All Articles

Nex AGI posted Nex-N2.5-mini under Apache 2.0. Official BF16 usedStorage is 70.24 GB of weights. Sep 9 fold: mradermacher GGUFs are real, Q2_K through Q8_0, 12.94 to 36.90 GB. All fit a GX10 as a file-size claim. Official serve is patched SGLang on 2×H100. I have not loaded it.

ggml-org tagged llama.cpp v0.4.0 on Sep 4, 2026. Nightly on that release is b10809. ggml is bumped to v0.23.0. This is a changelog note, not a GX10 sit.

Official FP8 and NVFP4 fit 128GB on disk. llama.cpp PR 28323 merged; first nightly with the expert-layer fix is b10796. I have not loaded this on the GX10. Not a sit.

IFM posted K2 Horizon under Apache 2.0: six sizes from 0.9B to 375B-A23B. Official under-~110 GB GGUF/FP8 siblings fit on paper. 375B FP8 does not. Not a bench. Not a GX10 load.

Microsoft posted streaming ASR weights today: 1.5B usedStorage 5.64 GB, 7B 17.35 GB. MIT. Both fit a GX10 as a file-size claim. This is not TTS, and I have not loaded it.

AngelSlim posted Hy4-preview GGUFs. Smallest card size is 213.66 GiB. Still not a GX10 sit. Stock llama.cpp does not run them.

XHToken posted Spark-X2.5 4B and 1.7B under Apache 2.0. Official 4B BF16 is 8.23 GB of weights. That fits a GX10. Native 1M is claimed. I have not timed it, and this is not NVIDIA Spark.

Unsloth IQ1_S is 82.44 GB. ggml-org Q2_K_S is 98.59 GB. llama.cpp b10762 has PR 28133. Still not a GX10 sit.

Tencent posted Hy4-preview on Aug 28: 770B / 49B active, Apache 2.0. Official usedStorage is 1.56 TB. Official FP8 is 814 GB. No public GGUF. Not a GX10 sit.

Qwen posted Flash-Next on Aug 26: 125B / 6B active, plus 51B n-gram and 4B MTP. Official BF16 is 360 GB. Unsloth GGUFs landed; the smallest is 72.5 GB and still not a GX10 sit.

Open WebUI tagged v0.11.1 on Aug 25. Admins can pause each tool call for allow or deny. Automations and temp chats stay out. Their 1000x stream claim is theirs.

IBM posted Granite 4.2 on Aug 25: Apache 2.0, dense 3B/8B/30B, thinking switch, 128K native. Official 30B BF16 is 58.55 GB of weights. That fits a GX10. I have not timed it.

ggml-org tagged llama.cpp v0.3.0 on Aug 25. It adds dots3-note, GLM-4.5-Air MTP, DeepSeek 4 -sm tensor, and a moov-at-EOF video fix. A GitHub tag is not a GX10 number.

SpaceXAI posted Grok 4.6 on Google's Model Garden Aug 21. Google's card says Preview, grok-4.6, global endpoint, 13 QPM. Same $2/$6 as Bedrock. Still not a bench.

I cheered Qwen 3.8 27B before the weights existed. Then I loaded NVFP4 on the ASUS Ascent GX10 and timed it. Threads, comments, and the numbers from that week.

I told Grok Build to recreate my link-in-bio page from live links. It published wikiwayne.grok.me. Zero monthly fee, full control — this is what 'just build it' looks like in 2026.

I run every X post through Grok's Enhance button before I hit send. It catches omitted words and missed punctuation. Here is why that tiny feature matters more than another chatbot.

Fantastic pricing is credible. Calling Grok 4.6 a solid Sonnet-class model on day one, ahead of Gemini, is not — not without real metrics. How I read launch-day AI posts.

I posted Qwen 3.8 27B three times and meant it. Why a 27B open-weight model still matters on a 128GB Spark-class box, and how I would actually load it.

Hermes agent, DeepSeek V4 Flash 0731, ASUS Ascent GX10 — local only. How I shipped the Moon Trail browser game without cheating on a cloud model.

I unboxed an ASUS Ascent GX10 — NVIDIA DGX Spark in a 150mm cube — and started running local models on 128GB of unified memory. No brochure. First notes from the bench.

Cornerstone WikiWayne guide: Best GPU for Local AI (2026). Open-weight, practitioner-tested local AI.

Shopping the secondary market without overspending on VRAM you cannot use.

Slower tokens, stronger air-gap story.

What `-ngl` / GPU layer sliders actually do.

Compose services for local chat without cloud relay.

Quant-specific VRAM bands for Meta Llama 3 8B class models.

Step-by-step Ollama install paths for the three major desktop OS families.

Compile with GPU backends for NVIDIA cards.

Cornerstone WikiWayne guide: llama.cpp Complete Guide. Open-weight, practitioner-tested local AI.

Leave the managed service when you need custom builds or flags.

Use the LM Studio catalog without guessing quant labels.

Cornerstone WikiWayne guide: LM Studio vs Ollama vs llama.cpp: Which Local AI Tool?. Open-weight, practitioner-tested local AI.

Cornerstone WikiWayne guide: Local AI Model Tracker (2026). Open-weight, practitioner-tested local AI.

Network egress, logging, and backup habits for homelabs.

CUDA maturity vs ROCm tradeoffs for GGUF stacks.

Point agents and UIs at `http://localhost:11434/v1`.

Cornerstone WikiWayne guide: Open WebUI for Local AI. Open-weight, practitioner-tested local AI.

Docker and bare-metal pairing for household chat.

From zero to a working chat with a small Ollama tag.

When to spend extra gigabytes on higher precision.

Cornerstone WikiWayne guide: Quantization Explained for Local AI. Open-weight, practitioner-tested local AI.

Cornerstone WikiWayne guide: Run Open-Weight Models Locally (2026). Open-weight, practitioner-tested local AI.

Cornerstone WikiWayne guide: VRAM Requirements for Local LLMs. Open-weight, practitioner-tested local AI.

GGUF packs tensors and metadata for llama.cpp-compatible runners.

Ollama favors CLI and API automation; LM Studio favors GUI model browsing. Compare setup, VRAM use, and GGUF workflows on real hardware.