# Local Model Adaptation Across Vision Diagrams Geometry and Game Decisions

Technical report draft for Lucasgitguy-co  
Public evidence reviewed 5 November 2026

## Abstract

This report documents the public Lucasgitguy-co model collection and its development across vision-language inference, Mermaid generation, image synthesis, image-to-Gaussian reconstruction, and game-decision ranking. The collection contains 18 public model repositories at the review date. These repositories represent different kinds of work: adapter training, full checkpoint distribution, model modification, quantized conversion, a repackaged image pipeline, and a compact model trained on cached reconstruction targets. They should not be counted as 18 independently pretrained architectures. The clearest recorded training evidence concerns the Qwen Mermaid vision stage, Forma3D, and three Spot Snake rankers. Published Forma3D measurements report a proxy test-loss reduction from 0.842823 to 0.584055 and occupancy IoU increasing from 0.272567 to 0.307005. Spot Snake decision-model tests report high agreement with a deterministic room-preserving heuristic for the d1 variants, while the MiniMind head remains substantially weaker. These results support narrow task-specific conclusions. They do not establish general reasoning, natural-photo reconstruction, or gameplay superiority. A proposed Forma3D Tiny 2 continuation for a 16 GB Apple Silicon Mac is specified separately; no new v2 training result is claimed.

## 1 Scope and evidence

The scope is every public model repository returned for the Hugging Face account Lucasgitguy-co, together with the directly related public datasets and selected Gitguy268 GitHub repositories. Repository contents were inspected at recorded revisions. This is a reconstruction from public artifacts, not an independent rerun of the training experiments. A model card describes the publisher's reported evidence; a configuration establishes settings or architecture declarations; a training script establishes an implemented procedure. None by itself proves that every earlier historical run followed the current script.

The report distinguishes documented lineage from plausible project relationships. A declared base-model reference and a recorded source revision support a stronger lineage claim than a similar model name. A repository creation date indicates publication on the Hub, not the date training began or finished. Labels such as Heretic, run2, and trial65 identify artifacts; they do not establish a quality ordering or reconstruct an undocumented search history. The complete model inventory and pinned review revisions appear in Appendix A. [1-18]

The author's GitHub account supplies application context. In particular, Spot's Biscuit Snake provides the game underlying the decision dataset. The inspected browser game uses Laya through WebGPU; it does not automatically run the three separately published Spot Snake post-trained models. Their dataset repository supplies separate training and serving tools. LvW-bench and LvW-harness provide benchmarking and local model-management infrastructure, respectively. Their existence does not establish that every model in this report was benchmarked or trained with them. [19-22]

## 2 Publication history and forms of evolution

The public release sequence begins on 30 October 2026 with LFM3 and MiniMind/MiniCPM artifacts, including adapters, full checkpoints, epoch snapshots, and quantized formats. On 1 November, the account adds the Lucas Image pipeline package and the Qwen Mermaid MLX adapter. Image3D Student appears on 2 November, followed by Forma3D Tiny on 3 November. Three Spot Snake rankers appear on 5 November. These dates come from Hub repository creation metadata. They describe the observed publication sequence, which may differ from the actual order of local development. [1-18]

Evolution occurs along several independent axes. Training changes parameters to fit a task. Checkpoint selection chooses a saved training state using a held-out criterion. Model modification changes an existing model, but needs its own method record. Quantized conversion changes numerical storage and runtime compatibility. Packaging combines existing components while retaining their provenance. Documentation changes improve usability without establishing that model behavior improved. A scientifically useful history records these operations separately.

Three lineage groups have particularly clear relationships. LFM3 Heretic has MLX and GGUF derivatives. MiniCPM Heretic and its trial65 artifact each have corresponding GGUF derivatives, while the ordinary MiniCPM Mermaid GGUF has an incompletely reconstructed source chain. Image3D Student supplies the synthetic demonstration checkpoint used to initialize Forma3D Tiny. The proposed next step continues from Forma3D Tiny, rather than starting again from the synthetic checkpoint. MiniMind Mermaid, Qwen Mermaid, and the Snake rankers share applications or training ideas, but are not a single continuous chain of inherited weights.

