capy-flash v0.7.1 runs a 125B MoE on 64GB of RAM
10/22/2026 — 10/22, 22:19·1 sources (all)·1 reports (all)
Key takeaways
- A developer released capy-flash v0.7.1, an open-source inference engine that runs Qwen3.8-Flash-Next, a 125B MoE model, on 64GB of RAM plus…
Background & analysis
On October 22, 2026, a developer announced capy-flash v0.7.1 on Reddit's LocalLLaMA community, describing it as an open-source inference engine that can run Qwen3.8-Flash-Next, a 125B MoE model, on 64GB of RAM with one or two 16GB consumer RTX 50 cards. According to the post, running this model previously required the developer's own setup of two 16GB cards plus 96GB of RAM, or something better. The v0.7 line targets smaller machines: users can start with the hardware they already have and add a second card later, keeping the same engine and the same model files.
The post reports roughly 234 tok/s with two cards. A 3-bit version of Qwen3.8-27B runs on a single 16GB card at about 174 tok/s. The quantized weights reuse ISTA-DASLab's GSQ-RCO scheme at roughly 3 bits; they are copied as-is rather than re-quantized. So far this single release post is the only source, and no further developments have been reported.
AI-generated from 1 reports · updated 14 hours ago
Latest turnThe developer of capy-flash shipped v0.7.1, an open-source inference engine that runs Qwen3.8-Flash-Next (125B MoE) on 64 GB of RAM with one or two 16 GB consumer GPUs, reaching about 234 tok/s on two cards. Qwen3.8-27B at 3-bit now fits on a single 16 GB card at roughly 174 tok/s. It reuses ISTA-DASLab's GSQ-RCO ~3-bit quants as-is rather than requantizing.
Related AI tools
- Heat index
- 66
- All sources
- 1
- All reports
- 1
- First seen
- 14 hours ago
Reports on this story headlines open the original
The developer of capy-flash shipped v0.7.1, an open-source inference engine that runs Qwen3.8-Flash-Next (125B MoE) on 64 GB of RAM with one or two 16 GB consumer GPUs, reaching about 234 tok/s on two cards. Qwen3.8-27B at 3-bit now fits on a single 16 GB card at roughly 174 tok/s. It reuses ISTA-DASLab's GSQ-RCO ~3-bit quants as-is rather than requantizing.
Reddit · LocalLLaMAAI score 59
Other stories people are talking about
- Heat index 30Developer benchmarks Qwen3.8-Flash-Next vs Opus 5.5 on Rust coding1 sources
- Heat index 152SurgeQwen launches N1 Pro AI glasses and clip-on AI earbuds2 sources
- Heat index 80Reddit User Charts Scores vs. Size for 11 Open-Weight Models1 sources
- Heat index 80Price list compares per-million-token costs across Qwen API models1 sources
- Heat index 42User logs Qwen3.5-4B failure modes across 10 agent harnesses1 sources
How is heat calculated?About the methodHide
Heat counts how many independent sources covered a story in the last 48 hours: one source counts once no matter how many posts it published, decaying with a 24-hour half-life. What ranks first is what many people are talking about.
This page aggregates public feeds. Headlines and summaries are machine-organized and remain the property of the original authors; verify important facts at the source. If you believe a headline or summary infringes your rights, email the contact address in the footer with the page URL and basis of your claim, and we will remove or replace it after review.
- Surge
- Discussion rising fast
- New
- First report within 6 hours
- Rising
- Still gathering discussion