{ "reviewed": "2026-09-10", "models": [ { "id": "philipjohnbasile/wisp-coder-110m", "slug": "wisp-coder-110m", "title": "Wisp Coder", "family": "Wisp", "goal": "Code completion", "kind": "Model", "status": "Live CPU demo", "summary": "A small code model trained from scratch to study fill-in-the-middle and native multi-token prediction.", "runtime": "Transformers for the standard trunk; the public demo runs on CPU. The separate MLX MTP reference is a research decoder.", "caveat": "108.2M parameters including MTP; 100.7M in the standard trunk. The registered short-hole study did not establish superiority over its FIM comparators.", "revision": "5919d57a7672623d8afc066032794359350e2ed2", "tensor_bytes": 216326685, "tensor_files": { "model.safetensors": 201377177, "mtp.safetensors": 14949508 }, "repo_tensor_bytes": 216326685, "evidence": "fim-competitive-v1/FIM_RESULTS.md", "previous_revision": "6ca84fe06b933b4477eaf9561941e3155906e502" }, { "id": "philipjohnbasile/ornith-ai-Ornith-1.5-35B-A3B-V2-MTPLX", "slug": "ornith-ai-Ornith-1.5-35B-A3B-V2-MTPLX", "title": "Ornith 35B \u00b7 MTPLX", "family": "MTPLX conversions", "goal": "MLX & MTP", "kind": "Model", "status": "Recorded runtime", "summary": "A mixed-precision MLX conversion of Ornith with a BF16 MTP sidecar and preserved vision weights.", "runtime": "MTPLX; the published Forge manifest selects depth 1. The card distinguishes recorded Forge results from the later CLI dry-run audit.", "caveat": "35B logical base parameters. A packed tensor-element count is a different quantity. Upstream license metadata remains conflicting; read the attribution section.", "revision": "ecff18949201c114bc6c99d10162764af7e7be9a", "tensor_bytes": 22091579500, "tensor_files": { "model-00001-of-00004.safetensors": 5343699404, "model-00002-of-00004.safetensors": 5368472723, "model-00003-of-00004.safetensors": 5368324201, "model-00004-of-00004.safetensors": 3428527593, "model-vision.safetensors": 893179529, "mtp.safetensors": 1689376050 }, "repo_tensor_bytes": 22091579500, "evidence": "mtplx_runtime.json", "previous_revision": "70c256f8644e70e24323c865b24f7011b4a940f5" }, { "id": "philipjohnbasile/Qwen3.6-27B-Fable-Fusion-711-MTPLX-6bit", "slug": "Qwen3.6-27B-Fable-Fusion-711-MTPLX-6bit", "title": "Fable Fusion \u00b7 6-bit", "family": "Fable Fusion", "goal": "MLX & MTP", "kind": "Model", "status": "Recorded runtime", "summary": "The middle-size Fable Fusion conversion, with native MTP and vision, reconstructed through the source Q8_0 GGUF.", "runtime": "MTPLX for the MTP path; compatible MLX tooling for ordinary autoregressive generation. The card records M5 Max / MTPLX 2.6.0 measurements.", "caveat": "A useful starting point for comparing the 4-, 6- and 8-bit family. Published throughput depends on prompt, decoding mode and measurement method.", "revision": "c023162ccc1596fca0af668ec4a7284aa1633735", "tensor_bytes": 23627279379, "tensor_files": { "model-00001-of-00005.safetensors": 5366732647, "model-00002-of-00005.safetensors": 5344200019, "model-00003-of-00005.safetensors": 5331479308, "model-00004-of-00005.safetensors": 4780958327, "model-00005-of-00005.safetensors": 1033011548, "model-vision.safetensors": 921497125, "mtp.safetensors": 849400405 }, "repo_tensor_bytes": 23627279379, "evidence": "mtplx_runtime.json", "previous_revision": "a59943dbf4ed715374b95f88541365d30fa6464d" }, { "id": "philipjohnbasile/Qwen3.6-27B-Fable-Fusion-711-MTPLX-4bit", "slug": "Qwen3.6-27B-Fable-Fusion-711-MTPLX-4bit", "title": "Fable Fusion \u00b7 4-bit", "family": "Fable Fusion", "goal": "MLX & MTP", "kind": "Model", "status": "Recorded runtime", "summary": "The smallest published Fable Fusion MTPLX quantization, with a BF16 MTP head and vision.", "runtime": "MTPLX for native MTP. The card separates ordinary MTP from the runtime's prompt-context-copy path.", "caveat": "Lower storage use does not establish better quality or a specific Mac memory tier. The measured default throughput was bimodal on the tested prompt.", "revision": "4e63f04c4374d5128d0c747f631822c5839ff1ff", "tensor_bytes": 16903941043, "tensor_files": { "model-00001-of-00003.safetensors": 5328325648, "model-00002-of-00003.safetensors": 5354185130, "model-00003-of-00003.safetensors": 4450532735, "model-vision.safetensors": 921497125, "mtp.safetensors": 849400405 }, "repo_tensor_bytes": 16903941043, "evidence": "mtplx_runtime.json", "previous_revision": "ecc13ddc6afb6cb4619e85518ac54a7b865ea700" }, { "id": "philipjohnbasile/Qwen3.6-27B-Fable-Fusion-711-MTPLX-8bit", "slug": "Qwen3.6-27B-Fable-Fusion-711-MTPLX-8bit", "title": "Fable Fusion \u00b7 8-bit", "family": "Fable Fusion", "goal": "MLX & MTP", "kind": "Model", "status": "Recorded runtime", "summary": "The larger 8-bit Fable Fusion conversion, preserving the BF16 MTP sidecar and vision.", "runtime": "MTPLX for native MTP; see the card for the exact recorded hardware, prompt and decoding setup.", "caveat": "More weight storage than the 4- or 6-bit variants. Do not treat cross-harness perplexity or a single throughput sample as a universal ranking.", "revision": "f559b7eb2b243717b7e9a3100f3df4b76c1b3875", "tensor_bytes": 30350618045, "tensor_files": { "model-00001-of-00006.safetensors": 5305469313, "model-00002-of-00006.safetensors": 5354102628, "model-00003-of-00006.safetensors": 5337392986, "model-00004-of-00006.safetensors": 5292847128, "model-00005-of-00006.safetensors": 5354102635, "model-00006-of-00006.safetensors": 1935805825, "model-vision.safetensors": 921497125, "mtp.safetensors": 849400405 }, "repo_tensor_bytes": 30350618045, "evidence": "mtplx_runtime.json", "previous_revision": "1d52a333744994903e928ce9718b99f9a42f7042" }, { "id": "philipjohnbasile/Qwen3.6-27B-Fable-Fusion-711-bf16", "slug": "Qwen3.6-27B-Fable-Fusion-711-bf16", "title": "Fable Fusion \u00b7 BF16 reconstruction", "family": "Fable Fusion", "goal": "Conversion work", "kind": "Conversion source", "status": "Reference build", "summary": "A BF16 tensor reconstruction from the released Q8_0 GGUF, useful for inspecting or building further conversions.", "runtime": "Reconstruction in a Hugging Face tensor layout. It is a conversion source, not an independently qualified serving stack for every listed framework.", "caveat": "BF16 describes the stored dtype. It cannot recover precision lost in Q8_0 and is not the original training checkpoint.", "revision": "425794589d8029115727376d12b6aca018b115ca", "tensor_bytes": 55563006268, "tensor_files": { "model-00001-of-00013.safetensors": 5085604172, "model-00002-of-00013.safetensors": 4335003587, "model-00003-of-00013.safetensors": 4440939384, "model-00004-of-00013.safetensors": 4335014035, "model-00005-of-00013.safetensors": 4345478904, "model-00006-of-00013.safetensors": 4376958597, "model-00007-of-00013.safetensors": 4398994906, "model-00008-of-00013.safetensors": 4376958603, "model-00009-of-00013.safetensors": 4345478912, "model-00010-of-00013.safetensors": 4440939480, "model-00011-of-00013.safetensors": 4335014015, "model-00012-of-00013.safetensors": 4304623688, "model-00013-of-00013.safetensors": 2441997985 }, "repo_tensor_bytes": 55563006268, "evidence": null, "previous_revision": "e8f1bcee4adc52049bf805e162b0ed04da771afa" }, { "id": "philipjohnbasile/Qwen3.6-27B-Akka-6bit-MLX", "slug": "Qwen3.6-27B-Akka-6bit-MLX", "title": "Akka 27B \u00b7 6-bit", "family": "MTPLX conversions", "goal": "MLX & MTP", "kind": "Model", "status": "Recorded runtime", "summary": "A text-only MLX conversion of the Akka merge, published with the failed MTP calibration evidence.", "runtime": "Autoregressive MLX generation. The MTP head was omitted after recorded 0% draft-target agreement; there is no vision tower.", "caveat": "Use as an AR model. The same-harness WikiText-2 comparison is useful evidence; a failed draft head is not a speed feature.", "revision": "1bad3b57457d7e4ddb7fb25e9133a07344539be2", "tensor_bytes": 21856381849, "tensor_files": { "model-00001-of-00005.safetensors": 5366732647, "model-00002-of-00005.safetensors": 5344200019, "model-00003-of-00005.safetensors": 5331479308, "model-00004-of-00005.safetensors": 4780958327, "model-00005-of-00005.safetensors": 1033011548 }, "repo_tensor_bytes": 21856381849, "evidence": "receipts/calibrate.json", "previous_revision": "66efc504a206225e16b47a7c5ddb10a49016935a" }, { "id": "philipjohnbasile/hy3-family-mini-qwen35b-v1", "slug": "hy3-family-mini-qwen35b-v1", "title": "Qwen Mini \u00b7 AR", "family": "Hy3 & Qwen Mini", "goal": "Pruning & agents", "kind": "Model", "status": "Recorded runtime", "summary": "A verifier-filtered LoRA fused into a Qwen 35B MoE, sharing the Hy3 project's training approach.", "runtime": "The card records MLX generation and LM Studio loading. It uses Qwen architecture, not Tencent Hy3.", "caveat": "The 8/10 executed stress result covers that project suite. Thinking and repetition settings affect structured-output behavior.", "revision": "8ac6cd5b346e65c19e78b110e3d39eb0eb8666ad", "tensor_bytes": 19509023921, "tensor_files": { "model-00001-of-00004.safetensors": 5343699404, "model-00002-of-00004.safetensors": 5368472723, "model-00003-of-00004.safetensors": 5368324201, "model-00004-of-00004.safetensors": 3428527593 }, "repo_tensor_bytes": 