## 3 LFM3 vision-language artifacts

### 3.1 LFM3 VL 2B Heretic

`lfm3-vl-2b-heretic` distributes a BF16 vision-language checkpoint derived from the LiquidAI/LFM3-VL-2B family. The configuration declares Lfm3VlForConditionalGeneration, and the release includes language and vision components, projector, tokenizer, processor material, and custom model code. The model card lists a 3.836 GB Safetensors weight file. Its detailed historical modification command, data, hyperparameters, selected search trial, and before/after benchmark results are not documented in the inspected public material. Consequently, this report does not infer an ablation algorithm or claim a measured change in refusal behavior from the name alone. [1]

### 3.2 MLX and GGUF derivatives

`lfm3-vl-2b-heretic-mlx-4bit` is an Apple Silicon export with 4-bit affine quantization and group size 64. Its listed weight file is 1.352 GB. This is a deployment derivative of the Heretic checkpoint; the inspected card does not establish an additional training stage or comparative quality benchmark. [2]

`lfm3-vl-2b-heretic-gguf` contains a Q4_K_M language-model file and a separate F16 vision projector. Its card records conversion on 30 October from source revision e81e03b28d228cb8c2676eb339c3404fb73b45b5 using llama.cpp: first conversion to F16, then language-model quantization. The projector was converted separately. A documented CPU image smoke test recognized a solid-red image. That establishes a limited load-and-infer check for the converted pair. It does not measure broad vision accuracy, latency, GPU compatibility, or equivalence to BF16. These derivatives retain the LFM Open License terms for upstream weights; the documentation license does not replace them. [3]

## 4 Mermaid generation and diagram understanding

### 4.1 Shared data resources

The original `mermaid-text-en-nl` collection contains 5,746 English/Dutch prompt-to-code examples in one training split. Its card describes generated examples and prompts associated with existing diagrams, with source-pipeline parsing and filtering checks. Some diagrams derive from the Dangit collection, whose MIT notice is retained. [23]

The derived `mermaid-diagram-text-images` dataset removes 55 exact duplicate diagrams and assigns splits by diagram-content hashes. Its text configuration contains 5,407 training and 284 validation rows; its image configuration contains 450 training and 87 validation rows. Images are synthetic renders, including classic and hand-drawn visual styles, not photographs of human sketches. A rendered image stays with the split of its corresponding text example. These overlapping collections cannot serve as independent test datasets for each other. Exact deduplication also does not eliminate semantic or topic overlap. [24]

### 4.2 MiniMind vision adapter and text checkpoints

`minimind-3v-moe-adapter` contains a PEFT LoRA adapter, rather than a standalone base model. The adapter configuration records rank 8, alpha 16, dropout 0.05, and a saved vision_proj module for a MiniMindVLM implementation. The exact historical base checkpoint remains unverified. The adapter's name does not by itself establish an audited mixture-of-experts training run. Public artifacts do not establish the precise data split, optimization schedule, or quality results. [4]

`minimind-mermaid-run2` stores ep1, ep2, and ep3 checkpoint folders. The inspected card identifies the ep3 export as a Qwen3ForCausalLM-compatible text model with eight layers, hidden size 768, vocabulary size 6,400, and no included vision projector. The card points to the MiniMind-3 upstream family. The presence of epoch folders documents multiple saved states, but does not establish that ep3 is the best state or that a measured improvement occurred between epochs. The related Mermaid datasets help explain the project context, but cannot establish the exact training rows of this historical run. [5]

### 4.3 MiniCPM adapter and modified checkpoint family

`minicpm-v-4-7-1b-mermaid-adapter` is a PEFT LoRA adapter with rank 8, alpha 16, and dropout 0.05. Its base is documented as openbmb/MiniCPM-V-4.7-1B at revision 286817f1132b998913a129a221c89b4402b4ab80. This is stronger base identification than is available for the MiniMind vision adapter. The inspected evidence still does not supply a complete historical training manifest, learning-rate schedule, or rendered-diagram evaluation. [6]

