Alibaba’s Qwen workforce has launched Qwen3.8-Flash-Subsequent, an open-weight multimodal Combination-of-Specialists mannequin constructed for price per token. The checkpoint pairs a 125B spine with a 51B N-gram embedding desk and a 4B multi-token prediction module. Solely 6B parameters activate per token. The workforce positions it as an early preview of the structure that may underpin Qwen4, the identical position Qwen3-Subsequent performed for Qwen3.5. 4 adjustments carry the discharge: a Gated DeltaNet and Qwen Sparse Consideration hybrid, Gated Residual, N-gram Embedding, and the Muon optimizer. Qwen workforce stories coaching price at roughly one-ninth that of Qwen3.7-Plus.
Is it deployable?
Sure however not on a workstation. The FP8 checkpoint is 172.78 GiB and the BF16 checkpoint is 335.28 GiB. Per vLLM recipes, TP2 is the minimal validated FP8 configuration on GB300 and TP4 is really helpful. On an 8×H200 node, use TEP8; plain TP8 is incompatible with the checkpoint’s 128-wide quantization blocks. Sparse activation cuts compute, not storage.
What is definitely new
Qwen3.8-Flash-Subsequent pairs a 125B predominant mannequin with 51B N-gram embedding parameters and a 4B multi-token prediction module, totaling 180B on disk. Solely 6B parameters activate per token. 4 adjustments drive this:
- Hybrid consideration (GDN + QSA): Three of each 4 layers use Gated DeltaNet, a linear-attention layer that compresses historical past right into a fixed-size recurrent state. The fourth layer runs Qwen Sparse Consideration (QSA), which makes use of a light-weight indexer to pick context at micro-block granularity reasonably than per token. The layer format is 12 × (3 × GDN → 1 × QSA) throughout 48 layers, with a QSA finances of 512 blocks or 2048 tokens.
- Gated Residual: The residual stream widens into 4 parallel branches, with an element-wise learn gate and a per-branch scalar write gate, at bottleneck rank 320.
- N-gram Embedding: A 20,000,000-entry bigram/trigram desk at layer 2 provides capability by means of deterministic lookups. It may be offloaded to host reminiscence with asynchronous prefetch — although offload at present runs solely on NVIDIA gadgets.
- Coaching recipe.:The Muon optimizer is utilized alongside AdamW to particular weight classes, with batch-size warmup eradicated and scaling legal guidelines refitted.
The MoE layer carries 512 specialists, activating 10 routed plus 1 shared, at knowledgeable intermediate dimension 640.
Benchmarks
Qwen stories 58.7 on DeepSWE 1.1, 62.5 on SWE-bench Professional, 81.0 on SWE-bench Multilingual, and 91.9 on LiveCodeBench v6. On agentic duties it posts 73.9 on CoWorkBench, 55.7 on JobBench, and 73.5 on Toolathlon Verified. Multimodal outcomes embody 84.5 on AndroidWorld, 76.6 on LVBench, 88.5 on RealWorldQA, and 95.7 on MathVision with code interpreter.
The mannequin doesn’t lead all over the place. Claude Opus 4.6 (Max) takes HLE at 40.0 towards Qwen’s 35.9, and DeepSeek-V4-Flash-0731 leads NL2Repo-Bench at 54.2 versus 48.1. Frontier reasoning stays the hole.
Effectivity
Qwen states coaching price roughly 1/9 that of Qwen3.7-Plus. On serving, the announcement cites QSA kernel speedups of as much as 7.6× prefill and 4.9× decode at 1M tokens, whereas the SGLang cookbook and vLLM recipes cite 10.2× and 6.6×. Deal with the vary as vendor-reported till independently measured. Qwen additionally stories 8.6× the prefill throughput of Qwen3.7-Plus at a 90% prefix-cache hit price.
Context is 262,144 tokens natively, extensible to 1,000,000 with YaRN.
Working it
The mannequin serves by means of vLLM, SGLang, TokenSpeed, transformers serve, and llama.cpp for GGUF quants. Fantastic-tuning is supported through Unsloth, Swift, and LLaMA-Manufacturing facility. It already powers the “Commonplace” mode on QwenWork and works with Qwen Code.
Considering mode is on by default, with reasoning_effort at xhigh, medium, or low. Qwen recommends temperature 1.0 and top_p 0.95 for pondering mode, and temperature 0.7 with top_p 0.80 for instruct mode.
Key Takeaways
- 125B spine + 51B N-gram embeddings + 4B MTP, with solely 6B parameters energetic per token.
- Three of 4 layers use Gated DeltaNet; the fourth runs Qwen Sparse Consideration at micro-block granularity.
- Skilled at roughly 1/9 the price of Qwen3.7-Plus, with 262K native context extensible to 1M through YaRN.
- FP8 weights are 172.78 GiB, so self-hosting wants a multi-GPU node, not a workstation.
- Licensed underneath qwen-community-1.0, not Apache-2.0 — confirm phrases earlier than industrial use.
Take a look at the GitHub Web page, HF Mannequin Card and Technical Particulars. Additionally, be happy to comply with us on Twitter and don’t overlook to affix our 150k+ML SubReddit and Subscribe to our E-newsletter. Wait! are you on telegram? now you’ll be able to be a part of us on telegram as nicely.
Asif Razzaq is the CEO of Marktechpost Media Inc.. As a visionary entrepreneur and engineer, Asif is dedicated to harnessing the potential of Synthetic Intelligence for social good. His most up-to-date endeavor is the launch of an Synthetic Intelligence Media Platform, Marktechpost, which stands out for its in-depth protection of machine studying and deep studying information that’s each technically sound and simply comprehensible by a large viewers. The platform boasts of over 2 million month-to-month views, illustrating its recognition amongst audiences.
