Qwen3.8-Flash-Next
Alibaba·
Overview
Open-weight multimodal MoE that Qwen ships as an early preview of the architecture behind Qwen4 — the same role Qwen3-Next played for Qwen3.5. A 125B main model plus a separate 51B N-gram embedding table, with 6B parameters activated per token. Four changes carry the release: a GDN + QSA hybrid attention stack (Gated DeltaNet compresses the history, Qwen Sparse Attention picks important context at micro-block granularity), a Gated Residual stream widened to four dynamically gated branches, the N-gram embedding table itself (looked up from local context, offloadable to host memory with asynchronous prefetching), and a Muon optimizer with a scaling law refitted for the new architecture. Native context is 262,144 tokens. Qwen reports roughly 1/9 the training cost of Qwen3.7-Plus. Because only 6B parameters are active and the embedding table can sit in system RAM, Unsloth's smallest quant runs on a 78 GB RAM / unified-memory machine without a discrete GPU. A production Qwen3.8-Flash for the QwenCloud API was announced at $0.16 / $0.47 per MTok but is not live yet, so no cloud entry is tracked; the license file ships with the weights and is not stated in the GitHub release.
Capabilities and innovations
Benchmarks
| Benchmark | Score | Source |
|---|---|---|
| GPQA Diamond | 91.7% | vendor GPQA Diamond 91.7 from Qwen's launch table. The blog and the Hugging Face model card are both unreachable from this environment (egress-blocked); the value was taken from consistent secondary transcriptions of the table and not independently reproduced. Same caveat as the Qwen3.8-27B entry. |
| Humanity’s Last Exam (no tools) | 35.9% | vendor HLE 35.9 from the same launch table. Caveat beyond the transcription one: Qwen states this run is graded by GPT-4o rather than the benchmark's default grader, so it is not strictly comparable to the HLE values stored for other models. Same transcription caveat as GPQA. |
| SWE-bench Pro | 62.5% | vendor SWE-bench Pro 62.5 from the same launch table (Qwen3.8-27B: 61.7, Qwen3.7-Plus: 55.8). Same transcription caveat as GPQA. |
- Terminal-Bench 2.x: not reported by the vendor — Qwen's launch table leans on its own agentic suite (DeepSWE 1.1 58.7, SWE-bench Multilingual 81.0, LiveCodeBench v6 91.9, JobBench 55.7) and publishes no Terminal-Bench figure. Neither Artificial Analysis nor Vals AI lists the model as of 2026-08-26.
- MMLU-Pro: not reported by the vendor — MMLU-Pro is only mentioned for the base model, as one of the 8 of 14 benchmarks Qwen3.8-Flash-Next-Base is said to lead; no number was published for the instruct model.
- MMMU-Pro: independent evaluation pending — The model is natively multimodal, but no MMMU-Pro figure appears in the launch table and no independent run is available yet.
- SWE-bench Verified: not reported by the vendor — Era-1 benchmark; Qwen reports SWE-bench Pro and its own coding suite instead.
- Terminal-Bench 3.0: not reported by the vendor
Architecture and hardware
- Parameters
- 176B total, 6B active per token (MoE)
- Estimated VRAM at Q4
- ~102 GB, Multi-GPU Self-Host class
- Quantization formats
- GGUF, FP8, NVFP4, BF16
- Recommended runtime
- llama.cpp
- License
- Open (final license pending)
Links
More from Alibaba
Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.