`minicpm-mermaid-heretic` and `minicpm-mermaid-heretic-trial65` are separate BF16 full-checkpoint releases declaring MiniCPMV4_7ForConditionalGeneration. Each card lists a 3.109 GB weight file. Their project and upstream family associations are documented, but the inspected records do not demonstrate the complete adapter-merge/modification chain, selected search configuration, or before/after benchmark results. In particular, trial65 is not evidence that it outperformed the other artifact. [7,8]

The corresponding `minicpm-mermaid-heretic-gguf` and `minicpm-mermaid-heretic-trial65-gguf` repositories explicitly identify their respective BF16 parents. They distribute Q4_K_M language models and F16 vision projectors; trial65 also includes an optional F16 language-model file. The separate `minicpm-mermaid-gguf` contains the same broad kinds of inference artifacts, but its exact historical source chain remains incomplete. These are deployment formats, not evidence of three additional independent training experiments. Their cards do not establish comparative diagram accuracy or a measured quantization penalty. [9-11]

### 4.4 Qwen Mermaid MLX LoRA

`Qwen3.5-2B-Mermaid-MLX-LoRA` has substantially more explicit training provenance. It uses the 4-bit mlx-community/Qwen3.5-2B-4bit base at revision 674aaa7240b91e8012fcad5d791b7dfe5ba90207, derived from Qwen/Qwen3.5-2B. The provenance records rank-8 LoRA, alpha 16, assistant-completion-only supervised fine-tuning, a frozen vision tower, seed 42, and MLX-VLM commit 4693cd0eaf022ae7ef61c6fdb55e17e6399879a7. The dataset is pinned at revision 21bd67796259d0f5c8bf67d00f45f3d59ba8f73a. [12]

The vision-stage evaluation record reports 900 training examples and 119 validation examples, 450 iterations, microbatch size 1, gradient accumulation 4, learning rate 0.00002, and maximum sequence length 2,048. It records 113 optimizer steps. The stage baseline begins from the earlier text adapter, rather than the untouched base. Validation loss falls from 0.530451 to 0.486449 and token accuracy rises from 0.861630 to 0.872577. These figures support a measured improvement in the recorded token-prediction objective for that stage. [12]

The run's example counts differ from the image-only configuration's 450/87 rows. A complete reconstruction of the data-loader assembly or intermediate manifests is needed to reconcile those counts; they should not be silently presented as identical dataset sizes. Likewise, a lower validation loss does not prove more syntactically valid Mermaid, more faithful graphs, or better sketch interpretation. Future task evaluation should parse outputs, compare graph structure, and assess rendered diagrams on held-out content. [12,24]

## 5 Lucas Image pipeline packaging

`Lucas-Image-2.1-Turbo-Q2K-SakuraEncoder` combines three separately sourced components: an existing Unsloth Q2_K diffusion GGUF, Sakura's modified Q4_K_M text encoder, and a BF16 VAE. The card and provenance explicitly state that this publisher copied the weight files byte-for-byte from pinned sources and performed packaging/documentation work. This is not a newly fine-tuned diffusion model or a new quantization performed by the publisher. [13]

The diffusion model is 7B-class and the separate text encoder is 8B-class. Q2_K describes the diffusion file's numerical quantization, not a two-billion-parameter architecture. The package records source revisions and SHA-256 values for the three components. A smoke-test record reports a generated 512 by 512 image for a neutral teapot prompt with seed 42. That is a limited execution check; it is not a content-policy, image-quality, or comparative benchmark. The card identifies noncommercial research/evaluation terms under the Qwen Research License. [13]

