Making clankers think and GPUs go BRRR
LLM engineer at the National Library of Norway, where I help pretrain, post-train and evaluate Norwegian large language models, most recently the Borealis family. Off the clock I run Heiervang Technologies, an open-source lab for AI agents, inference tooling, and forks of the libraries I use daily. MSc in Informatics (NLP) from the University of Oslo. I stream a lot of the process on Twitch and YouTube.
Borealis: The National Library of Norway's newly released family of Norwegian language models. Built on Gemma 3, spanning 270M to 27B parameters, vision-language capable, fluent in Bokmål and Nynorsk, and shipped in open and full-data variants with GGUF and MLX builds. I'm part of the team behind it.
unleash: One TUI/CLI for all your agent harnesses. (⭐ 14)
I get brand-new models running on open inference stacks, then handle the lifecycle around them: quantize, serve, evaluate, and wrap in agents.
- Model enablement & inference: New architectures (GLM-4.7, Qwen3-Omni/TTS, GPT-OSS) brought up in vLLM, vllm-omni, and Hugging Face Transformers.
- Quantization: FP8, MXFP4, and GGUF to squeeze large models onto smaller and non-NVIDIA GPUs (Vulkan, ZLUDA).
- Speech, voice & audio: Speaker/voice embeddings, TTS, ASR, music generation.
- LLM training & fine-tuning: Pre/post-training Norwegian LLMs, LoRA, prompt baking, synthetic data.
- Evaluation & benchmarking: EuroEval, lm-eval-harness, NorEval, MTEB.
- AI agents & dev tooling: Agent orchestration, context compaction, CLI tools.
Getting brand-new models and fixes into the core ML stack:
- GLM-4.7-Flash → vLLM: Taught vLLM's MLA detection about
glm4_moe_liteso the model runs on the engine. vllm#32614 - Qwen3-TTS → vllm-omni: Added speaker-embedding passthrough for the speech & voices APIs, a
speakeralias, and a fix for the Qwen3-TTS-0.6B profile-run hang. #1227 · #2424 · #1082 - GPT-OSS MXFP4 quant → Transformers: Fixed the MXFP4 quantizer to handle variable
num_local_experts/hidden_size, so GPT-OSS models quantize correctly. transformers#41795 - OpenAI-compatible API eval → lm-evaluation-harness: Fixed the assertion error that broke evaluation against OpenAI-compatible endpoints. lm-eval#3356
Plus smaller fixes to NVIDIA/Megatron-LM (mcore example training loop), EuroEval (fp16 loading for bf16 models on pre-Ampere GPUs), and ltgoslo/noreval.
A running index of my public work. (HT = under Heiervang Technologies.)
Agents & LLM tooling
- am-i-openai-compatible: Probe any HTTP server for OpenAI-API compliance, with an honest matrix of how OSS implementations actually behave. HT · ⭐ 16
- supercompact: Harder, better, faster, stronger context compaction for AI agents. HT · ⭐ 8
- sorting-hat: AI file-naming CLI with animated ASCII art; works with any OpenAI-compatible LLM. ⭐ 16
- bakery: Prompt baking & context distillation for LLMs. ⭐ 5
- chatfiles: The dumbest and smartest way to make agents communicate. ⭐ 2
- syntk: Synthetic-data toolkit.
- exex: Expert exchange.
- markutils: Python utilities I reach for constantly.
Local inference & models
- localllama-with-vulkan: Run LLM inference on any GPU; tutorials and benchmarks.
- replications: Install diaries, container images, and guides for running or training the model that just dropped.
- norbert3-modernbert-conversion: Convert NorBERT3 weights into a ModernBERT architecture.
- ollama-modelfiles: A collection of Ollama Modelfiles.
- ht-* forks: Maintained forks of
vllm-omni,llama.cpp,pytorch,vllm,unsloth,mergekit,ACE-Step,codex. HT
Speech, audio & vision
- semantic-sound-search: Semantic search over audio samples.
- spritegrid: Corrects AI-generated pixel art by applying grid sampling. ⭐ 3
- autonomous-pokerogue: Turns PokéRogue into an idle clicker driven by vision-language models.
MCP & data
- dhlab-mcp: MCP server for the National Library's DH-lab.
- ssb_mcp: MCP server for Statistics Norway (SSB).
Writing, web & meta
- Advent-of-AI-papers: An advent calendar for AI and NLP research papers.
- opensourcefriendly.dev: An index of open-source-friendly companies and developer integrations.
- python-template: My Python project template.
- dotfiles-public: My dotfiles, with an optional private extension.
- studies-portfolio: Compiled programming work from my CS/NLP studies.
A live index of every domain pointed at the placeholder slam page. All of them serve the same animated frontend; the keys of sites.js are the source of truth, and the list below is regenerated daily from it.
bomboc.lat · clanker.cam · clanker.foundation · clanker.lifestyle · clanker.productions · clanker.surf · clanker.wiki · cssus.live · hei.ceo · hei.codes · hei.expert · hei.tools · heiervang.com · heiervang.tech · hover.dog · hoverboard.dog · hyperpla.net · marku.sh · markus.consulting · markus.sh · markus.solutions · markus.work · opensourcefriendly.dev · ossfriendly.dev · pkill.games · pkill.io · pkill.live · pkill.sh · pkill.space · pkill.team · retard.mx · rugpull.in · savethe.click
33 domains · synced from sites.js.
31 models · 7 datasets · 6 spaces · 490 likes · 220k+ all-time downloads
Live all-time download counters (auto-updating):
Models
- GLM-4.7-Flash-FP8: FP8 quantization · 140k downloads, 19 likes.
- Qwen3-Voice-Embedding-12Hz (0.6B & 1.7B): Speaker/voice embeddings · ~34k downloads, 48 likes.
- Qwen3-Omni-30B-A3B-FP8: FP8 any-to-any omni model · 30k downloads.
- LCO-Embedding-Omni GGUF series: Multimodal embeddings in GGUF.
- modern-norbert3: Norwegian ModernBERT conversion.
- t5-base-define: Definition generation.
Datasets: wordnet-definitions-en-2021 (11 likes) · clickbait_title_classification (6 likes)
Spaces: transformer-training-visualized · explore-embedding-inversion · saved-you-a-click
- Clickbait spoiling / "Saved You A Click": Abstractive title-answering. MSc thesis line: the reddit-syac dataset, pegasus/t5 spoiler models, a Gradio demo, and a browser extension.
- Definition modeling: Generating dictionary definitions with T5 + WordNet.
- Norwegian NLP: NorBERT/ModernBERT conversions and benchmark work.
I build a lot of this live on twitch.tv/hyperplane_t. Come say hi.
