Releases·Tool Release
R9V Update: ~100 tok/s TG on Qwen3.8 Flash Next IQ4_XS, dual R9700 + 128GB RAM; crash fixes via n-gram SSD streaming, Q4_K_XL support
The R9V developer released an update improving RDNA4 inference performance and stability, reporting ~100 tok/s in llama.cpp text generation on Qwen3.8 Flash Next IQ4_XS across two R9700 GPUs with 128GB RAM. The update fixes crashes via n-gram SSD streaming and adds Q4_K_XL support. The developer says further tuning will pause while they ship a separate inference engine and a research/site-builder app.
76 engagement1 sourceMon, Sep 14, 2026, 03:01 AM
AI Overview
- R9V fixes crash instances via n-gram SSD streaming and improved diagnostics; dev torture-tested ~12 hours with no instability
- Q4_K_XL quant now supported but needs more fine-tuning, yielding ~50 tok/s
- Dev is also building a separate inference engine and a deep research/site builder app, ~6 months in, before refocusing on R9V
- Launch80 discord reports even better RDNA4 numbers on other quants; configs nearing ceiling where model isn't fully VRAM resident
Related
Social·Community
Reddit build shares 4x Radeon R9700 Threadripper server serving Qwen 3.8 27b and DSV4 Flash to a 10-person startup
A Reddit user shares a client build for a 10-person startup: a Threadripper 9970x machine with four AMD Radeon R9700 AI Pro GPUs (128GB total VRAM) running a fork of Radiance to serve Qwen 3.8 Next flash, 27b, and DSV4 Flash. The post details throughput numbers at multiple context lengths, showing on-prem multi-GPU AMD setups can sustain 16 concurrent sessions within a 1600W power budget.
A Reddit user shares a client build for a 10-person startup: a Threadripper 9970x machine with four AMD Radeon R9700 AI Pro GPUs (128GB total VRAM) running a fork of Radiance to serve Qwen 3.8 Next flash, 27b, and DSV4 Flash. The post details throughput numbers at multiple context lengths, showing on-prem multi-GPU AMD setups can sustain 16 concurrent sessions within a 1600W power budget.
1.3k engagement1 sourceredditSun, Oct 11, 2026, 12:20 AM
Social·Community
Strata Qwen 3.8 Flash Next matches Qwen 3.5-35B-A3B speed, outperforms 27B, runs on 2x3090s
Reddit users report running Qwen 3.8 Flash Next through the Strata engine on modest hardware, using a community Q4_K_XL quant by bitlamas on Hugging Face. One user gets 80-110 tok/s on dual 3090s, claiming Flash Next matches Qwen 3.5-35-A3B speed with higher intelligence than Qwen 3.8 27B. Another runs a q2_0 quant on a 2060 laptop with 6 GB VRAM, achieving 10 tok/s at 50k context.
Reddit users report running Qwen 3.8 Flash Next through the Strata engine on modest hardware, using a community Q4_K_XL quant by bitlamas on Hugging Face. One user gets 80-110 tok/s on dual 3090s, claiming Flash Next matches Qwen 3.5-35-A3B speed with higher intelligence than Qwen 3.8 27B. Another runs a q2_0 quant on a 2060 laptop with 6 GB VRAM, achieving 10 tok/s at 50k context.
Updated Sat, Oct 3, 2026, 11:44 PM — A second user confirms Flash Next runs at 10 tok/s at 50k context on a 2060 laptop with only 6 GB VRAM via Strata.
12 engagement2 sourcesredditSat, Oct 3, 2026, 02:05 PM
Social·Community
Qwen3.8 Flash Next on 5060 Ti 16GB - 55 tok/s average, and a few demos
A Reddit user tested qwen3.8-flash-next-coder-iq1_m on an RTX 5060 Ti 16GB with 32GB RAM, getting ~55 tok/s generation and ~1.5k tok/s prompt processing. One-shot demos — a landing page and an interactive 3D globe — produced output the user rated above typical local Q3 models.
A Reddit user tested qwen3.8-flash-next-coder-iq1_m on an RTX 5060 Ti 16GB with 32GB RAM, getting ~55 tok/s generation and ~1.5k tok/s prompt processing. One-shot demos — a landing page and an interactive 3D globe — produced output the user rated above typical local Q3 models.
7 engagement1 sourceredditTue, Oct 6, 2026, 05:43 PM
Releases·Tool Release
Developer releases R9V custom RDNA4 kernels, tripling Qwen3.8-Flash-Next speed on dual R9700s
A developer released R9V, a set of hand-tuned kernels for AMD RDNA4 (R9700) GPUs, plus quantized model packages on Hugging Face. Applied to vLLM-Radiance, the kernels push Qwen3.8-Flash-Next to 78 tok/s TG256 on two R9700s, roughly 3x faster than before. The release targets local-inference practitioners on AMD hardware and includes an early-stage inference engine for dense models.
A developer released R9V, a set of hand-tuned kernels for AMD RDNA4 (R9700) GPUs, plus quantized model packages on Hugging Face. Applied to vLLM-Radiance, the kernels push Qwen3.8-Flash-Next to 78 tok/s TG256 on two R9700s, roughly 3x faster than before. The release targets local-inference practitioners on AMD hardware and includes an early-stage inference engine for dense models.
49 engagement1 sourceredditMon, Aug 31, 2026, 01:13 AM
Social·Community
zzpanic ships final vllm-radiance build for single R9700 GPU running qwen3.8-27b as StillDeadcode's new radiance engine supersedes old builds
A community developer (zzpanic) published the final revision of a personal vllm-radiance inference build targeting a single AMD R9700 GPU, as StillDeadcode's standalone radiance inference engine supersedes old vllm-radiance builds. The build adds startup caching and a KV-offload feature enabling long-context multi-agent work on one card.
A community developer (zzpanic) published the final revision of a personal vllm-radiance inference build targeting a single AMD R9700 GPU, as StillDeadcode's standalone radiance inference engine supersedes old vllm-radiance builds. The build adds startup caching and a KV-offload feature enabling long-context multi-agent work on one card.
15 engagement1 sourceredditSun, Oct 11, 2026, 08:47 AM