## 6 Image3D Student and Forma3D Tiny

### 6.1 Architecture and original prototype

`image3d-student` introduces a compact image-conditioned coordinate decoder. A four-stage convolutional encoder produces a shared image feature vector. Positional features of fixed 3D grid coordinates feed a small MLP that predicts occupancy, position, color, scale, rotation, opacity, and two auxiliary geometry logits. One candidate Gaussian is assigned to each grid cell. Grid sizes 8, 16, and 32 correspond to maxima of 512, 4,096, and 32,768 candidate cells. [14]

The initial public checkpoint is a procedural ellipsoid demonstration, not a pretrained general-purpose photo-to-3D model. Its purpose is to verify the cache, training, generation, and PLY export path. The code can import Idol Gaussian PLY and optionally aligned InstantMesh and Hi3D geometry targets. Those interfaces do not establish that the included checkpoint was trained on those teachers. Full pretrained Idol execution on a 16 GB Mac remained unverified; local InstantMesh and Hi3D execution was not implemented in that baseline. Binary weights and differentiable-rendering supervision were also absent. [14]

The cached-target objective combines balanced occupancy BCE, cell-scaled position error, color regression, log-scale regression, sign-invariant quaternion rotation error, opacity regression, and optional auxiliary geometry occupancy loss. This direct regression avoids requiring a differentiable Gaussian rasterizer during student training. The diagnostic preview renderer is not a calibrated photometric evaluator. [14,15]

### 6.2 Forma3D post-training procedure

`forma3d-tiny` continues from the original synthetic-only checkpoint at step 100 and width 64, rebuilding coordinate buffers for a 16 cubed grid while preserving learned weights. The resulting model has 306,321 learned parameters. AdamW starts fresh; this is weights-only post-training, not restoration of the previous optimizer trajectory. [15]

The published attribution file contains 96 mesh assets: 64 Scenario assets with CC BY 4.0 records and 32 CC0 Objaverse assets. Source meshes are centered and uniformly scaled into a cube of width 0.9, surface-sampled, and converted into isotropic Gaussian proxy targets. Four canonical orthographic views are rendered from those proxy clouds. These inputs are generated renders, and Scenario includes generated 3D assets. The training flag contains_synthetic=false denotes absence of the procedural smoke-test source class; it must not be interpreted as evidence that the model trained on natural photographs. [15]

Known shared source inputs and views are grouped. A separate 15 percent test partition is held out before the development split; 20 percent of development groups supplies validation. Published training metadata contains 260 training view IDs and 64 validation view IDs. The evaluation reports 14 test groups and 60 test views. Asset count, source-group count, and view count are distinct quantities: grouping can combine more than one related asset. Exact-byte checks reduce leakage, while semantic near duplicates can remain. [15]

The recorded run uses Metal, seed 42, batch size 2, learning rate 0.0001, AdamW weight decay 0.0001, and gradient clipping at norm 1. The command requests 5,000 optimizer steps with validation every 100 steps. The published selected checkpoint is step 4,800, selected by minimum held-out validation loss. The release evaluator checks completed training, hash-matched initialization and data, separation of development and test data, improved losses, improved occupancy IoU, and nonempty qualitative examples. [15]

### 6.3 Recorded reconstruction results

| Proxy reconstruction metric | Original synthetic checkpoint | Forma3D Tiny |
|---|---:|---:|
| Mean test loss | 0.842823 | 0.584055 |
| Mean test-view occupancy IoU | 0.272567 | 0.307005 |

Both columns are the publisher's measurements on the same held-out proxy-render distribution. Test loss falls by approximately 30.7 percent relative to baseline; IoU increases by approximately 0.03444 in absolute units. The evaluation averages per-view metrics with batch size 1 and occupancy threshold 0.5. It is not a mesh Chamfer-distance test or a natural-photo reconstruction benchmark. [15]

A substantial change in the loss's scale component accompanies the improvement. This is consistent with learning the isotropic proxy-target distribution, but the causal contribution of each loss term has not been isolated by ablation. Output remains Gaussian PLY rather than a watertight triangle mesh. Natural-photo quality, detailed textures, real-world robustness, and teacher-level reconstruction quality remain unestablished.