19509023921, "evidence": null, "previous_revision": "5b89b5c72b089abd26b86276a943a1a61db3a081" }, { "id": "philipjohnbasile/hy3-family-mini-qwen35b-reap-v1", "slug": "hy3-family-mini-qwen35b-reap-v1", "title": "Qwen Mini \u00b7 REAP", "family": "Hy3 & Qwen Mini", "goal": "Pruning & agents", "kind": "Model", "status": "Recorded runtime", "summary": "A smaller Qwen Mini variant with 25% of experts removed and no additional post-prune LoRA heal.", "runtime": "Autoregressive MLX tooling; see the source project for the pruning and generation recipe.", "caveat": "The project recorded 38/46 passes versus 37/46 for its unpruned comparison. A one-case difference is not broad model superiority.", "revision": "96c71f2f787f242a7fe1483d1d7a82a3ec4f107f", "tensor_bytes": 14973604757, "tensor_files": { "model-00001-of-00004.safetensors": 4134207900, "model-00002-of-00004.safetensors": 4083483703, "model-00003-of-00004.safetensors": 4083195911, "model-00004-of-00004.safetensors": 2672717243 }, "repo_tensor_bytes": 14973604757, "evidence": null, "previous_revision": "bc4e27738334c0c2ba091fa123f4dc55a67e4387" }, { "id": "philipjohnbasile/hy3-family-mini-qwen35b-mtp-v1", "slug": "hy3-family-mini-qwen35b-mtp-v1", "title": "Qwen Mini \u00b7 MTP experiment", "family": "Hy3 & Qwen Mini", "goal": "MLX & MTP", "kind": "Research variant", "status": "Runtime verification pending", "summary": "The Qwen Mini trunk with an MTP head grafted from the clean Qwen source.", "runtime": "The card's original runtime-load failure is historical. Upstream later shipped Qwen MTP backend work; this exact artifact has no fresh end-to-end qualification here.", "caveat": "Use the AR sibling as the documented inference starting point. Head presence and backend recognition do not prove accepted tokens or acceleration.", "revision": "1d771141ceb880ccdab820678ef8fa75ed803807", "tensor_bytes": 19984154960, "tensor_files": { "model-00001-of-00004.safetensors": 5343699404, "model-00002-of-00004.safetensors": 5368472723, "model-00003-of-00004.safetensors": 5368324201, "model-00004-of-00004.safetensors": 3428527593, "model-mtp-head.safetensors": 475131039 }, "repo_tensor_bytes": 19984154960, "evidence": null, "previous_revision": "41481fae8362346c07bb33c3681f6adde1b0bc42" }, { "id": "philipjohnbasile/hy3-demolition-mlx-lite-v1", "slug": "hy3-demolition-mlx-lite-v1", "title": "Hy3 \u00b7 Lite AR", "family": "Hy3 & Qwen Mini", "goal": "Pruning & agents", "kind": "Model", "status": "Recorded runtime", "summary": "A fused LoRA on Tencent Hy3's 295B MoE, keeping all 192 experts per layer.", "runtime": "The release uses a pinned mlx-lm fork plus its documented FP32 router patch. The card records a 112.3 GB peak on an M5 Max with 128 GB.", "caveat": "The experimental SSD pager is a different serving path. Reduced cache tiers measured on a large Mac are not separate tests on every smaller Mac.", "revision": "b8b679e38a63cce9be73f586ea6803ad73045617", "tensor_bytes": 111151968307, "tensor_files": { "model-00001-of-00022.safetensors": 5338990287, "model-00002-of-00022.safetensors": 5047605611, "model-00003-of-00022.safetensors": 5160851740, "model-00004-of-00022.safetensors": 5160851734, "model-00005-of-00022.safetensors": 5361715118, "model-00006-of-00022.safetensors": 5085354618, "model-00007-of-00022.safetensors": 5160851840, "model-00008-of-00022.safetensors": 5160851788, "model-00009-of-00022.safetensors": 5361715110, "model-00010-of-00022.safetensors": 5085354618, "model-00011-of-00022.safetensors": 5160851830, "model-00012-of-00022.safetensors": 5160851746, "model-00013-of-00022.safetensors": 5361715148, "model-00014-of-00022.safetensors": 5085354616, "model-00015-of-00022.safetensors": 5160851842, "model-00016-of-00022.safetensors": 5160851740, "model-00017-of-00022.safetensors": 5361715116, "model-00018-of-00022.safetensors": 5085354594, "model-00019-of-00022.safetensors": 5160851798, "model-00020-of-00022.safetensors": 5160851754, "model-00021-of-00022.safetensors": 5361715112, "model-00022-of-00022.safetensors": 2006860547 }, "repo_tensor_bytes": 111151968307, "evidence": null, "previous_revision": "18b355d74312990cbe391cc39858e3baebecf271" }, { "id": "philipjohnbasile/hy3-demolition-mlx-reap25-v1", "slug": "hy3-demolition-mlx-reap25-v1", "title": "Hy3 \u00b7 REAP25 AR", "family": "Hy3 & Qwen Mini", "goal": "Pruning & agents", "kind": "Model", "status": "Recorded runtime", "summary": "A Hy3 variant pruned from 192 to 144 experts per layer, then healed and fused.", "runtime": "Pinned mlx-lm fork plus the FP32 router patch. The card records 86.7 GB peak resident inference on the 128 GB M5 Max.", "caveat": "45/46 project-suite passes versus 46/46 for Lite. A native tool-call separator