Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
fix(registry): the GLM family names both models it loads
JustVugg3a9a3da54b80
diff --git a/README.it.md b/README.it.mdindex 9a22b719..ed81a6c3 100644--- a/README.it.md+++ b/README.it.md@@ -12,7 +12,7 @@ miliardi a 2,8 mila miliardi di parametri** — su hardware consumer ed eterogenin C puro e senza dipendenze del motore, trattando storage, RAM e VRAM comeun'unica gerarchia di inferenza.-Oggi girano otto famiglie: **GLM-5.2** (744B), **GLM-5.3-Flash** (321B, con+Oggi girano otto famiglie: **GLM-5.2/5.3** (744B), **GLM-5.3-Flash** (321B, convision), **Inkling** (975B), **Kimi K3** (2,8T), **DeepSeek V4 Flash** (284B),**Qwen3.8-Flash-Next** (125B + 51B n-gram), **Qwen3.6** (35B-A3B) e**OLMoE** (7B) — undiff --git a/README.md b/README.mdindex 8df8c496..126c5c6e 100644--- a/README.md+++ b/README.md@@ -18,7 +18,7 @@ parameters** — on consumer and heterogeneous hardware, in pure C with zeroengine dependencies, by treating storage, RAM, and VRAM as a single inferencehierarchy (AI memory multitiering).-Eight families run today: **GLM-5.2** (744B), **GLM-5.3-Flash** (321B, with+Eight families run today: **GLM-5.2/5.3** (744B), **GLM-5.3-Flash** (321B, withvision), **Inkling** (975B), **Kimi K3** (2.8T), **DeepSeek V4 Flash** (284B),**Qwen3.8-Flash-Next** (125B + 51B n-gram), **Qwen3.6** (35B-A3B) and**OLMoE** (7B) —@@ -405,7 +405,7 @@ the model's `config.json`):> | Model | Disk for the weights | RAM | GPU |> |---|---|---|---|> | **OLMoE** | ~7 GB (int8 container) | 8 GB | not needed |-> | **GLM-5.2** | ~372 GB | 16 GB min, 24 GB comfortable | not needed |+> | **GLM-5.2/5.3** | ~372 GB | 16 GB min, 24 GB comfortable | not needed |> | **GLM-5.3-Flash** | ~195 GB converted | 25 GB (12 GB weights at int4 + expert cache) | not needed |> | **Inkling** | ~469 GB | 25 GB with the int4 dense container, ~120 GB without | not needed |> | **Kimi K3** | ~1.6 TB | 32 GB+ | not needed |@@ -419,7 +419,7 @@ the model's `config.json`):| Family | Total / active | Weights | Build | Docs ||---|---|---|---|---|-| **GLM-5.2** | 744B / 40B | [`mastouri/…-int4-g64-with-int8-mtp`](https://huggingface.co/mastouri/GLM-5.2-colibri-int4-g64-with-int8-mtp) (372 GB) | `make -C c glm` | this page |+| **GLM-5.2/5.3** | 744B / 40B | [`mastouri/…-int4-g64-with-int8-mtp`](https://huggingface.co/mastouri/GLM-5.2-colibri-int4-g64-with-int8-mtp) (372 GB) | `make -C c glm` | this page || **Inkling** (Thinking Machines) | 975B / 41B | [`nbeerbower/Inkling-colibri-int4`](https://huggingface.co/nbeerbower/Inkling-colibri-int4) (469 GB) | `make -C c inkling` | [inkling.md](docs/inkling.md) || **GLM-5.3-Flash** (Z.ai) | 321B / 40B | [`zai-org/GLM-5.3-Flash`](https://huggingface.co/zai-org/GLM-5.3-Flash) — converted to **int4-gs64** routed experts, dense stays BF16 and the precision is a load-time choice; vision included | `make -C c glm53` | [glm53-flash.md](docs/glm53-flash.md) || **Kimi K3** (Moonshot) | 2.8T / 104B | [`moonshotai/Kimi-K3`](https://huggingface.co/moonshotai/Kimi-K3) — original checkpoint, routed experts stay **native MXFP4** | `make -C c kimi_k3` | [kimi_k3.md](docs/kimi_k3.md) |diff --git a/README.zh-CN.md b/README.zh-CN.mdindex 66edb96d..bb075688 100644--- a/README.zh-CN.md+++ b/README.zh-CN.md@@ -10,7 +10,7 @@**小巧引擎,庞大模型。