How little machinery useful local intelligence needs: genesys, C-Genisys and Noema, preregistered, failures included. Research artifacts only.
Joakim Palm
Joakimpalm-Zen
AI & ML interests
Building Xyntetik Runner, a local inference engine, and using it as a research platform for quantization, LoRA on GGUF weights, model surgery and reproducible inference. This profile is the evidence trail: runnable releases, compatibility reports, experiments and failures. If you just want something to run, start with the Runner Releases collection.
Recent Activity
updated a model 1 day ago
Joakimpalm-Zen/Mellum2.1-12B-A2.5B-Thinking-selective-Q8_0-expQ5_K-GGUF published a model 1 day ago
Joakimpalm-Zen/Mellum2.1-12B-A2.5B-Thinking-selective-Q8_0-expQ5_K-GGUF updated a model 3 days ago
Joakimpalm-Zen/Qwen3.8-Flash-Next-MTP-GGUFOrganizations
Dense students distilled from larger parents, released only after a preregistered gate. Kvist-14B: model, GGUFs, training record, checkpoints.
Third-party files run on Xyntetik Runner and compared with llama.cpp and the publisher's reference, each bound to the exact file by SHA-256.
-
Joakimpalm-Zen/granite-4.0-h-small-runner-report
Preview • Updated • 51 -
Joakimpalm-Zen/Nemotron-3.5-Lightning-30B-A3B-runner-report
Preview • Updated • 63 -
Joakimpalm-Zen/Qwen3-speculative-pair-report
Viewer • Updated • 1 • 58 -
Joakimpalm-Zen/NVIDIA-Nemotron-Nano-9B-v2-Q8_0-GGUF
Text Generation • 9B • Updated • 419 • 1
Sublayer removal, healing and distillation-recovered quantization scales, each artifact beside the evidence that measured it. Negative results kept.
-
Joakimpalm-Zen/Muse-Glimmer-30B-surgery-report
Updated • 68 -
Joakimpalm-Zen/Muse-Glimmer-30B-Surgical-Q4_K-GGUF
Text Generation • 28B • Updated • 286 -
Joakimpalm-Zen/Muse-Glimmer-30B-Surgical-Q8_0-GGUF
Text Generation • 28B • Updated • 378 • 1 -
Joakimpalm-Zen/Muse-Glimmer-30B-Surgical-6p34-Q4_K-recovered-GGUF
28B • Updated • 304 • 1
Genesis: preregistered research on machine-only learning in three tracks (language, existence, crossing), failures included. Research artifacts only.
Adapters trained through the quantized GGUF that serves them, byte-reproducible, with the finding that a 4-bit merge erases the fine-tune.
Where expert pruning and selective quantization stop working, measured against the parent at 400 teacher-forced positions. Negative results kept.
-
Joakimpalm-Zen/Nemotron-3.5-Lightning-30B-A3B-prune-frontier-report
Updated • 146 -
Joakimpalm-Zen/Muse-Glimmer-30B-runner-quant-frontier-report
Updated • 117 -
Joakimpalm-Zen/Hermes-4-14B-quant-fidelity-report
Viewer • Updated • 5 • 69 -
Joakimpalm-Zen/gpt-oss-20b-keep30-MXFP4-GGUF
Text Generation • 20B • Updated • 1.46k • 2
GGUF models and adapters meant to be downloaded and served with Xyntetik Runner. Each card states its measured status against the parent.
-
Joakimpalm-Zen/Qwen3-30B-A3B-selective-attnQ8_0-expQ4_0-GGUF
Text Generation • 31B • Updated • 718 -
Joakimpalm-Zen/Qwen3-Coder-30B-A3B-Instruct-keep120-Q4_K_M-GGUF
Text Generation • 29B • Updated • 1.02k • 1 -
Joakimpalm-Zen/NVIDIA-Nemotron-Nano-9B-v2-Q8_0-GGUF
Text Generation • 9B • Updated • 419 • 1 -
Joakimpalm-Zen/gemma-4-E2B-it-Q4_0-GGUF
Text Generation • 5B • Updated • 513 • 2
How little machinery useful local intelligence needs: genesys, C-Genisys and Noema, preregistered, failures included. Research artifacts only.
Genesis: preregistered research on machine-only learning in three tracks (language, existence, crossing), failures included. Research artifacts only.
Dense students distilled from larger parents, released only after a preregistered gate. Kvist-14B: model, GGUFs, training record, checkpoints.
Adapters trained through the quantized GGUF that serves them, byte-reproducible, with the finding that a 4-bit merge erases the fine-tune.
Third-party files run on Xyntetik Runner and compared with llama.cpp and the publisher's reference, each bound to the exact file by SHA-256.
-
Joakimpalm-Zen/granite-4.0-h-small-runner-report
Preview • Updated • 51 -
Joakimpalm-Zen/Nemotron-3.5-Lightning-30B-A3B-runner-report
Preview • Updated • 63 -
Joakimpalm-Zen/Qwen3-speculative-pair-report
Viewer • Updated • 1 • 58 -
Joakimpalm-Zen/NVIDIA-Nemotron-Nano-9B-v2-Q8_0-GGUF
Text Generation • 9B • Updated • 419 • 1
Where expert pruning and selective quantization stop working, measured against the parent at 400 teacher-forced positions. Negative results kept.
-
Joakimpalm-Zen/Nemotron-3.5-Lightning-30B-A3B-prune-frontier-report
Updated • 146 -
Joakimpalm-Zen/Muse-Glimmer-30B-runner-quant-frontier-report
Updated • 117 -
Joakimpalm-Zen/Hermes-4-14B-quant-fidelity-report
Viewer • Updated • 5 • 69 -
Joakimpalm-Zen/gpt-oss-20b-keep30-MXFP4-GGUF
Text Generation • 20B • Updated • 1.46k • 2
Sublayer removal, healing and distillation-recovered quantization scales, each artifact beside the evidence that measured it. Negative results kept.
-
Joakimpalm-Zen/Muse-Glimmer-30B-surgery-report
Updated • 68 -
Joakimpalm-Zen/Muse-Glimmer-30B-Surgical-Q4_K-GGUF
Text Generation • 28B • Updated • 286 -
Joakimpalm-Zen/Muse-Glimmer-30B-Surgical-Q8_0-GGUF
Text Generation • 28B • Updated • 378 • 1 -
Joakimpalm-Zen/Muse-Glimmer-30B-Surgical-6p34-Q4_K-recovered-GGUF
28B • Updated • 304 • 1
GGUF models and adapters meant to be downloaded and served with Xyntetik Runner. Each card states its measured status against the parent.
-
Joakimpalm-Zen/Qwen3-30B-A3B-selective-attnQ8_0-expQ4_0-GGUF
Text Generation • 31B • Updated • 718 -
Joakimpalm-Zen/Qwen3-Coder-30B-A3B-Instruct-keep120-Q4_K_M-GGUF
Text Generation • 29B • Updated • 1.02k • 1 -
Joakimpalm-Zen/NVIDIA-Nemotron-Nano-9B-v2-Q8_0-GGUF
Text Generation • 9B • Updated • 419 • 1 -
Joakimpalm-Zen/gemma-4-E2B-it-Q4_0-GGUF
Text Generation • 5B • Updated • 513 • 2