Locally Explains

7 Qwen 3.8 models, and the biggest is not the best

Sources checked on 2 October 2026. Every benchmark figure below is a score Qwen or Alibaba published about its own models, in its own tables, with its own harnesses. None of them was measured independently.

Qwen3.8 27B

• Dense causal language model with a vision encoder; 27B parameters; image and video input; native context 262,144 tokens, “extensible up to 1,000,000 tokens”. huggingface.co/Qwen/Qwen3.8-27B
• Text performance table, Coding rows (Qwen3.8-27B · Qwen3.6-27B · Qwen3.7-Plus · Muse Glimmer-30B · Opus4.6 Max):
  • Terminal Bench 2.1 (Terminus): 73.0 · 63.4 · 64.0 · 51.7 · 78.2
  • SWE-bench Pro: 61.7 · 53.5 · 57.6 · 51.2 · 53.4
  • DeepSWE 1.1: 42.2 · 13.3 · 14.2 · – · –
• Footnote: “Except for Opus4.6 Max, which uses the officially reported score, all models are evaluated with the Claude Code harness at temp=1.0, top_p=0.95, and a 256K context window. Problematic tasks were corrected, and all baseline models were re-evaluated on the refined benchmark.”
• Quantized downloads: Unsloth’s GGUF repository lists UD-Q4_K_M at 16.5 GB, Q4_1 at 17.5 GB and UD-Q4_K_XL at 17.6 GB. Ollama’s qwen3.8:27b package is listed at 18 GB. huggingface.co/unsloth/Qwen3.8-27B-GGUF · ollama.com/library/qwen3.8/tags
• The 2.4T flagship against 27B: 2,400 / 27 = 88.9, so “almost ninety times” the parameters.

Qwen3.8 Flash Next

• “125B with 6B activated, plus 51B n-gram embedding and 4B MTP”; Gated DeltaNet and Qwen Sparse Attention; native context 262,144, extensible to 1,000,000; text, image and video input. Hugging Face lists the model size as 180B params. huggingface.co/Qwen/Qwen3.8-Flash-Next
• “a multimodal MoE model that also serves as an early preview of the architecture used in Qwen4”, carrying forward “the hybrid Gated DeltaNet + Gated Attention design”; “three out of every four layers use Gated DeltaNet”; N-gram Embedding “performs lookups using the local context formed by the current token and several preceding tokens”, “effectively adding a large-scale local-pattern memory”; the MTP module is trained for multi step speculative decoding. Published 26 August 2026. qwen.ai/blog?id=qwen3.8-flash-next
• “Qwen3.8-Flash is the official version based on Qwen3.8-Flash-Next with more production features, e.g., 1M context length by default, official built-in tools.” huggingface.co/Qwen/Qwen3.8-Flash-Next
• Coding rows (Flash Next · 27B · Qwen3.7-Plus · DeepSeek-V4-Flash-0731 · Claude Opus 4.6 Max): DeepSWE 1.1 58.7 · 42.2 · 16.5 · 54.4 · –; SWE-bench Pro 62.5 · 61.7 · 55.8 · 56.0 · 53.4. Also CoWorkBench 73.9 and NL2Repo-Bench 48.1 for Flash Next.

Qwen3.8 Max

• Open checkpoint Qwen3.8-2.4T-A95B: “2.4T in total and 95B activated”, 512 experts with 10 routed and 1 shared active; native context 262,144, extensible to 1,010,000; “a text-only model that requires thinking mode for all interactions”. huggingface.co/Qwen/Qwen3.8-2.4T-A95B
• “Qwen3.8-Max is the official version based on Qwen3.8-2.4T-A95B with more features, such as vision input & non-thinking support, 1M context length by default, official built-in tools.”
• Full benchmark table (Opus4.8 · Fable5 · GPT5.6 Sol (max) · Qwen3.7-Max · Qwen3.8-Max): Terminal Bench 2.1 84.6 · 84.6 · 88.8 · 74.5 · 86.6; SWE-bench Pro 69.2 · 80.0 · 64.6 · 60.6 · 67.7; PaperBench 80.3 · 88.8 · 90.5 · 64.8 · 93.0. qwen.ai/blog?id=qwen3.8 (3 August 2026)
• “As of July 30, 2026, after approximately 16 days of fully autonomous AI operation, the repository had accumulated 265 commits, 127 PRs, and 151 issues.” The project is the oh-my-cli command line harness, run by Qwen3.8-Max. Same page.
• Serving recipe: the NVFP4 W4A4 build is 1.32 TiB of weights and fits on 8 B300 GPUs. recipes.vllm.ai/Qwen/Qwen3.8-2.4T-A95B
• NVIDIA DGX B300: eight B300 GPUs in one system. nvidia.com/en-us/data-center/dgx-b300

Prices

• Alibaba Cloud Model Studio, per million tokens, 2 October 2026: qwen3.8-max, International deployment, $2 input and $6 output; qwen3.8-flash, Global deployment, $0.113 input and $0.382 output (International $0.15 and $0.47). Prices differ by region, deployment scope, batch and caching discounts. alibabacloud.com/help/en/model-studio/model-pricing

Qwen3.8 Omni Flash and Omni Flash Realtime

• “Text, image, audio, and video inputs with a 1M-token context window”; workflows such as video editing, translation and building reports from meetings; across 29 evaluations its average score improves by more than 25% over Qwen3.5-Omni-Plus. qwen.ai/blog?id=qwen3.8-omni-flash (18 September 2026)
• UniClawBench 69.6 (Qwen3.5-Omni-Plus 67.1). LVBench 76.9. AndroidWorld: Omni Flash 87.1, Flash 84.5, 27B 81.9, Qwen3.7-Plus 81.0, Claude Opus 4.6 (Max) 62.0; 87.1 minus 62.0 is 25.1 points. ClawEval-MM: pass@3 60.4, average 61.9.
• Text table: SWE-bench Pro Omni Flash 63.3, Flash 62.5, 27B 61.7, Qwen3.7-Plus 55.8, DeepSeek-V4-Flash-0731 56.0, Opus 4.6 Max 53.4; DeepSWE 1.1 Omni Flash 57.8, Flash 58.7.
• Realtime: “It perceives and responds while receiving live audio-visual streams”; “supports connections over WebSocket and WebRTC”. Same page.

Qwen3.8 LiveTranslate

• “Average lagging (LAAL) drops from 2.8 seconds to 2.3 seconds”; an Interleave architecture “improving faithfulness, fluency, and conciseness”; real time speaker separation; synchronized source and translation output; 60 languages. On FLEURS across 70 language directions it “leads both the previous generation and current mainstream real-time interpretation systems”. qwen.ai/blog?id=qwen3.8-livetranslate (18 September 2026)

A discrepancy

• The narration says Qwen3.7 Plus scores 78.2 on Terminal Bench 2.1, ahead of 27B’s 73.0. In Qwen’s 27B model card, 78.2 is Claude Opus 4.6 Max’s score; Qwen3.7-Plus is listed at 64.0, below the 27B. The picture shows the table as published.