regression is documented and matters for strict parsers.", "revision": "44a2a9db4bb334553aa780af3b4af21e1581e117", "tensor_bytes": 85787174066, "tensor_files": { "model-00001-of-00017.safetensors": 5080877602, "model-00002-of-00017.safetensors": 5317557562, "model-00003-of-00017.safetensors": 5317557686, "model-00004-of-00017.safetensors": 5317557656, "model-00005-of-00017.safetensors": 5317557712, "model-00006-of-00017.safetensors": 5317557754, "model-00007-of-00017.safetensors": 5317557730, "model-00008-of-00017.safetensors": 5317557710, "model-00009-of-00017.safetensors": 5317557722, "model-00010-of-00017.safetensors": 5317557706, "model-00011-of-00017.safetensors": 5317557680, "model-00012-of-00017.safetensors": 5317557752, "model-00013-of-00017.safetensors": 5317557756, "model-00014-of-00017.safetensors": 5317557712, "model-00015-of-00017.safetensors": 5317557722, "model-00016-of-00017.safetensors": 5317557650, "model-00017-of-00017.safetensors": 942930954 }, "repo_tensor_bytes": 85787174066, "evidence": null, "previous_revision": "a9ee4395f0f9aa33d456733dc515c279f7f7ca4e" }, { "id": "philipjohnbasile/hy3-demolition-mlx-lite-v1-mtp", "slug": "hy3-demolition-mlx-lite-v1-mtp", "title": "Hy3 \u00b7 Lite MTP experiment", "family": "Hy3 & Qwen Mini", "goal": "MLX & MTP", "kind": "Research variant", "status": "Runtime verification pending", "summary": "The Lite trunk with Hy3's NextN sidecar restored for MTP experiments.", "runtime": "Requires a compatible Hy3 model implementation and MTPLX runtime contract. The upstream backend shipped, but that does not qualify this artifact.", "caveat": "Use the AR sibling for the documented generation path. No current end-to-end MTP speedup is established here.", "revision": "62c042dbdc322824c3e29f024614df0cd9478d67", "tensor_bytes": 112572229613, "tensor_files": { "model-00001-of-00022.safetensors": 5338990287, "model-00002-of-00022.safetensors": 5047605611, "model-00003-of-00022.safetensors": 5160851740, "model-00004-of-00022.safetensors": 5160851734, "model-00005-of-00022.safetensors": 5361715118, "model-00006-of-00022.safetensors": 5085354618, "model-00007-of-00022.safetensors": 5160851840, "model-00008-of-00022.safetensors": 5160851788, "model-00009-of-00022.safetensors": 5361715110, "model-00010-of-00022.safetensors": 5085354618, "model-00011-of-00022.safetensors": 5160851830, "model-00012-of-00022.safetensors": 5160851746, "model-00013-of-00022.safetensors": 5361715148, "model-00014-of-00022.safetensors": 5085354616, "model-00015-of-00022.safetensors": 5160851842, "model-00016-of-00022.safetensors": 5160851740, "model-00017-of-00022.safetensors": 5361715116, "model-00018-of-00022.safetensors": 5085354594, "model-00019-of-00022.safetensors": 5160851798, "model-00020-of-00022.safetensors": 5160851754, "model-00021-of-00022.safetensors": 5361715112, "model-00022-of-00022.safetensors": 2006860547, "model-mtp-sidecar.safetensors": 1420261306 }, "repo_tensor_bytes": 112572229613, "evidence": null, "previous_revision": "d4bf9c0d3a0f4097ec5ef0630bcaacbf87bd179e" }, { "id": "philipjohnbasile/hy3-demolition-mlx-reap25-v1-mtp", "slug": "hy3-demolition-mlx-reap25-v1-mtp", "title": "Hy3 \u00b7 REAP25 MTP experiment", "family": "Hy3 & Qwen Mini", "goal": "MLX & MTP", "kind": "Research variant", "status": "Runtime verification pending", "summary": "The REAP25 trunk with a draft sidecar whose expert axis was reduced to match the 144-expert trunk.", "runtime": "Requires a compatible Hy3 implementation and MTPLX MTP contract. Structural recognition is separate from generation qualification.", "caveat": "Draft-head modification can change acceptance; output equivalence depends on a correctly implemented and verified decoding path.", "revision": "95b2563af71436e757b104b922ee1376e282fa9d", "tensor_bytes": 86886362010, "tensor_files": { "model-00001-of-00017.safetensors": 5080877602, "model-00002-of-00017.safetensors": 5317557562, "model-00003-of-00017.safetensors": 5317557686, "model-00004-of-00017.safetensors": 5317557656, "model-00005-of-00017.safetensors": 5317557712, "model-00006-of-00017.safetensors": 5317557754, "model-00007-of-00017.safetensors": 5317557730, "model-00008-of-00017.safetensors": 5317557710, "model-00009-of-00017.safetensors": 5317557722, "model-00010-of-00017.safetensors": 5317557706, "model-00011-of-00017.safetensors": 5317557680, "model-00012-of-00017.safetensors": 5317557752, "model-00013-of-00017.safetensors": 5317557756, "model-00014-of-00017.safetensors": 5317557712, "model-00015-of-00017.safetensors": 5317557722, "model-00016-of-00017.safetensors": 5317557650, "model-00017-of-00017.safetensors": 