Nemotron-3.5-Bolt-GameDev

A locally trained GameSpec configuration specialist based on NVIDIA Nemotron 3.5 Bolt E4B. It produces bounded JSON for the included tested browser game engines: Flappy, Breakout, 3D Runner and 3D Collector. The engines use JavaScript Canvas and Three.js. Assembling the JSON with these engines produces a self-contained offline HTML game.

This is a community fine-tune. It is not an official NVIDIA model. It specializes in configuring four defined game types; arbitrary game-code generation, new mechanics, visual design quality and general game development have not been qualified. The engine/compiler is a separate part of the package and is not embedded in the model weights.

Run with Ollama

Download this repository, including the GGUF, and run these commands from its directory with a recent Gemma 4 capable Ollama:

ollama create Nemotron-3.5-Bolt-GameDev -f Modelfile
ollama run Nemotron-3.5-Bolt-GameDev

For a playable game, use the included Python standard-library client. It requests schema-constrained JSON, strictly validates it, and assembles a trusted engine:

python make_game.py --archetype flappy --prompt "Create a flappy game titled Copper Flight. Use sunset theme and seed 42. Use default gameplay settings." --out copper-flight.html

Open the resulting HTML in a modern browser. The file includes its dependencies and needs no model or internet connection to play. The generator also writes the specification and Ollama response beside the HTML. Supported --archetype values: flappy, breakout, runner, collector.

Interactive Ollama chat produces JSON. Use make_game.py to turn it into a game. The client supplies the training system prompt and archetype-specific schema; unrestricted chat has weaker constraints. Unsupported requested mechanics cannot be represented and may be omitted by the model. Check the specification against your request.

Files and export format

File	Purpose
Nemotron-3.5-Bolt-GameDev-Q8_0.gguf	Standalone merged model; no separate base download needed
Modelfile / system.txt	Ollama import settings and specialization instructions
make_game.py, engines/	Strict Ollama client, schema, renderer, Canvas and Three.js engines
adapter/	Original PEFT adapter and F16 GGUF LoRA for reproducibility
data/	Original synthetic specification conversations and fixed train/validation split
export-integrity.json	Base/adapter provenance, merge equation, per-tensor integrity checks
provenance/	Training details, original model cards and original evaluation audit summaries
ollama-validation.json	Deployment-specific regression results; separate from original evaluation

The filename uses Q8_0 to identify the precision of the 258 adapted matrices. It is a mixed-quantization export: these matrices are Q8_0(dequantize(evaluated base) + 2 * B_f16 @ A_f16). All 462 other tensors retain their original types and bytes, including original Q8 embeddings, remaining Q4_K/Q6_K matrices, BF16 and F32 tensors. It is not an all-Q8 quantization of a BF16 merge. Tokenizer, architecture and native chat-template metadata are preserved. The merge uses the same quantized base that was evaluated with the separate adapter, minimizing changes unrelated to the deployment export.

Base provenance:

• Training: nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-BF16, revision 49843eed8db1f1c616c47af75cfd36e07aec0667.
• Evaluated GGUF: nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB, revision 1db54d5b12533ff1bcaa8e42cf4c667988a1f1be.
• Underlying architecture: Gemma 4 E4B, approximately 8B total parameters / 4B effective.

Training

The selected LoRA trained locally on an RTX 5070 12 GB using 256 original synthetic conversations: 208 training / 48 validation. Rank 8, alpha 16, 17,440,768 trainable parameters, 64 optimizer steps, learning rate 0.0001, maximum sequence length 2048, NF4 base and BF16 compute. Loss was masked to assistant completions. Related validation token loss changed from 0.0999496356 to 0.0001186695. These parameterized examples share four fixed output formats; the loss does not establish general game-development ability. No paid training or inference service was used.

See provenance/training.json, the dataset and split for recorded versions, hashes and settings. The PEFT adapter config references the original BF16 training base. The merged GGUF instead uses the evaluated quantized native base described above; those are distinct artifacts.

Evaluation and limits

The original Q4 base + separate F16 LoRA + schema + strict validator + frozen engines achieved 32/32 exact configuration matches versus 21/32 for the base on paired related held-out requests. There were 11 specialist-only successes and zero baseline-only successes; exact two-sided McNemar p = 0.0009765625. This evaluates requested configuration, not new mechanics or human enjoyment.

The original system's second and final amended end-to-end suite passed 172/172 cases (86/86 in each dimension). A pass required both exact requested specification values and all automated engine checks. That suite covered only its prospectively specified explicit all-parameter request distribution, four known archetypes, fixed Validation Game title, bounded numeric/theme/seed mixture and recorded local Edge environment. There were no model repairs, fallbacks or discarded cases.

The first frozen harness failed (2D 6/86; 3D 0/86). Timing waits and an offscreen touch coordinate were corrected in a separately frozen amendment, and wholly fresh requests were drawn. Engine code did not change. Both attempts remain reported; they were not pooled. The four reported one-sided rates each received alpha 0.0125, giving the original second-suite rates 95.0322576% lower bounds with simultaneous confidence at least 95%, under the amended protocol's assumptions. Summaries, correction protocol and audit hashes are in provenance/. These are original-run audit summaries with private workspace prefixes redacted, not a full standalone reproduction of all browser evidence.

Those confidence bounds do not apply to this merged GGUF, Ollama deployment, unrestricted prompts, arbitrary games, human aesthetic judgments or other devices. The merge introduces Q8 rounding and Ollama changes the runtime. Deployment regression checks are reported separately in ollama-validation.json; they reuse earlier requests and do not establish an independent reliability guarantee. No near-guarantee of general game-making ability is claimed.

This export passed 32/32 exact configuration regression checks on Ollama 0.40.8 and 4/4 normal offline-file browser integration smokes covering assembly, titles, start/pause/resume/restart, mobile fit and no remote requests or JavaScript errors. The packaged command-line client also generated the requested Copper Flight specification and HTML successfully. The initial browser smoke incorrectly assumed that Canvas Restart immediately starts a round; Canvas intentionally returns to Ready. That test assertion was corrected, with the initial report retained in provenance/ollama-browser-smoke-initial.json. Model weights, engine code, specifications and generated HTML did not change. See ollama-browser-smoke.json for the corrected checks. These integration smokes do not repeat the original full gameplay validation or measure human quality.

Like its base model, this small specialization can misunderstand requests, omit unsupported requirements or produce unsuitable output. The included client rejects invalid JSON/specifications without silently repairing or substituting a game. Valid JSON alone does not establish that the user's intent was followed. The limited themes and fixed engines do not provide arbitrary artwork, assets, genres or mechanics.

Licensing and origin

NVIDIA model materials are supplied under the OpenMDW License Agreement 1.1. See NOTICE, the original model cards, Apache 2.0 for the underlying Gemma materials, and the included Three.js MIT license. The original adapter, new dataset and game-engine additions are supplied under OpenMDW-1.1. Retain applicable license and origin notices when redistributing.