Corey Higgins

Founder, Wynfall AI
Building local models and agents. Testing what actually runs.

[Corey Higgins @coreywynfall]

@iAmDon_eth Blender plus Unity is the useful stack here. Can it fix one mechanic after the first prompt without wrecking the rest?

[Corey Higgins @coreywynfall]

@RoundtableSpace Four models on one flight-sim prompt still isn't a ranking. The useful follow-up is who recovers when the runway state changes.

[Corey Higgins @coreywynfall]

@MiaAI_lab @mrexodia @CMichaelGibbs @plotarmordev @jvr0x Two useful local AI test lanes here: DGX Spark and RTX 5090. I'd want tokens per second at the same context length before comparing results.

[Corey Higgins @coreywynfall]

@rapidmlx 4B params on 18GB M3 Pro is the part worth copying. Live tool calls on a local Mac are a much more useful constraint than another cloud demo.

[Corey Higgins @coreywynfall]

@googlegemma Raspberry Pi + llama.cpp is the receipt that edge Gemma actually ships. Cleaning speech and routing intent on-device beats cloud-first for local public services.

[Corey Higgins @coreywynfall]

@atomicagent_io Coding mode plus a hard context budget is the right pair. The silent failure mode is still fat tool descriptions eating the window before the repo map does. Audit tools the same way you audit memory size.

[Corey Higgins @coreywynfall]

@yume_arasaki @MiaAI_lab On a 3090 the win is usually memory layout plus the chat template, not a magic quant tag. If decode jumps but tool calls get weird, check the template before you chase another GGUF.

[Corey Higgins @coreywynfall]

@jlucas @exolabs Right read. PAIR schedules whole jobs onto nodes that already have the model in Ollama or LM Studio. Exo-style tensor split across the LAN is a different product. Both help agents; only one of them is fake-shared VRAM.

[Corey Higgins @coreywynfall]

@wallstengine PAIR routes whole jobs to free Ollama or LM Studio nodes. It does not merge VRAM across the LAN. That is the useful line for multi-agent work: parallel subagents, not fake tensor parallel. Are people pairing a gaming PC plus a MacBook first, or two desktops?

[Corey Higgins @coreywynfall]

@thatroblennon The 1% skill budget in Claude Code (about 8k tokens once basics are reserved) is why fat AGENTS.md packs silently vanish. Audit tool descriptions the same way you audit context.

[Corey Higgins @coreywynfall]

@Chromadera @Eschalabs Full 262K context on a stock 7900 XTX in ~20.7GB at ~26 tok/s is the AMD card people forget. A W2 build that still ships a usable chat template is the bar that matters.

[Corey Higgins @coreywynfall]

@Andy_ShuoYang 68.3 tok/s on a single 5090 with NVFP4 and no speculative decode is the honest desktop number. Parking the 51GB n-gram table on NVMe at about 0.5% throughput cost is the trick worth copying.

[Corey Higgins @coreywynfall]

@mfranz_on Breakeven lands fast. At $10 per million output, a 3090 pays for itself around 100M output tokens, which is a few months of daily agent runs.

[Corey Higgins @coreywynfall]

@awnihannun 105 tok/s on a dense 27B is the M5 Max number. Spec decode at 2.9x on MLX is doing the work people keep pinning on bigger MoE.

[Corey Higgins @coreywynfall]

@MiaAI_lab Same 😔 @QwenDevs

[Corey Higgins @coreywynfall]

@jnardiello Great setup!

[Corey Higgins @coreywynfall]

@Da7_Tech Some users are going to claim AGI 🙈

[Corey Higgins @coreywynfall]

@sfxnz Amazing

[Corey Higgins @coreywynfall]

I love machines like this but the memory bandwidth is what usually gets me and is a huge trade off for me. Sure the memory size is good for larger parameters but imo it’s not gonna matter with slower tok speeds or larger projects. However I would love to just use it as an agent box or something to run for my homelab server. Good option for some users tho!!