H2LooP

Civilization runs on system software. It cannot fail. Most AI coding tools were not built for this domain. H2LooP was. #h2loopai #EmbeddedSystems #SystemSoftware #AIInfrastructure

H2LooP Spark CPT (Preview) is available now on HuggingFace under a Research Only License. Works with vLLM and 🤗 Transformers. Single H100, bfloat16. This is an early checkpoint Paper → arxiv.org/abs/2603.11139 Request access → huggingface.co/h2loop-ai/spar…

Introducing H2LooP Spark: the first domain-specialized autocomplete model for embedded software. A 7B model that beats Claude Opus 4.6 and Qwen3-Coder-30B on embedded code completion. Not fine-tuned. Continually pre-trained on 23B tokens of firmware, datasheets, and vendor SDKs

We built SpecMap: an agentic pipeline that maps vendor datasheets directly to code symbols, across 13 embedded domains 100B raw tokens curated down to 23B. The result: a model that knows the exact register offset, the exact intrinsic opcode, and the exact pin mapping.

General LLMs fail at embedded code because - Infineon TriCore intrinsics - NXP eDMA scatter/gather docs, and - AURIX ATOM timer pin maps simply don't exist in standard pre-training data.

Token accuracy on held-out embedded code (13 domains, 9 repos never seen during training): → H2LooP Spark 7B: 34.1% → Qwen3-Coder-30B: 24.6% → Claude Opus 4.6: 24.1% → Base OLMo-7B: 16.8% +108% over base. Leads frontier models that are 4–50x larger.

#EmbeddedSystems #FunctionalSafety #MISRA #SafetyCritical #AIResearch #SLM #DeepTech #h2loopai

MISRA compliance is mandatory in safety-critical C. We built a compact SLM to automate it. Then benchmarked against frontier models 100x its size.

Full blog: beyond.h2loop.ai/autotr If this is the kind of problem you want to work on: traceability, embedded systems, AI that reasons about hardware then we are hiring. beyond.h2loop.ai/hiring

Runs locally. No code leaves your environment. Domain-specialized SLMs matching frontier models at 1/1000th the size.

Rule family performance: Pointer safety (11.x, 18.x): 100% fix rate Control flow (15.x): matches Gemini Pro, fewer edits Initialization (9.x): parity with domain experts Type model (10.x): 67% fix rate

The real test: fix violations without rewriting the codebase. Character delta on fixes: Sanitizr: ~12% Gemini 2.5 Flash: 25-31% Surgical edits. Near-identical to expert corrections.

Full blog: beyond.h2loop.ai/convrg If this is the kind of problem you want to work on: traceability, embedded systems, AI that reasons about hardware then we are hiring. beyond.h2loop.ai/hiring

Common assumption: full fine-tuning performs better, LoRA is the efficient trade-off. We tested this. The assumption is wrong.

We independently reproduced the Thinking Machines Lab "LoRA Without Regret" findings. The results hold across our evaluation. #MachineLearning #FineTuning #LoRA #LLM #AIResearch #ModelTraining #DeepTech #h2loopai

The more relevant weight matrices you apply LoRA to, the better the result. The instinct to restrict LoRA to attention layers is leaving performance on the table.

With proper hyperparameter setup, LoRA outperforms full fine-tuning while keeping its compute advantages. What "proper" means: Learning rates 10x higher than full SFT Higher rank selection LoRA on all weight matrices, not just attention and proper warmup scheduling + weight decay

Full paper: beyond.h2loop.ai/convrg If this is the kind of problem you want to work on — traceability, embedded systems, AI that reasons about hardware — we are hiring. beyond.h2loop.ai/hiring

None of it was obvious before the runs.

High-rank LoRA beats low-rank. Domain-only data beats mixed corpora. Full-module targeting is optimal.