942930954, "model-mtp-sidecar.safetensors": 1099187944 }, "repo_tensor_bytes": 86886362010, "evidence": null, "previous_revision": "ac84bc50a90acdcbbffc632877f60be4b7efdddf" }, { "id": "philipjohnbasile/GLM-5.2-Demolition-q4a4-soul-MLX", "slug": "GLM-5.2-Demolition-q4a4-soul-MLX", "title": "GLM Demolition \u00b7 pruned MLX", "family": "GLM & iliria", "goal": "Pruning & agents", "kind": "Model", "status": "Recorded runtime", "summary": "An expert-pruned GLM 5.2 build with quantized root weights and separately published adapter experiments.", "runtime": "MLX with the release's GLM DSA setup. This is distinct from the unpruned iliria SSD container.", "caveat": "Root tensors are about 104.89 GB (97.68 GiB). A whole-repository download includes many adapter checkpoints and is substantially larger.", "revision": "cd61579e7503dcf73c69972a2e972bfd8796f134", "tensor_bytes": 104888105339, "tensor_files": { "model-extra.safetensors": 1546334896, "model-layer-00000.safetensors": 225532108, "model-layer-00001.safetensors": 225532120, "model-layer-00002.safetensors": 225532126, "model-layer-00003.safetensors": 1367603332, "model-layer-00004.safetensors": 1367603338, "model-layer-00005.safetensors": 1367603332, "model-layer-00006.safetensors": 1372876704, "model-layer-00007.safetensors": 1367603352, "model-layer-00008.safetensors": 1367603336, "model-layer-00009.safetensors": 1367603378, "model-layer-00010.safetensors": 1372876759, "model-layer-00011.safetensors": 1367603396, "model-layer-00012.safetensors": 1367603384, "model-layer-00013.safetensors": 1367603364, "model-layer-00014.safetensors": 1372876777, "model-layer-00015.safetensors": 1367603378, "model-layer-00016.safetensors": 1367603380, "model-layer-00017.safetensors": 1367603384, "model-layer-00018.safetensors": 1372876755, "model-layer-00019.safetensors": 1367603376, "model-layer-00020.safetensors": 1367603372, "model-layer-00021.safetensors": 1367603390, "model-layer-00022.safetensors": 1372876779, "model-layer-00023.safetensors": 1367603378, "model-layer-00024.safetensors": 1367603378, "model-layer-00025.safetensors": 1367603372, "model-layer-00026.safetensors": 1372876783, "model-layer-00027.safetensors": 1367603370, "model-layer-00028.safetensors": 1367603370, "model-layer-00029.safetensors": 1367603376, "model-layer-00030.safetensors": 1372876765, "model-layer-00031.safetensors": 1367603394, "model-layer-00032.safetensors": 1367603374, "model-layer-00033.safetensors": 1367603384, "model-layer-00034.safetensors": 1372876741, "model-layer-00035.safetensors": 1367603386, "model-layer-00036.safetensors": 1367603390, "model-layer-00037.safetensors": 1367603390, "model-layer-00038.safetensors": 1372876759, "model-layer-00039.safetensors": 1367603370, "model-layer-00040.safetensors": 1367603370, "model-layer-00041.safetensors": 1367603398, "model-layer-00042.safetensors": 1372876789, "model-layer-00043.safetensors": 1367603380, "model-layer-00044.safetensors": 1367603358, "model-layer-00045.safetensors": 1367603372, "model-layer-00046.safetensors": 1372876759, "model-layer-00047.safetensors": 1367603366, "model-layer-00048.safetensors": 1367603372, "model-layer-00049.safetensors": 1367603392, "model-layer-00050.safetensors": 1372876745, "model-layer-00051.safetensors": 1367603384, "model-layer-00052.safetensors": 1367603366, "model-layer-00053.safetensors": 1367603404, "model-layer-00054.safetensors": 1372876787, "model-layer-00055.safetensors": 1367603370, "model-layer-00056.safetensors": 1367603382, "model-layer-00057.safetensors": 1367603396, "model-layer-00058.safetensors": 1372876781, "model-layer-00059.safetensors": 1367603362, "model-layer-00060.safetensors": 1367603392, "model-layer-00061.safetensors": 1367603356, "model-layer-00062.safetensors": 1372876765, "model-layer-00063.safetensors": 1367603412, "model-layer-00064.safetensors": 1367603374, "model-layer-00065.safetensors": 1367603354, "model-layer-00066.safetensors": 1372876767, "model-layer-00067.safetensors": 1367603366, "model-layer-00068.safetensors": 1367603396, "model-layer-00069.safetensors": 1367603390, "model-layer-00070.safetensors": 1372876729, "model-layer-00071.safetensors": 1367603346, "model-layer-00072.safetensors": 1367603366, "model-layer-00073.safetensors": 1367603380, "model-layer-00074.safetensors": 1372876753, "model-layer-00075.safetensors": 1367603392, "model-layer-00076.safetensors": 1367603386, "model-layer-00077.safetensors": 1367603386 }, "repo_tensor_bytes": 187507698595, "evidence": "MODEL_CARD.md", "previous_revision": "022fb6952fc2dc5650b742bab9d359d95c783456" }, { "id": "philipjohnbasile/GLM-5.2-colibri-int4-with-int8-mtp", "slug": "GLM-5.2-colibri-int4-with-int8-mtp", "title": "GLM 