**在消费级与异构硬件上运行**前沿 MoE 模型——从 744B 到2.8T 参数**——以引擎零依赖的纯 C 实现,将存储、RAM 与 VRAM 视为统一的推理层级。-目前可运行八个模型家族:**GLM-5.2**(744B)、**GLM-5.3-Flash**(321B,含视觉)、+目前可运行八个模型家族:**GLM-5.2/5.3**(744B)、**GLM-5.3-Flash**(321B,含视觉)、**Inkling**(975B)、**Kimi K3**(2.8T)、**DeepSeek V4 Flash**(284B)、**Qwen3.8-Flash-Next**(125B + 51B n-gram)、**Qwen3.6**(35B-A3B)与 **OLMoE**(7B)——各自一个 C 文件,共用同一套 `coli chat` / `coli serve` / `coli web` 前端。[完整列表](README.md#other-supported-models)diff --git a/README.zh-TW.md b/README.zh-TW.mdindex 3c1a9363..35406f33 100644--- a/README.zh-TW.md+++ b/README.zh-TW.md@@ -10,7 +10,7 @@**小巧引擎,龐大模型。**在消費級與異質硬體上執行**前沿 MoE 模型——從 744B 到2.8T 參數**——以引擎零相依套件的純 C 實作,將儲存、RAM 與 VRAM 視為統一的推論階層。-目前可執行八個模型家族:**GLM-5.2**(744B)、**GLM-5.3-Flash**(321B,含視覺)、+目前可執行八個模型家族:**GLM-5.2/5.3**(744B)、**GLM-5.3-Flash**(321B,含視覺)、**Inkling**(975B)、**Kimi K3**(2.8T)、**DeepSeek V4 Flash**(284B)、**Qwen3.8-Flash-Next**(125B + 51B n-gram)、**Qwen3.6**(35B-A3B)與 **OLMoE**(7B)——各自一個 C 檔案,共用同一套 `coli chat` / `coli serve` / `coli web` 前端。[完整清單](README.md#other-supported-models)diff --git a/c/family_registry.py b/c/family_registry.pyindex 7ab26a48..aa4cad62 100644--- a/c/family_registry.py+++ b/c/family_registry.py@@ -928,7 +928,23 @@ FAMILIES = (FamilyDescriptor(id="glm",model_types=("glm_moe_dsa", "glm5_moe", "glm"),- display_name="GLM-5.2",+ # Two model names, one family, and that is not a shortcut. Z.ai state+ # it on GLM-5.3's own card: "GLM-5.3 uses the same base model as+ # GLM-5.2 -- every gain comes from post-training." The checkpoints+ # agree. Diffing the two config.json leaves exactly one extra key+ # (moe_router_dtype) and the transformers_version that wrote the file:+ # same 78 layers, 256 experts, hidden 6144, moe_intermediate 2048, same+ # parameter count. There is nothing architectural to tell them apart,+ # so no rule over the configuration can name one and not the other, now+ # or later. #1326's approach for Qwen -- name a checkpoint by its+ # geometry -- cannot apply here, because the geometry is identical.+ #+ # Announcing a GLM-5.3 container as "GLM-5.2" was wrong in the one way+ # that matters: the engine ran the right weights and the banner named a+ # different model. Naming both is the honest statement of what this+ # family loads. If a future release ever adds a real discriminator,+ # split this then, on evidence.+ display_name="GLM-5.2/5.3",display_scale="744B",engine_artifact="colibri",engine_aliases=("glm",),diff --git a/c/tests/test_family_registry.py b/c/tests/test_family_registry.pyindex 782dcdaa..836bde90 100644--- a/c/tests/test_family_registry.py+++ b/c/tests/test_family_registry.py@@ -1412,5 +1412,59 @@ class FamilyRegistryTest(unittest.TestCase):"l'archivio non spedirebbe family_registry.py")+class GlmFamilyNamesBothModelsTest(unittest.TestCase):+ """#1365 follow-up: a GLM-5.3 container was announced as "GLM-5.2 744B".++ The engine loaded the right weights; only the banner named a different+ model. The fix is the name, and the reason the name covers two models+ rather than choosing between them is worth pinning here, because the+ obvious "improvement" later is to add detection.++ There is nothing to detect. Z.ai say so on GLM-5.3's own card -- "GLM-5.3+ uses the same base model as GLM-5.2; every gain comes from post-training"+ -- and the checkpoints agree: diffing the two real config.json leaves one+ extra key and the transformers_version that wrote the file. If a release+ ever ships a genuine discriminator, split the family then, on evidence,+ and this test is where to record it.