flash-next
qwen3.8-flash-next-expert-traces
Qwen3.8-Flash-Next expert routing traces
Token-level routing traces of a deployed MoE model: for every token and every one of the
48 MoE layers, which experts the router chose, the top-32 router logits behind that choice,
and the exact hidden state the router read — plus, in v3, the state at many layers per token,
the post-final-norm state the LM head consumes, and the LM head's top-8 next-token candidates.
The corpus exists to answer one question: how well can the next tokens'… See the full description on the dataset page: https://huggingface.co/datasets/aswinkumar99/qwen3.8-flash-next-expert-traces.Qwen3.8-Flash-Next-GGUF-metricsqwen38-flash-next-int2-rtx5090-research
Qwen3.8-Flash-Next Mixed INT2 AutoRound on a Single RTX 5090
This benchmark and reproducibility artifact documents SGLang inference for the mixed-INT2 AutoRound Qwen3.8-Flash-Next checkpoint on one NVIDIA RTX 5090 Blackwell GPU. It covers a validated 256K / 262,144-token long context, MoE autotuning, tiered KV cache, CPU offload, and the device-local evidence showing why another Blackwell GPU's tuning configuration should not be copied blindly.
Headline inference… See the full description on the dataset page: https://huggingface.co/datasets/Jerrybro/qwen38-flash-next-int2-rtx5090-research.Qwen3.8-Flash-Next-Abliterated-Uncensored-GGUF-metrics
Qwen3.8-Flash-Next-Abliterated-Uncensored-GGUF: measurements
Everything behind the numbers on the
model card,
from one run on 2026-10-07/08: Qwen/Qwen3.8-Flash-Next@de4b8e4d, llama.cpp
980aef8c, 8x RTX PRO 6000 (sm_120), CUDA 13.
Path
What
kld/
The original BF16 model's logits over the held-out neutral and code sets (87 chunks at 4096 context), the reference for every KLD.
ablit/data/manifest.json
Prompt sources with revisions, the split sizes and the sha256 of… See the full description on the dataset page: https://huggingface.co/datasets/AtomicChat/Qwen3.8-Flash-Next-Abliterated-Uncensored-GGUF-metrics.gameworld-qwen38-flash-next-trajectory
GameWorld: Qwen3.8-Flash-Next Trajectories
This dataset is organized for inspecting trajectories.
It contains 1,020 completed trajectories from Qwen/Qwen3.8-Flash-Next
(both General and Computer-Use (CUA) interfaces) on the GameWorld 400-step benchmark: 170 tasks, and 3 runs per task.
Leaderboard
Each model was evaluated on 170 tasks with a 400-step limit and three runs per task (510 trajectories per interface).
SR is the success rate, and PG is mean task… See the full description on the dataset page: https://huggingface.co/datasets/yyyang/gameworld-qwen38-flash-next-trajectory.Qwen3.8-Flash-Next-expert-activation-map
Qwen3.8-Flash-Next Expert Activation Map
Per-(layer, expert) routing and output-importance statistics for
Qwen3.8-Flash-Next (Qwen4Exp architecture, 48 MoE layers x 512 routed
experts, top-k 10), measured on the unquantized bf16 checkpoint over a
2751-prompt, 28-domain calibration corpus.
The purpose is to answer, per layer, which experts carry the model's routed
output so that expert-level decisions (bf16 protection under quantization,
offload residency, pruning, warm-start… See the full description on the dataset page: https://huggingface.co/datasets/tcclaviger/Qwen3.8-Flash-Next-expert-activation-map.