{"id":13138,"date":"2026-11-17T11:02:06","date_gmt":"2026-11-17T02:02:06","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/?p=13138"},"modified":"2026-11-17T13:11:47","modified_gmt":"2026-11-17T04:11:47","slug":"qwen38-flash-next-e384-fp4-local","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/qwen38-flash-next-e384-fp4-local\/","title":{"rendered":"Qwen3.8-Flash-Next-E384-FP4-Local Vision-Language Model: ~96GB Memory"},"content":{"rendered":"<p><!-- lmw:facts --><\/p>\n<h2>At a Glance<\/h2>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Item<\/th>\n<th>Value<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Repository<\/td>\n<td><a href=\"https:\/\/huggingface.co\/autotrust\/Qwen3.8-Flash-Next-E384-FP4-Local\">autotrust\/Qwen3.8-Flash-Next-E384-FP4-Local<\/a><\/td>\n<\/tr>\n<tr>\n<td>Publisher guide<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/publisher-alibaba-en\/\">Alibaba (Qwen): models and licenses<\/a><\/td>\n<\/tr>\n<tr>\n<td>Published<\/td>\n<td>2026-10-22<\/td>\n<\/tr>\n<tr>\n<td>License<\/td>\n<td>nvidia-open-model-agreement<\/td>\n<\/tr>\n<tr>\n<td>Formats<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/format-safetensors-en\/\">safetensors<\/a><\/td>\n<\/tr>\n<tr>\n<td>Source type<\/td>\n<td>Unverified (not confirmed by a primary source)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><em>Values determined by this site&#8217;s code when the information was collected. Dates are JST.<\/em><\/p>\n<p><!-- \/lmw:facts --><\/p>\n<h2>Overview<\/h2>\n<p>It is reported by autotrust that &#8220;Qwen3.8-Flash-Next-E384-FP4-Local&#8221;, an unofficial derivative of NVIDIA&#8217;s published &#8220;Qwen3.8-Flash-Next-FP4-Local&#8221;, has been released. Please note that this information has not been officially confirmed at this time and is based on unverified community reports. The model is said to use Neural Architecture Search (NAS) to reduce the number of experts in the Mixture of Experts (MoE) structure, aiming for lightweight inference.<\/p>\n<p>According to the model card, it is configured to retain 384 out of the original 512 experts in each decoder MoE layer and MTP (Multi-Token Prediction) layer. It is reported to be built with the aim of maintaining performance on specific tasks such as coding and tool use while reducing the model size by approximately 16%.<\/p>\n<h2>Specifications<\/h2>\n<p>Specifications listed in the documentation are as follows. This model is an unofficial derivative with some modified configurations.<\/p>\n<ul>\n<li>Parameter count: Total parameters 125B, activated parameters 6B (both figures from the base model)<\/li>\n<li>Architecture: Mixture of Experts (MoE) \/ Qwen4ExpForConditionalGeneration<\/li>\n<li>Number of experts: 384 routed experts per layer (reduced from the original 512)<\/li>\n<li>Activated experts: 10 routed experts + shared expert<\/li>\n<li>Context length: Up to 262,144 tokens<\/li>\n<li>Quantization format: NVFP4<\/li>\n<\/ul>\n<h2>Performance<\/h2>\n<p>Measurement results reported by the publisher autotrust are available. The comparison target is the pre-reduction original model (512-expert version). Scores are reported to have been measured using a B200 GPU and vLLM 0.31.0.<\/p>\n<h3>Major Benchmark Comparison<\/h3>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<p class=\"lmw-table-hint\" style=\"margin:0 0 4px;font-size:0.85em;opacity:0.7;\">\u2192 Scroll horizontally to see all columns<\/p>\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>capability<\/th>\n<th>benchmark<\/th>\n<th>setting<\/th>\n<th>original<\/th>\n<th><strong>E384<\/strong><\/th>\n<th>\u0394<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>reasoning (science)<\/td>\n<td>GPQA-Diamond (198)<\/td>\n<td>effort=low, 32k<\/td>\n<td>85.9<\/td>\n<td>77.3<\/td>\n<td>\u22128.6<\/td>\n<\/tr>\n<tr>\n<td>English knowledge<\/td>\n<td>MMLU test (14,042)<\/td>\n<td>effort=low, 16k<\/td>\n<td>92.0<\/td>\n<td>86.9<\/td>\n<td>\u22125.1<\/td>\n<\/tr>\n<tr>\n<td>Chinese knowledge<\/td>\n<td>C-Eval val (1,346)<\/td>\n<td>effort=low, 16k<\/td>\n<td>93.5<\/td>\n<td>89.0<\/td>\n<td>\u22124.5<\/td>\n<\/tr>\n<tr>\n<td>math<\/td>\n<td>AIME 2026 (30 \u00d7 4, pass@1)<\/td>\n<td>effort=xhigh, 64k<\/td>\n<td>95.0<\/td>\n<td>93.3<\/td>\n<td>\u22121.7<\/td>\n<\/tr>\n<tr>\n<td>vision<\/td>\n<td>MMMU