Releases·Tool Release
Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s
A Hacker News post links to a GitHub project claiming Qwen 3.8 Flash Next (125B) can run on a consumer RTX 4090 at 100 tokens/s. The post carries no substantive excerpt, so specifics about the method cannot be confirmed from the available material.
1214 engagement1 sourceSun, Oct 4, 2026, 12:51 PM
AI Overview
- Qwen 3.8 Flash Next, a 125B-parameter model, runs on an RTX 4090 at a reported 100 tokens/s, per the linked GitHub project
Entities
Related
Social·Community
Developer runs Qwen3.8-27B at 256K context on 24GB RTX PRO 4000 SFF, 50 tok/s with MTP
A developer demonstrates running Qwen3.8-27B with full 256K context on a 24GB RTX PRO 4000 SFF GPU, achieving 50 tokens per second using multi-token prediction. The setup uses 97.9% of GPU memory and leverages the card's 432 GB/s bandwidth, showing large-context inference is feasible on consumer hardware.
A developer demonstrates running Qwen3.8-27B with full 256K context on a 24GB RTX PRO 4000 SFF GPU, achieving 50 tokens per second using multi-token prediction. The setup uses 97.9% of GPU memory and leverages the card's 432 GB/s bandwidth, showing large-context inference is feasible on consumer hardware.
103 engagement2 sourceshackernews, redditMon, Aug 17, 2026, 02:29 PM
Developments·Benchmark
Qwen3.8-Flash-Next-oQ4e-mtp: 45 tok/s on M4 Max, 25 tok/s on M2 Ultra for local inference — llm-bench.io
llm-bench.io published local inference benchmarks for a quantized Qwen3.8-Flash-Next build on Apple Silicon, reporting 45 tok/s on an M4 Max and 25 tok/s on an M2 Ultra. The build is quantized to oQ4e with multi-token prediction and delivers speeds similar to Qwen3.8 27B on Apple Silicon.
llm-bench.io published local inference benchmarks for a quantized Qwen3.8-Flash-Next build on Apple Silicon, reporting 45 tok/s on an M4 Max and 25 tok/s on an M2 Ultra. The build is quantized to oQ4e with multi-token prediction and delivers speeds similar to Qwen3.8 27B on Apple Silicon.
44 engagement1 sourceredditSun, Sep 6, 2026, 08:42 AM
Social·Community
Kimi K3 (2.8T) at 1 token/s on a MacBook Pro, streamed from four SSDs
Kimi K3 (2.8T) runs at 1 token/s on a MacBook Pro with weights streamed from four SSDs
Kimi K3 (2.8T) runs at 1 token/s on a MacBook Pro with weights streamed from four SSDs
426 engagement1 sourcehackernewsTue, Sep 8, 2026, 08:07 PM
Social·Community
ROCm vs Vulkan on R9700+Strix Halo: DSv4 Flash 0731 and Qwen 3.8 Flash Next benchmarked in llama.cpp
A user ran llama.cpp inference benchmarks comparing ROCm 10.0 and Vulkan (RADV) on an R9700 + Strix Halo rig, testing DeepSeek v4 Flash 0731 and Qwen 3.8 Flash Next with updated kernel, firmware, Mesa 26.2.3, and recent llama.cpp PRs. ROCm remains faster for DeepSeek, but Vulkan has narrowed the gap; for Qwen, ROCm leads prompt processing while Vulkan leads token generation. The results matter to local-inference practitioners choosing a backend on AMD iGPU-class hardware.
A user ran llama.cpp inference benchmarks comparing ROCm 10.0 and Vulkan (RADV) on an R9700 + Strix Halo rig, testing DeepSeek v4 Flash 0731 and Qwen 3.8 Flash Next with updated kernel, firmware, Mesa 26.2.3, and recent llama.cpp PRs. ROCm remains faster for DeepSeek, but Vulkan has narrowed the gap; for Qwen, ROCm leads prompt processing while Vulkan leads token generation. The results matter to local-inference practitioners choosing a backend on AMD iGPU-class hardware.
19 engagement1 sourceredditTue, Sep 22, 2026, 05:15 AM
Social·Community
What model sits between Qwen 3.8 27b and Flash next for coding?
A Reddit user testing Qwen 3.8 27B and Flash next locally on an RTX 5090 with 96GB DDR5 asks the community for a model balancing coding capability and decode speed, targeting 75-100 TPS. Qwen 3.8 27B runs at 200+ TPS but lacks the coding strength of Flash next, which manages only ~50 TPS. If no middle option exists, the user plans to use Flash next for planning and the 27B for implementation.
A Reddit user testing Qwen 3.8 27B and Flash next locally on an RTX 5090 with 96GB DDR5 asks the community for a model balancing coding capability and decode speed, targeting 75-100 TPS. Qwen 3.8 27B runs at 200+ TPS but lacks the coding strength of Flash next, which manages only ~50 TPS. If no middle option exists, the user plans to use Flash next for planning and the 27B for implementation.
316 engagement2 sourcesredditMon, Sep 28, 2026, 01:25 PM