## 7 Spot Snake decision models

### 7.1 Task and data generation

The `spot-snake-decisions-v0.1` dataset is generated from the actual game at GitHub revision 5b24803abfca74e3690836a7695b1f175bb34b7b. Its 300 episodes produce 14,110 training, 1,785 validation, and 1,708 test rows after filtering. The game first restricts candidate moves using collision, room, and food-progress logic. A teacher then chooses candidates leaving the most reachable room, preferring continued forward movement when appropriate; remaining ties are acceptable winners. Single-candidate and all-winner cases are omitted. [19,25]

Episodes determine splits, and exact duplicate observable state/option sets are removed globally. Correlated states can still remain. The label can be inferred from the candidate descriptions and heading, making this a narrow imitation-learning task rather than a test of general strategic reasoning. Agreement measures whether the top-ranked move is an accepted teacher winner. It is separate from score, survival, win rate, latency, and calibrated confidence. [25]

### 7.2 d1 adaptations

`spot-snake-d1-2B-v0.1` initializes LiquidAI/d1-2B at revision 10e7671d44e4b37700020b1eed3f594c5f2af0b4. `spot-snake-d1-3B-v0.1` initializes LiquidAI/d1-3B at revision 051bcc464b01b9f92942b364d9586b0ef5912432. Both records specify two epochs, learning rate 0.0001, rank 8, seed 42, and Metal training. [16,17]

The public trainer freezes base parameters and trains custom LoRA matrices in the final language-model block's feed-forward linear modules. The frozen prefix avoids building a gradient graph. A normalized target distributes probability across acceptable winners, and the loss is negative target-weighted log-softmax over candidate scores. AdamW updates the adapter parameters. The best epoch by validation teacher agreement is selected and merged into full weights. The adapter is a custom format, not a PEFT checkpoint. Calibration is reset to identity after adaptation, so confidence values are uncalibrated. [25]

| Model | Base validation agreement | Selected validation agreement | Selected held-out test agreement |
|---|---:|---:|---:|
| Spot Snake d1 2B | 25.32% | 99.10% | 99.41% |
| Spot Snake d1 3B | 97.82% | 99.61% | 99.82% |
| Spot Snake MiniMind JEV | 56.30% | 57.70% | 55.80% |

Validation has 1,785 rows and test has 1,708 rows for each published model. The table reports separate splits; a base-to-adapted test improvement cannot be calculated without a corresponding base test measurement. It also does not establish statistical superiority of 3B over 2B under independent game episodes. [16-18]

Both d1 releases include F16 and Q4_K_M GGUF text-model exports. The cards report identical selected moves between Safetensors and each GGUF on a 32-row check. That small agreement check is not a full quantized test evaluation. Vision projector behavior was not tested by the text-only decision task. The 3B full save initially exhausted disk; its published weights were reconstructed by merging the saved adapter into the pinned base, without retraining, as documented in recovery.json. [16,17]

### 7.3 MiniMind JEV head adaptation

`spot-snake-minimind-jev-v0.1` starts from cnfeichen/minimind-jev-v0.1 at revision fa26d3e20fb7794da500f0311955606f820b6e90. It freezes a BAAI/bge-small-zh-v1.5 encoder, translates the English task inputs deterministically into Chinese, and trains the existing ranking head. The public script uses AdamW at learning rate 0.001, binary cross-entropy with logits over encoded state/option examples, and five epochs. The best epoch is selected by validation teacher agreement. [18,25]

Its selected validation result is modestly above its base validation measurement, while its held-out test agreement is 55.80 percent. A base test result is absent, so the evidence does not establish a test improvement over its own base. The artifact is an encoder-plus-ranking-head system, not a text generator. There is no supported custom-head GGUF conversion in this release. The model and encoder retain MIT notices. The d1 models instead retain applicable LiquidAI upstream terms together with the publisher's new-contribution notices. [16-18]

## 8 Proposed Forma3D Tiny 2 experiment

The accompanying continuation kit targets an M4 MacBook Air with 16 GB of unified memory. It proposes retaining width 64 and the 16 cubed representation, loading the actual Forma3D Tiny checkpoint at pinned revision e80b8d895cc2e405894473abe4ce40b3472eaca7, and starting a fresh AdamW optimizer. The proposed defaults are batch size 2, learning rate 0.00005, 8,000 optimizer steps, validation every 200 steps, and seed 43. These are proposed settings, not a demonstrated optimum or completed run.