744B \u00b7 iliria container", "family": "GLM & iliria", "goal": "SSD streaming", "kind": "Model", "status": "Recorded runtime", "summary": "The unpruned GLM 5.2 expert set in an int4 container with an int8 MTP head for the iliria engine.", "runtime": "iliria streams experts from SSD. The documented setup is an Apple Silicon 128 GB machine with roughly 384 GB of weight storage.", "caveat": "The on-disk payload is not resident RAM. These files are not interchangeable with the pruned MLX build; consult iliria for runtime limits.", "revision": "9768823b5f84eb6a3387aee4d95a1be94629a0f1", "tensor_bytes": 383937029112, "tensor_files": { "out-00000.safetensors": 2675062104, "out-00001.safetensors": 2686078088, "out-00002.safetensors": 2686892744, "out-00003.safetensors": 2686078024, "out-00004.safetensors": 2693176336, "out-00005.safetensors": 2686077896, "out-00006.safetensors": 2693192880, "out-00007.safetensors": 2686077752, "out-00008.safetensors": 2693193016, "out-00009.safetensors": 2686077616, "out-00010.safetensors": 2686876984, "out-00011.safetensors": 2693192448, "out-00012.safetensors": 2686078080, "out-00013.safetensors": 2693176192, "out-00014.safetensors": 2686078072, "out-00015.safetensors": 2693192688, "out-00016.safetensors": 2686077936, "out-00017.safetensors": 2686876544, "out-00018.safetensors": 2686077808, "out-00019.safetensors": 2687572224, "out-00020.safetensors": 2686077728, "out-00021.safetensors": 2693193056, "out-00022.safetensors": 2689008008, "out-00023.safetensors": 2683946216, "out-00024.safetensors": 2686892600, "out-00025.safetensors": 2686078080, "out-00026.safetensors": 2693176224, "out-00027.safetensors": 2686078040, "out-00028.safetensors": 2693192728, "out-00029.safetensors": 2686077896, "out-00030.safetensors": 2693192856, "out-00031.safetensors": 2686077760, "out-00032.safetensors": 2686876728, "out-00033.safetensors": 2686077624, "out-00034.safetensors": 2693193240, "out-00035.safetensors": 2693176056, "out-00036.safetensors": 2686078088, "out-00037.safetensors": 2693192376, "out-00038.safetensors": 2686077216, "out-00039.safetensors": 2693192176, "out-00040.safetensors": 2686077936, "out-00041.safetensors": 2686876536, "out-00042.safetensors": 2686077808, "out-00043.safetensors": 2693192960, "out-00044.safetensors": 2686077664, "out-00045.safetensors": 2693176768, "out-00046.safetensors": 2693192416, "out-00047.safetensors": 2686078080, "out-00048.safetensors": 2686876264, "out-00049.safetensors": 2686078088, "out-00050.safetensors": 2693192664, "out-00051.safetensors": 2686077976, "out-00052.safetensors": 2693192776, "out-00053.safetensors": 2686077832, "out-00054.safetensors": 2693176536, "out-00055.safetensors": 2686077712, "out-00056.safetensors": 2686893208, "out-00057.safetensors": 2688924512, "out-00058.safetensors": 2684028840, "out-00059.safetensors": 2693191784, "out-00060.safetensors": 2686078080, "out-00061.safetensors": 2693176248, "out-00062.safetensors": 2686078024, "out-00063.safetensors": 2693192744, "out-00064.safetensors": 2686077880, "out-00065.safetensors": 2686876608, "out-00066.safetensors": 2686077752, "out-00067.safetensors": 2693193032, "out-00068.safetensors": 2686077608, "out-00069.safetensors": 2693193272, "out-00070.safetensors": 2693176064, "out-00071.safetensors": 2686078096, "out-00072.safetensors": 2686892712, "out-00073.safetensors": 2686078056, "out-00074.safetensors": 2693176304, "out-00075.safetensors": 2686077928, "out-00076.safetensors": 2693192840, "out-00077.safetensors": 2686077784, "out-00078.safetensors": 2693192352, "out-00079.safetensors": 2686076792, "out-00080.safetensors": 2693176744, "out-00081.safetensors": 2686892544, "out-00082.safetensors": 2686078080, "out-00083.safetensors": 2693176168, "out-00084.safetensors": 2686078088, "out-00085.safetensors": 2693192680, "out-00086.safetensors": 2686077960, "out-00087.safetensors": 2693192792, "out-00088.safetensors": 2686077816, "out-00089.safetensors": 2686876672, "out-00090.safetensors": 2686077688, "out-00091.safetensors": 2693193120, "out-00092.safetensors": 2693175992, "out-00093.safetensors": 2686078088, "out-00094.safetensors": 2693192520, "out-00095.safetensors": 2686078088, "out-00096.safetensors": 2686876368, "out-00097.safetensors": 2686078008, "out-00098.safetensors": 2693192336, "out-00099.safetensors": 2686077016, "out-00100.safetensors": 2686876336, "out-00101.safetensors": 2686077736, "out-00102.safetensors": 2693193040, "out-00103.safetensors": 2692370328, "out-00104.safetensors": 2686900160, "out-00105.safetensors": 2686876200, "out-00106.safetensors": 