metadata
title: README
emoji: ⚙️
colorFrom: indigo
colorTo: gray
sdk: static
pinned: false
RESMP.DEV
Evidence-first model systems for real hardware
Inference engineering · calibrated compression · code intelligence · agent traces
RESMP.DEV is an independent applied research and engineering organization focused on making capable models measurable, reproducible, and useful across heterogeneous hardware. We publish quantized weights, native inference artifacts, retrieval encoders, accessibility models, and agentic datasets—together with the evidence needed to evaluate them.
What we work on
| Area | Work |
|---|---|
| Low-precision inference | Architecture-aware NVFP4, FP8, MXFP4/MXFP8, GPTQ, and mixed-precision releases for CUDA and Apple Silicon. |
| Code intelligence | Multilingual code-retrieval encoders, calibrated quantization, and OpenAI-compatible local serving. |
| Agentic systems | Large-scale reasoning and coding traces, continuous distillation, model mutation, and adversarial workload research. |
| Accessible computing | Speech recognition and models trained for accessibility-centered tasks. |
Featured releases
Frontier inference
- Qwen3-Next 80B-A3B NVFP4: Instruct · Thinking
- GLM-4.7-Flash mixed precision: Trellis · Marlin MMFP4
- MiMo-V2.5 ASR FP8: Mandarin, English, Cantonese, and code-switched speech
Code retrieval
- Jina Code Embeddings 1.5B: calibrated block-GPTQ MXFP4 for MLX
- LFM2.5 Code Encoders: 230M and 350M releases in BF16, MXFP8-GPTQ, and MXFP4-GPTQ
Models and data for difficult workloads
- Fable GPT-5.5 Distillation Traces: public agentic reasoning, code, math, and science traces
- LLaMutation Qwen2.5 14B: experimental code translation and completion
- Accessible Qwen 4B: accessibility-centered model · dataset
How we publish
Our model cards aim to distinguish measured results from hypotheses. Where the artifact permits, releases include:
- pinned upstream revisions and derivative-license attribution;
- matched BF16, native round-to-nearest, or architecture-appropriate controls;
- explicit hardware and runtime boundaries;
- held-out evaluations, hashes, manifests, and machine-readable receipts;
- limitations and negative results—not just headline numbers.
Experimental means experimental. Check each repository's model card, required runtime, license, and validation scope before deploying it.