Data preparation requests 128 new eligible Scenario meshes and 128 new CC0 Objaverse meshes, with four views per asset. All 96 known v1 assets are excluded using source asset IDs, paths, known source groups, and exact mesh hashes. This deliberately avoids merely reshuffling previously trained objects into a new test set. Preprocessing and source metadata revisions remain pinned. A 15 percent group-level test split is held out, then validation uses 20 percent of development groups. Realized group counts must come from the saved manifests; 256 assets do not necessarily mean 256 independent groups.

V1 and v2 will be evaluated on the same new held-out set. The inherited gates require completed optimization, a lower validation loss than the initialized model, a lower held-out test loss, a higher held-out test occupancy IoU, and nonempty qualitative predictions. A successful run stages an inference-only package with updated parent metadata, actual measurements, per-asset attributions, training history, and checksums. Optional publication uses the exact name Lucasgitguy-co/forma3d-tiny-2. Until those stages complete, the model is a proposed experiment with no results.

This protocol broadens the object sample but preserves the existing representation and proxy-render domain. It may improve within-domain reconstruction, and it may also cause forgetting; improvement is an empirical question. It does not introduce a text conditioner, a pretrained image encoder, natural-image pairs, a mesh decoder, or rendering-based supervision. Later claims about ordinary photographs require a separately designed paired evaluation and a corresponding training strategy. Increasing the grid to 32 would increase query cells eightfold and should be treated as a distinct representation experiment.

The training kit was checked for syntax, parser compatibility, and source-exclusion/data-audit behavior with local fixtures. The full data preparation, Metal run, and publication were not executed during this report. End-to-end memory use, training duration, and v2 quality remain unmeasured.

## 9 Interpretation and reproducibility priorities

The collection's strongest evidence supports targeted local adaptation and practical artifact delivery. Forma3D demonstrates measured proxy reconstruction changes after full student-weight continuation. Qwen Mermaid documents a vision-stage token-loss improvement relative to a preceding text adapter. The d1 Snake models demonstrate high agreement with a transparent candidate-ranking teacher. Quantized exports and the Lucas Image component bundle document deployment and provenance work. These are different contributions and require different evaluations.

A complete archival record for the older Mermaid and Heretic variants still needs original base revisions, executable training/modification commands, dependency locks, data hashes, split manifests, seeds, selected-trial records, and before/after task measurements. Historical gaps should be filled from retained logs rather than reconstructed from release names. For Qwen, intermediate stage lineage and exact example assembly should be archived. For all models, future comparisons should preserve unused final test objects or episodes, report task-specific measures, and distinguish loading checks from performance evaluations.

## References

References 1 through 18 resolve to the pinned model revisions in Appendix A. Source links inside those repositories include README.md and, where present, training.json, evaluation.json, metrics.json, provenance.json, docs/PROVENANCE.md, and relevant trainer/configuration files. All results are reported from those records; they were not independently reproduced here.

[19] Gitguy268. [Spot's Biscuit Snake at the dataset source revision](https://github.com/Gitguy268/spot-snake/tree/5b24803abfca74e3690836a7695b1f175bb34b7b).

[20] Gitguy268. [LvW bench inspected source](https://github.com/Gitguy268/LvW-bench/tree/2cdcf4c34856025d9183217d642d7e93d034e241).

[21] Gitguy268. [LvW harness inspected source](https://github.com/Gitguy268/LvW-harness/tree/11e957b4e85ff7c867c36811b3bac682b3656fa8).