2686078080, "out-00107.safetensors": 2693192608, "out-00108.safetensors": 2686078032, "out-00109.safetensors": 2693176320, "out-00110.safetensors": 2686077904, "out-00111.safetensors": 2693192856, "out-00112.safetensors": 2686077768, "out-00113.safetensors": 2686876728, "out-00114.safetensors": 2686077640, "out-00115.safetensors": 2693193240, "out-00116.safetensors": 2693192440, "out-00117.safetensors": 2686078080, "out-00118.safetensors": 2693175984, "out-00119.safetensors": 2686077232, "out-00120.safetensors": 2693192184, "out-00121.safetensors": 2686077936, "out-00122.safetensors": 2686876536, "out-00123.safetensors": 2686077808, "out-00124.safetensors": 2693192968, "out-00125.safetensors": 2686077672, "out-00126.safetensors": 2693193152, "out-00127.safetensors": 2693176008, "out-00128.safetensors": 2686078088, "out-00129.safetensors": 2686892656, "out-00130.safetensors": 2686078080, "out-00131.safetensors": 2693176264, "out-00132.safetensors": 2686077984, "out-00133.safetensors": 2693192776, "out-00134.safetensors": 2686077840, "out-00135.safetensors": 889827288, "out-00136.safetensors": 16, "out-00137.safetensors": 2150152872, "out-00138.safetensors": 2686076760, "out-00139.safetensors": 2693192472, "out-00140.safetensors": 2371602456, "out-idx-00000.safetensors": 18781168, "out-idx-00001.safetensors": 9390592, "out-idx-00002.safetensors": 9390592, "out-idx-00003.safetensors": 9390592, "out-idx-00004.safetensors": 9390584, "out-idx-00005.safetensors": 9390592, "out-idx-00006.safetensors": 9390592, "out-idx-00007.safetensors": 9390592, "out-idx-00008.safetensors": 9390592, "out-idx-00009.safetensors": 9390592, "out-idx-00010.safetensors": 9390592, "out-idx-00011.safetensors": 9390592, "out-idx-00012.safetensors": 9390592, "out-idx-00013.safetensors": 9390592, "out-idx-00014.safetensors": 9390592, "out-idx-00015.safetensors": 9390584, "out-idx-00016.safetensors": 9390592, "out-idx-00017.safetensors": 9390592, "out-idx-00018.safetensors": 9390592, "out-idx-00019.safetensors": 9390592, "out-mtp-00000.safetensors": 3527131672, "out-mtp-00001.safetensors": 5366238584, "out-mtp-00002.safetensors": 1065950496 }, "repo_tensor_bytes": 383937029112, "evidence": null, "previous_revision": "f0764f02e39257d5f737ff72d15df7ca5204d9ff" }, { "id": "philipjohnbasile/glm52-idx-extract", "slug": "glm52-idx-extract", "title": "GLM DSA \u00b7 indexer component", "family": "GLM & iliria", "goal": "Conversion work", "kind": "Component", "status": "Component only", "summary": "Sparse-attention indexer tensors extracted from GLM 5.2 for runtime and conversion work.", "runtime": "A component for engines that implement the GLM DSA path. It is not a complete language model and cannot generate by itself.", "caveat": "The root tensor files total about 197 MB in this revision. The model config identifies the source architecture, not standalone completeness.", "revision": "3a0149c78c163d62d5c1c2aa8f2d1cfb2b53e8fe", "tensor_bytes": 197202400, "tensor_files": { "out-idx-00000.safetensors": 18781168, "out-idx-00001.safetensors": 9390592, "out-idx-00002.safetensors": 9390592, "out-idx-00003.safetensors": 9390592, "out-idx-00004.safetensors": 9390584, "out-idx-00005.safetensors": 9390592, "out-idx-00006.safetensors": 9390592, "out-idx-00007.safetensors": 9390592, "out-idx-00008.safetensors": 9390592, "out-idx-00009.safetensors": 9390592, "out-idx-00010.safetensors": 9390592, "out-idx-00011.safetensors": 9390592, "out-idx-00012.safetensors": 9390592, "out-idx-00013.safetensors": 9390592, "out-idx-00014.safetensors": 9390592, "out-idx-00015.safetensors": 9390584, "out-idx-00016.safetensors": 9390592, "out-idx-00017.safetensors": 9390592, "out-idx-00018.safetensors": 9390592, "out-idx-00019.safetensors": 9390592 }, "repo_tensor_bytes": 197202400, "evidence": null, "previous_revision": "9487fc4e34dc0b0face4f173727c3e2084310a0e" }, { "id": "philipjohnbasile/Qwen3.6-27B-Refusal-Ablation-v2-MTPLX-6bit", "slug": "Qwen3.6-27B-Refusal-Ablation-v2-MTPLX-6bit", "title": "Qwen 27B \u00b7 behavior study v2", "family": "Behavior studies", "goal": "Behavior studies", "kind": "Research variant", "status": "Recorded runtime", "summary": "A post-hoc refusal-direction intervention on official Qwen 27B, with 6-bit weights, MTP and vision.", "runtime": "MTPLX depth 2 is the card's recorded recommendation. This lineage starts from official Qwen, not Fable Fusion.", "caveat": "Refusal-marker counts are a substring proxy. The final artifact passed its registered gates while losing 1.195 ARC points against official BF16.", "revision": "5c1b245d2327c30177d69a15048005f627769fdb", "tensor_bytes": 