Skip to content

Contribute

Contribute to Tokenstead

Local AI is built by the people running it.

Every rig, benchmark, and guide here started as something a reader sent in. Corrections, price reports, and tool tips need no account. Everything lands in a moderator queue before it reaches a page.

135
Guides published
14
Rigs shared
1
Benchmark logged

Ways to contribute

Pick the one that matches what you have. Each opens a form built for that kind of data.

Open calls

Concrete things we are missing right now. Pick one and send what you have; a partial answer still helps.

  1. Benchmark request Open call 01

    DeepSeek V4 Flash on a 24GB desktop card

    The 128GB unified-memory numbers are in. A clean q4_k_m run on a single 24GB card, with prompt processing and the context length you used, is not.

  2. Price report Open call 02

    Your last hosted invoice for an open-weight model

    List prices and billed prices drift once caching and batch discounts land. Send the provider, the line item, and the date on the invoice.

  3. Correction Open call 03

    A guide step that no longer works

    Runtimes, quants, and flags move. If a setup step on a guide failed for you this month, send the page and the error text.

  4. Tool or harness Open call 04

    The harness you actually drive every day

    Name the harness, the model behind it, and the failure that made you switch. We are building the harness catalog from people who run one for real.

  5. Rig Open call 05

    A two-card rig that earns the second card

    Tok/s before and after the second GPU, plus the interconnect and the workload. The per-watt math is the part most build logs skip.

  6. Policy tip Open call 06

    A rule change with a date and a primary source

    Export controls, license terms, data-center rules. Link the document and say what changes for someone running models at home.

  7. Guide pitch Open call 07

    A build log with the part that did not fit

    The return you ate, the riser that did not clear the case, the BIOS setting that saved the boot. That detail saves someone a weekend.

  8. Speech Open call 08

    A 24GB Nemotron Transcribe run

    NVIDIA's 34B speech model ships NVFP4 and GGUF. Run one hour of overlapping audio on a 24GB card with NeMo-Speech.cpp or vLLM and report peak VRAM, seconds of audio per second, speaker and timestamp quality, and WER on your sample.

  9. Source check Open call 09

    Date the GLM-5.5 lecture slides

    A viral post attributes a six-source data pipeline and 32B/122B/256B sizes to a Jie Tang talk this week. The photographed slide footer reads Tsinghua, 2024 InterTalk, and Z.ai has no public GLM-5.5 model card. Find the recording or the official deck and send the date and link.

  10. Models Open call 10

    A small model that replaced a big one

    If a 12B model took over a 70B workload, show the prompt, the output, and the hardware it ran on.

Recent contributors

The latest bylined work from the community. Every one of these is live on the site.

@placeholder-niklaslenz-ai shared a rig

Rig4x DGX Spark 15U rack - Niklas

12 days ago Read

cdnsteve shared a rig

Rig8x RTX 3090 open frame - AxiTek

16 days ago Read

cdnsteve published a guide

GuideAlibaba Cloud puts $3M into Omarchy and names it the OS target for Qwen Book, an ARM-first agentic computer

2 months ago Read

Tobi Lutke shared a rig

RigJeffrey Guenther Rack Cellar

2 months ago Read

Tobi Lutke shared a rig

RigTobi Lutke Rack Cellar

2 months ago Read

@placeholder-typecraft-dev shared a rig

RigTypecraft's homelab

2 months ago Read

cdnsteve shared a rig

RigRyzen AI Max 395+ / Radeon PRO R9700 lab - Martin Mielke

2 months ago Read

cdnsteve logged a benchmark

BenchmarkDeepSeek V4 Flash 0731 on GB10 (DGX Spark)

4 months ago Read

cdnsteve published a guide

GuideRun the Grok CLI on Ollama Cloud and custom providers

4 months ago Read

Start somewhere small

A correction takes two minutes and saves the next reader a dead end. If you would rather write the whole thing, the guide editor is one click away. If you built something and have numbers, the rig and benchmark forms take the raw data.