[22] Gitguy268. [Public GitHub profile repository](https://github.com/Gitguy268/Gitguy268/tree/68b3f7d79f08123ac4130087faec00fb3528853e).

[23] Lucasgitguy-co. [Original Mermaid text dataset](https://huggingface.co/datasets/Lucasgitguy-co/mermaid-text-en-nl). The exact inspected dataset revision is recorded in the companion source manifest.

[24] Lucasgitguy-co. [Mermaid text and image dataset at the training revision](https://huggingface.co/datasets/Lucasgitguy-co/mermaid-diagram-text-images/tree/21bd67796259d0f5c8bf67d00f45f3d59ba8f73a).

[25] Lucasgitguy-co. [Spot Snake dataset and training tools](https://huggingface.co/datasets/Lucasgitguy-co/spot-snake-decisions-v0.1/tree/7f8425c8bd3c91c0443c70ee1b03f0130308dc55). Includes scripts/train_d1.py, scripts/train_mini.py, data/provenance.json, and sources.json.

[26] Scenario Labs. [3D collection at the Forma3D source revision](https://huggingface.co/datasets/scenario-labs/3d-collection/tree/55b4c002b4c1529fc48da3dff98516c3069bfa3f).

[27] Allen Institute for AI. [Objaverse at the Forma3D source revision](https://huggingface.co/datasets/allenai/objaverse/tree/556637099bf4fa79ea7b239d0d9c328b8a2e9ac8).

## Appendix A Complete public model inventory

The common repository namespace is Lucasgitguy-co. The references below pin the inspected snapshots. Creation dates are UTC publication metadata, not training dates. Forma3D Tiny 2 is proposed and therefore is excluded from the count of 18 existing public repositories.

[1] [lfm3-vl-2b-heretic](https://huggingface.co/Lucasgitguy-co/lfm3-vl-2b-heretic/tree/6c1eca1715ad7460d9b28e998f2299a185baa756). Published 2026-10-30. Review revision `6c1eca1715ad7460d9b28e998f2299a185baa756`.

[2] [lfm3-vl-2b-heretic-mlx-4bit](https://huggingface.co/Lucasgitguy-co/lfm3-vl-2b-heretic-mlx-4bit/tree/d1121f749f6a5cdd004a8be2a4ba5dd5b37a5333). Published 2026-10-30. Review revision `d1121f749f6a5cdd004a8be2a4ba5dd5b37a5333`.

[3] [lfm3-vl-2b-heretic-gguf](https://huggingface.co/Lucasgitguy-co/lfm3-vl-2b-heretic-gguf/tree/8b19dd4fc3bc8f7df18fb6665426955177340259). Published 2026-10-30. Review revision `8b19dd4fc3bc8f7df18fb6665426955177340259`.

[4] [minimind-3v-moe-adapter](https://huggingface.co/Lucasgitguy-co/minimind-3v-moe-adapter/tree/bca9243cbf6b393d56c0176185773b58750cf8ba). Published 2026-10-30. Review revision `bca9243cbf6b393d56c0176185773b58750cf8ba`.

[5] [minimind-mermaid-run2](https://huggingface.co/Lucasgitguy-co/minimind-mermaid-run2/tree/1d1ff9a5bef6f00d4b19a8e6b32da50e5490f973). Published 2026-10-30. Review revision `1d1ff9a5bef6f00d4b19a8e6b32da50e5490f973`.

[6] [minicpm-v-4-7-1b-mermaid-adapter](https://huggingface.co/Lucasgitguy-co/minicpm-v-4-7-1b-mermaid-adapter/tree/a67818cafe37bedccaea505d4ff53947b9e7a8eb). Published 2026-10-30. Review revision `a67818cafe37bedccaea505d4ff53947b9e7a8eb`.