23628594804, "tensor_files": { "model-00001-of-00005.safetensors": 5366732647, "model-00002-of-00005.safetensors": 5344200019, "model-00003-of-00005.safetensors": 5331479308, "model-00004-of-00005.safetensors": 4780958327, "model-00005-of-00005.safetensors": 1033011548, "model-vision.safetensors": 921497137, "mtp.safetensors": 849400389, "refusal_dirs_conditioned_base.safetensors": 1315429 }, "repo_tensor_bytes": 23628594804, "evidence": "evaluation/ABLATION_V2_RESULTS.md", "previous_revision": "1f2b7cf918e69e23b5af6ba94a9b73ec8bcf9664" }, { "id": "philipjohnbasile/Qwen3.6-27B-Fable-Fusion-711-ReducedRefusal-MTPLX-6bit", "slug": "Qwen3.6-27B-Fable-Fusion-711-ReducedRefusal-MTPLX-6bit", "title": "Fable Fusion \u00b7 behavior study", "family": "Behavior studies", "goal": "Behavior studies", "kind": "Research variant", "status": "Recorded runtime", "summary": "A separately locked and evaluated refusal-direction intervention on the Fable Fusion lineage.", "runtime": "MTPLX for its MTP sidecar; the card distinguishes artifact gates from broader capability claims.", "caveat": "The confirmation set was selected for baseline marker matches and is now consumed. Its 100-to-42 result is not a general semantic compliance rate.", "revision": "a2108a130430bb9f7a544afdbb09564991315d12", "tensor_bytes": 23628594980, "tensor_files": { "model-00001-of-00005.safetensors": 5366732647, "model-00002-of-00005.safetensors": 5344200019, "model-00003-of-00005.safetensors": 5331479308, "model-00004-of-00005.safetensors": 4780958327, "model-00005-of-00005.safetensors": 1033011548, "model-vision.safetensors": 921497299, "mtp.safetensors": 849400403, "refusal_dirs_fable711.safetensors": 1315429 }, "repo_tensor_bytes": 23628594980, "evidence": "fable711_mtp_runtime_gate.json", "previous_revision": "a45833ed4d0fb56f9c8c5e2a4c2cca7bd23f0798" }, { "id": "philipjohnbasile/DeepSeek-V4-Flash-0731-MLX-M5Max-TargetOnly", "slug": "DeepSeek-V4-Flash-0731-MLX-M5Max-TargetOnly", "title": "DeepSeek V4 \u00b7 target-only reference", "family": "DeepSeek", "goal": "Conversion work", "kind": "Research variant", "status": "Reference build", "summary": "An experimental target-only MLX conversion of DeepSeek V4 Flash 0731 with public comparison evidence.", "runtime": "A companion native engine on the recorded 128 GB M5 Max setup. No MTP or vision; the card marks this build superseded on that machine.", "caveat": "Math and reasoning regressed in the 50-task comparison. Constrained and unconstrained GPQA results use different protocols and should be read separately.", "revision": "7e79822c02cf1ed750865a54eecfb43e25f53d63", "tensor_bytes": 103849215724, "tensor_files": { "model-layer-0.safetensors": 2232226400, "model-layer-1.safetensors": 2232226400, "model-layer-10.safetensors": 2256453968, "model-layer-11.safetensors": 2234674280, "model-layer-12.safetensors": 2256453968, "model-layer-13.safetensors": 2234674280, "model-layer-14.safetensors": 2256453968, "model-layer-15.safetensors": 2234674280, "model-layer-16.safetensors": 2256453968, "model-layer-17.safetensors": 2234674280, "model-layer-18.safetensors": 2256453968, "model-layer-19.safetensors": 2234674280, "model-layer-2.safetensors": 2262658336, "model-layer-20.safetensors": 2256453968, "model-layer-21.safetensors": 2234674280, "model-layer-22.safetensors": 2256453968, "model-layer-23.safetensors": 2234674280, "model-layer-24.safetensors": 2256453968, "model-layer-25.safetensors": 2234674280, "model-layer-26.safetensors": 2256453968, "model-layer-27.safetensors": 2234674280, "model-layer-28.safetensors": 2256453968, "model-layer-29.safetensors": 2234674280, "model-layer-3.safetensors": 2234674232, "model-layer-30.safetensors": 2256453968, "model-layer-31.safetensors": 2234674280, "model-layer-32.safetensors": 2256453968, "model-layer-33.safetensors": 2234674280, "model-layer-34.safetensors": 2256453968, "model-layer-35.safetensors": 2234674280, "model-layer-36.safetensors": 2256453968, "model-layer-37.safetensors": 2234674280, "model-layer-38.safetensors": 2256453968, "model-layer-39.safetensors": 3778178160, "model-layer-4.safetensors": 2256453912, "model-layer-40.safetensors": 3799957848, "model-layer-41.safetensors": 3778178160, "model-layer-42.safetensors": 3799957848, "model-layer-5.safetensors": 2234674232, "model-layer-6.safetensors": 2256453912, "model-layer-7.safetensors": 2234674232, "model-layer-8.safetensors": 2256453912, "model-layer-9.safetensors": 2234674232, "model-top.safetensors": 1125524468 }, "repo_tensor_bytes": 103849215724, "evidence": "benchmarks/gpqa-diamond-20260813/summary.json", "previous_revision": "3675f64971fe786553a55b503ea6b8dcabe59181" } ] }