndexCache arXiv:2603.12201v1; official GLM 5.2 BF16 and FP8 artifacts at the reviewed commits |
GLM-5 code repository is Apache-2.0; both reviewed GLM 5.2 weight artifacts declare MIT | Required current endpoint for the GLM course; freeze the GLM 5 report revision, 5.2 release page, cards/configs, serving assumptions, and evaluation protocols separately |
| Course | Candidate sources | Required freeze decision |
|---|---|---|
| Open Frontier Models: Shared Techniques | Small configuration and artifact excerpts from several rows | Select only examples whose reuse terms support the shared teaching fixtures |
| Inside DeepSeek | DeepSeek V4 Pro/Flash, with V3 and R1 as lineage | Pin exact report/card revisions and separate architecture, optimisation, post-training, reasoning, code, weight, and harness claims |
| Inside Kimi | Kimi K2, Kimi Linear, and Kimi K3 | Resolve K2 terms; pin the Kimi Linear and K3 attention sources; review the K3 custom licence; distinguish architecture, multimodal, quantisation, and benchmark claims |
| Inside Qwen | Qwen 3 | Select exact dense and MoE sizes plus base/instruction cards and licences |
| Inside MiniMax | MiniMax M1 | Pin the hybrid-attention report, code revision, weights, and long-context protocol |
| Inside GLM | GLM 5.2, with GLM 5 as lineage | Pin BF16/FP8 artifacts, distinguish the release's IndexShare name from the linked IndexCache paper, and pin serving and agentic evaluation protocols |
The first DeepSeek course now ends at V4 rather than V3/R1. V3 and R1 remain necessary lineage sources: V4 changes the attention and residual architecture while building on sparse-expert and reasoning work that those releases document. The course must compare Pro and Flash as two sizes in one V4 family, not treat their benchmark difference as a controlled architecture ablation.
The first GLM course now ends at GLM 5.2 rather than GLM 4.5. GLM 5 supplies the available architecture and post-training report; GLM 5.2 supplies the current configuration, weights, IndexShare deployment pattern, MTP changes, and long-horizon evaluation claims. Z.ai links the name IndexShare to the paper titled IndexCache: Accelerating Sparse Attention via Cross-Layer Index Reuse. Course prose must preserve both names and their scopes rather than silently treating them as independently documented mechanisms.
These are candidate-source updates, not immutable course-local manifests. The reviewed commits make the planning decision auditable, but each course still must complete its source-freeze and licence gates before technical lessons are implemented.
A model becomes a course case study only when:
- its public artifact demonstrates a distinct, teachable mechanism;
- an official technical report, repository, or model card documents the mechanism;
- a small deterministic implementation can expose a meaningful trade-off;
- the course can state the simplification without implying full-model reproduction;
- code, prose, figures, and weights used by the course have compatible licences;
- the lesson remains useful after the named model is no longer current.
Very recent model releases should not replace a stable source merely because they report a higher benchmark score. They enter only when they add a mechanism that improves the educational comparison.