+ """++ #: The two configs, reduced to what the registry reads plus the only two+ #: keys that actually differ between the real files.+ BASE = {+ "model_type": "glm_moe_dsa",+ "num_hidden_layers": 78,+ "n_routed_experts": 256,+ "hidden_size": 6144,+ "moe_intermediate_size": 2048,+ }++ def test_both_checkpoints_resolve_to_the_same_family(self):+ v52 = dict(self.BASE, transformers_version="5.12.0")+ v53 = dict(self.BASE, transformers_version="5.15.0",+ moe_router_dtype="float32")+ self.assertEqual(family_for_config(v52).id, family_for_config(v53).id)+ self.assertEqual(family_for_config(v53).engine_artifact, "colibri")++ def test_the_only_differences_are_not_discriminators(self):+ """Neither key can carry the decision, so neither may be read as if it+ could. `transformers_version` records the library that wrote the file+ and changes when anyone re-exports; `moe_router_dtype` is a precision+ hint that a re-export of GLM-5.2 could equally carry."""+ source = Path(__file__).resolve().parents[1] / "family_registry.py"+ text = source.read_text(encoding="utf-8")+ for key in ("transformers_version", "moe_router_dtype"):+ self.assertNotIn(f'"{key}"', text,+ f"{key} is not a version discriminator; naming a "+ f"checkpoint from it would be a guess wearing a "+ f"rule's clothes")++ def test_the_name_states_both_models(self):+ family = family_for_config(dict(self.BASE))+ for model in ("5.2", "5.3"):+ self.assertIn(model, family.display_name,+ "a user running one of the two must not read the "+ "name of the other")++if __name__ == "__main__":unittest.main()diff --git a/site/index.html b/site/index.htmlindex 74d3f331..7172cd29 100644--- a/site/index.html+++ b/site/index.html@@ -263,7 +263,7 @@ footer li a:hover{color:var(--on-dark)}<div class="overline">The models</div><h2 class="sec-h">Eight families run today, one engine each.</h2><div class="cards">- <div class="card"><h3>GLM-5.2</h3><p class="desc">The flagship target: an int4 container, token-exact versus reference, with the full expert atlas published.</p>+ <div class="card"><h3>GLM-5.2/5.3</h3><p class="desc">The flagship target: an int4 container, token-exact versus reference, with the full expert atlas published.</p><div class="tags"><span class="tag">reasoning</span><span class="tag">tools</span></div><div class="meta"><div class="row"><span class="k">Maker</span><span class="v">Z.ai</span></div><div class="row"><span class="k">Size · RAM</span><span class="v">744B MoE · from 16 GB</span></div></div><a class="btn ghost" href="https://github.com/JustVugg/colibri#quick-start">Run it →</a></div>@@ -364,7 +364,7 @@ footer li a:hover{color:var(--on-dark)}<li><a href="https://github.com/JustVugg/colibri/blob/main/docs/benchmarks.md">Benchmarks</a></li></ul></div><div><h4>Models</h4><ul>- <li><a href="#models">GLM-5.2</a></li><li><a href="#models">GLM-5.3-Flash</a></li>+ <li><a href="#models">GLM-5.2/5.3</a></li><li><a href="#models">GLM-5.3-Flash</a></li><li><a href="#models">Qwen3.8-Flash-Next</a></li><li><a href="#models">Kimi K3</a></li><li><a href="#models">Inkling</a></li><li><a href="#models">DeepSeek V4 Flash</a></li><li><a href="#models">Qwen3.6</a></li><li><a href="#models">OLMoE</a></li>