Qwen3.8 Flash Next

Andrew Zhu's Medium piece dates the Qwen3.8-Flash-Next release to August 26, 2026, characterizing it as the third Qwen release in three weeks and describing the 'Next' suffix as Qwen's research-preview brand, the same role Qwen3-Next played ahead of Qwen3.5. The article emphasizes that only 6B parameters activate per t The piece quotes Qwen's launch blog explaining the early-release rationale — letting the community examine architectural changes before the full Qwen4 family is built on top of them — and reports that on several agentic-coding measurements the model is better than Claude Opus 4.6. The author notes a multi-gigabyte manu

Kaitchup's review characterizes Qwen3.8-Flash-Next as a multimodal Mixture-of-Experts model with a 125B-parameter main model, only 6B parameters activated per token, an additional 51B parameters dedicated to a new n-gram embedding system, and a 4B MTP module for speculative decoding. The article emphasizes that the 'Ne The piece notes that the model's Transformers configuration already identifies the architecture as qwen4_exp, and outlines planned coverage comparing Qwen3.8-Flash-Next against Qwen3.8 27B, explaining the new components such as Qwen Sparse Attention, Gated Residual, and the 51B-parameter n-gram embedding table, plus an

Atomic Chat's guide, published six days after the August 26, 2026 weights drop, lays out what developers actually need to run Qwen3.8 Flash Next locally. It documents the model's key specs: 125B total parameters with 6B active per token, 48 layers of 512 routed experts (10 active + 1 shared), a 262,144-token native con Beyond the spec sheet, the guide frames Qwen3.8 Flash Next as the first open model built on the Qwen4 architecture that can be run locally, and it provides concrete llama.cpp launch steps alongside the Atomic Chat workflow. It positions the model as a cost-efficient workhorse rather than a Max-class flagship, and it li

The MindStudio overview of Qwen3.8-Flash-Next characterizes it as an open-weight preview model from Alibaba's Qwen team, listed under a new qwen4_exp model type on Hugging Face, signaling it as an experimental step toward Qwen 4. It describes the architecture as roughly 125B total parameters with about 6B active per to MindStudio reports that the model combines gated-delta layers maintaining compressed running memory with occasional Qwen sparse attention layers scanning context in chunks, alongside gated residual pathways that preserve information deeper into the network. Benchmark coverage in the overview has Qwen3.8-Flash-Next beat

DataCamp reports that Qwen3.8-Flash-Next was released on August 26, 2026, the same day as GLM-5.3 Flash, as an open-weight 125B MoE model with 6B active parameters per token that previews the Qwen4 architecture. It is positioned as a cost-efficient workhorse below Max-class intelligence, with a native 262,144-token con The piece cites benchmark results including SWE-bench Pro at 62.5 versus Claude Opus 4.6 Max at 53.4, CoWorkBench at 73.9 versus 68.2, and JobBench at 55.7 versus 36.6, while trailing on Humanity's Last Exam (35.9 versus 40.0) with reported brittleness on long agent chains. Pricing is stated as $0.16 per million input

Local AI Zone's deep-dive describes Qwen3.8-Flash-Next as the second-generation 'Next' architecture preview from Alibaba's Qwen team, released August 26, 2026, positioned as the open-weight counterpart to the production API SKU qwen3.8-flash. The architecture is laid out as 48 layers arranged as 12 macro-blocks of 3×(G Headline self-reported benchmarks listed include DeepSWE 1.1 at 58.7, SWE-bench Pro at 62.5 (reportedly beating Claude Opus 4.6 Max at 53.4), CoWorkBench at 73.9, GPQA Diamond at 91.7, LiveCodeBench v6 at 91.9, and AndroidWorld at 84.5, with weaker results on NL2Repo-Bench (48.1 versus DeepSeek-V4-Flash at 54.2), HLE (

An August 26, 2026 NVIDIA NIM forum post describes Qwen3.8-Flash-Next 176B as a multimodal MoE with a 125B main model plus 51B N-gram embeddings and 6B active parameters per token, and a native 262,144-token context extensible to 1M with YaRN. Weights are offered on Hugging Face and ModelScope, with a managed API via QwenCloud. Performance figures from NVIDIA cite the FP8 variant running on GB300 NVL72 (72 Blackwell Ultra GPUs, 130 TB/s NVLink) at 16,000 tokens/sec per GPU and 200 tokens/sec per user. Day-0 inference stacks include SGLang, vLLM, TensorRT-LLM, NeMo AutoModel fine-tuning, NeMo RL recipes, and TokenSpeed; the model also runs locally on DGX Spark, DGX Station, and 4× RTX PRO 6000 Blackwell workstations.

AIreiter's guide frames Qwen3.8-Flash-Next as a multimodal Mixture-of-Experts model built on the architecture intended for the upcoming Qwen4 family, explicitly noting it is not Qwen4 itself but a technology demonstrator that lets the open-source inference ecosystem build kernels, quantizations, and serving support ahe The article distinguishes confirmed facts (open-weight release on ModelScope and Hugging Face, multimodal vision-plus-text, likely Apache 2.0 license, GDN hybrid plus Qwen Sparse Attention, FP8 variant Qwen/Qwen3.8-Flash-Next-FP8, API pricing/availability, Qwen4 architecture preview status) from community speculation (