Quant Factory
Turn source checkpoints into practical local deployments and measure the quality cost instead of guessing.
- Q8 / Q5 / Q4 controls
- Hashes + toolchain provenance
- Quality, speed, memory, size
Independent local AI research
Local models,
measured in the open.
Quants, visual stress tests, and reproducible inference across Apple Silicon and NVIDIA DGX Spark. The hardware changes. The evidence standard does not.
Formalizing the automation layer now
LOCAL COMPUTEREPRODUCIBLE OUTPUT
2 architecturesApple + NVIDIA
76 scenariosStable control suite
Originals keptRepairs stay separate
2× dailyAutomation target
Three research lanes
One lab, not one machine.
The DGX cluster is one part of the system. Mac Studio testing, quantization research, model archiving, browser validation, and public evidence all belong under the same lab identity.
Visual Arena
Give local models the same uncapped creative task, preserve every token, and test what actually renders.
- Matched prompts + release
- Browser and shader validation
- Interactive pages + videos
Deployment Bench
Treat engine, quant, context, and speculative decoding as part of the measured system.
- Apple Silicon + DGX Spark
- Tool use + long context
- Qualified evidence only
Compute fabric
Use each system for what it does best.
No artificial hardware hierarchy. Apple Silicon, NVIDIA Grace Blackwell, and network storage each answer a different part of the local-model question.
Apple Silicon
Mac Studio
Local inference, MLX and llama.cpp experiments, grading, browser validation, orchestration, and media production.
Independent single-system lane
Grace Blackwell
4× DGX Spark
NVFP4 and FP8 serving, long-context tests, tensor-parallel deployments, speculative decoding, and multi-node research.
One, two, and four-node lanes
Research archive
UGREEN NAS
Source checkpoints, verified quants, immutable outputs, hash manifests, videos, and evidence that survives the active job.
20 TB source of truth
Recent visual research
The output is part of the evidence.
These are not screenshots generated for a post. Each scene began as one model response, was preserved as HTML, opened in a real browser, and published with repairs disclosed.
Visual Arena · Deployment Scale
Flash vs Full 743B
The same uncapped solar-system brief on two-Spark TP2 and four-Spark TP4 deployments.
Open experiment ↗
Visual Arena · WebGL
Galaxy Battle
Two long-reasoning local models build complete interactive universes.
Open experiment ↗
Visual Arena · Metamorphosis
Origami Dragon
Metallic paper folds, transforms, flies, and breathes fire.
Open experiment ↗
Visual Arena · Simulation
Storm in a Glass Cube
Two models turn a calm museum object into a living supercell.
Open experiment ↗Evidence chain
Reproducible from source to screen.
A result is only useful when the checkpoint, quant, runtime, request, raw stream, browser behavior, and any intervention can all be traced.
Pin
Model revision, license, tokenizer, and hashes.
Configure
Quant, engine, context, topology, and thinking policy.
Run
Matched request, uncapped output, synchronized release.
Preserve
Raw stream, reasoning, answer, usage, and original HTML.
Validate
Browser errors, shaders, motion, framing, and grading.
Disclose
Publish evidence and keep display repairs separate.
Visual standard
Original means original.
Model output is never silently repaired. If a small display correction is needed, the untouched artifact remains available and the repair gets its own copy, diff, badge, and disclosure.
Finish reason recorded
Token classes separated
JavaScript checked
Blank frames detected
Motion verified
Repair diff preserved
Quant standard
Loading is not validation.
A quant must survive tokenizer checks, tool use, instruction following, long-context retrieval, coding, visual generation, and comparison with a higher-quality control.
Source revision pinned
Files hashed
Quant recipe saved
Memory measured
Speed measured
Quality tax reported
Lab roadmap
From experiments to an operating system.
The public work already exists. The next step is connecting it with a safe controller, node leases, twice-daily jobs, and a consistent review package.
NOWEstablished
Public research archive
SparkBench evidence, visual battles, reproducibility guides, and deployment notes.
PHASE 1Building
Controller and node leasing
One job registry across Mac Studio, DGX Spark, NAS, and multiple agents.
PHASE 2Planned
Twice-daily Visual Arena
Permanent control prompt, rotating challenge, browser QA, video, and Telegram review.
PHASE 3Planned
Quant Factory
Verified GGUF and native NVIDIA lanes with quant-tax reports and NAS archiving.
Open evidence
Run it yourself.
The portable visual harness and benchmark history are public. Machine-specific infrastructure stays private; methods and results do not.