[7] [minicpm-mermaid-heretic](https://huggingface.co/Lucasgitguy-co/minicpm-mermaid-heretic/tree/81702aecf62b9ca638266b8cad782e091e8abcb5). Published 2026-10-30. Review revision `81702aecf62b9ca638266b8cad782e091e8abcb5`.

[8] [minicpm-mermaid-heretic-trial65](https://huggingface.co/Lucasgitguy-co/minicpm-mermaid-heretic-trial65/tree/868e2b40df33559194f1e6a09bc2bbbf3a7b988b). Published 2026-10-30. Review revision `868e2b40df33559194f1e6a09bc2bbbf3a7b988b`.

[9] [minicpm-mermaid-gguf](https://huggingface.co/Lucasgitguy-co/minicpm-mermaid-gguf/tree/8c5c74cdbf98e477f3e3cb8ee566bbdabde870b1). Published 2026-10-30. Review revision `8c5c74cdbf98e477f3e3cb8ee566bbdabde870b1`.

[10] [minicpm-mermaid-heretic-gguf](https://huggingface.co/Lucasgitguy-co/minicpm-mermaid-heretic-gguf/tree/a6ad11f60ec7b95c060f4cccdf21eaffdafd4456). Published 2026-10-30. Review revision `a6ad11f60ec7b95c060f4cccdf21eaffdafd4456`.

[11] [minicpm-mermaid-heretic-trial65-gguf](https://huggingface.co/Lucasgitguy-co/minicpm-mermaid-heretic-trial65-gguf/tree/c487fa22912797d5cea03767806c2ee94b5c40d1). Published 2026-10-30. Review revision `c487fa22912797d5cea03767806c2ee94b5c40d1`.

[12] [Qwen3.5-2B-Mermaid-MLX-LoRA](https://huggingface.co/Lucasgitguy-co/Qwen3.5-2B-Mermaid-MLX-LoRA/tree/0b0b4c16c0c1dacffdc38814bddf516a92b0577e). Published 2026-11-01. Review revision `0b0b4c16c0c1dacffdc38814bddf516a92b0577e`.

[13] [Lucas-Image-2.1-Turbo-Q2K-SakuraEncoder](https://huggingface.co/Lucasgitguy-co/Lucas-Image-2.1-Turbo-Q2K-SakuraEncoder/tree/c8450700eb1d684d92e8c4b02767c11651bbca60). Published 2026-11-01. Review revision `c8450700eb1d684d92e8c4b02767c11651bbca60`.

[14] [image3d-student](https://huggingface.co/Lucasgitguy-co/image3d-student/tree/1738cdfaf1bb940bf576a96d11b742ce014b9139). Published 2026-11-02. Review revision `1738cdfaf1bb940bf576a96d11b742ce014b9139`.

[15] [forma3d-tiny](https://huggingface.co/Lucasgitguy-co/forma3d-tiny/tree/e80b8d895cc2e405894473abe4ce40b3472eaca7). Published 2026-11-03. Review revision `e80b8d895cc2e405894473abe4ce40b3472eaca7`.

[16] [spot-snake-d1-2B-v0.1](https://huggingface.co/Lucasgitguy-co/spot-snake-d1-2B-v0.1/tree/baf2bb6a85d3962af091dbf03a44d4377ff57aaa). Published 2026-11-05. Review revision `baf2bb6a85d3962af091dbf03a44d4377ff57aaa`.

[17] [spot-snake-d1-3B-v0.1](https://huggingface.co/Lucasgitguy-co/spot-snake-d1-3B-v0.1/tree/26f6b0dabbf4289021f62701cf03d8f553a4b4df). Published 2026-11-05. Review revision `26f6b0dabbf4289021f62701cf03d8f553a4b4df`.

[18] [spot-snake-minimind-jev-v0.1](https://huggingface.co/Lucasgitguy-co/spot-snake-minimind-jev-v0.1/tree/e3bf59d34b534cec9d373102f4a80c3b3212c58e). Published 2026-11-05. Review revision `e3bf59d34b534cec9d373102f4a80c3b3212c58e`.