val (900)<\/td>\n<td>effort=low, 16k<\/td>\n<td>81.1<\/td>\n<td>79.4<\/td>\n<td>\u22121.7<\/td>\n<\/tr>\n<tr>\n<td>code<\/td>\n<td>HumanEval (164)<\/td>\n<td>greedy, effort=low<\/td>\n<td>97.0<\/td>\n<td>97.6<\/td>\n<td>+0.6<\/td>\n<\/tr>\n<tr>\n<td>tool use<\/td>\n<td>BFCL v4 Non-Live AST<\/td>\n<td>effort=low<\/td>\n<td>88.2<\/td>\n<td>88.6<\/td>\n<td>+0.4<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>BFCL v4 Live AST<\/td>\n<td><\/td>\n<td>79.3<\/td>\n<td>78.9<\/td>\n<td>\u22120.4<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>BFCL v4 Multi-Turn Base (200)<\/td>\n<td><\/td>\n<td>62.5<\/td>\n<td>71.5<\/td>\n<td>+9.0<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>BFCL v4 relevance \/ irrelevance<\/td>\n<td><\/td>\n<td>81.3 \/ 80.1<\/td>\n<td>81.3 \/ 78.6<\/td>\n<td>0 \/ \u22121.5<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>From this table, it can be said that the impact of reducing the number of experts by 25% varies greatly by field. Knowledge-intensive multiple-choice tasks such as MMLU, C-Eval, and GPQA-Diamond show a notable drop of 4.5 to 8.6 points. On the other hand, items such as HumanEval (coding) and parts of BFCL (tool use) record scores equivalent to or exceeding the original. A significant improvement of 9.0 points is reported particularly in the multi-turn item of BFCL, which is speculated to be the result of optimization by NAS.<\/p>\n<h3>Hard Science Scores Under High-Load Inference<\/h3>\n<p>A comparison when increasing the thinking budget (token count) in GPQA-Diamond is also reported.<\/p>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>setting<\/th>\n<th>original (512)<\/th>\n<th><strong>E384<\/strong><\/th>\n<th>\u0394<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>effort=low, 32k budget<\/td>\n<td>85.9<\/td>\n<td>77.3<\/td>\n<td>\u22128.6<\/td>\n<\/tr>\n<tr>\n<td><strong>effort=xhigh, 259,000-token budget<\/strong><\/td>\n<td><strong>90.9 (180\/198)<\/strong><\/td>\n<td><strong>85.9 (170\/198)<\/strong><\/td>\n<td><strong>\u22125.1<\/strong><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>These results suggest that providing sufficient thinking time can compensate to some extent for the loss of knowledge due to expert reduction. The difference of 8.6 points during low load shrinks to 5.1 points during high load. However, this model tends to think longer than the original, and is reported to have required a median of 16,817 tokens for inference.<\/p>\n<h3>MTP Speculative Decoding Acceleration<\/h3>\n<p>This model includes an MTP layer to which a unique acceleration recipe has been applied. Actual measured values on a B200 are as follows.<\/p>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<p class=\"lmw-table-hint\" style=\"margin:0 0 4px;font-size:0.85em;opacity:0.7;\">\u2192 Scroll horizontally to see all columns<\/p>\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>config<\/th>\n<th>k<\/th>\n<th>tok\/s<\/th>\n<th>\u03c4<\/th>\n<th>vs no MTP<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>no MTP<\/td>\n<td>0<\/td>\n<td>92.6<\/td>\n<td>\u2014<\/td>\n<td>1.00\u00d7<\/td>\n<\/tr>\n<tr>\n<td>native MTP<\/td>\n<td>7<\/td>\n<td>300.1<\/td>\n<td>5.14<\/td>\n<td>3.24\u00d7<\/td>\n<\/tr>\n<tr>\n<td><strong>this recipe<\/strong><\/td>\n<td><strong>7<\/strong><\/td>\n<td><strong>308.9<\/strong><\/td>\n<td>4.93<\/td>\n<td><strong>3.34\u00d7<\/strong><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>Applying this recipe is said to achieve up to 3.34x inference speed improvement compared to when MTP is not used. This is explained as the result of significantly reducing the number of read bytes per step through draft-time vocabulary restriction and quantization optimization.<\/p>\n<h2>Strengths and Use Cases<\/h2>\n<p>This model is optimized with the goal of exhibiting high efficiency in specific tasks. Based on the description in the model card and <code>capability_tags<\/code>, utilization is expected particularly in the following fields.<\/p>\n<p>First, a major feature is that it maintains performance equivalent to or higher than the base model in coding and agent (tool use) tasks. The HumanEval benchmark reports scores slightly surpassing the base model, making it suitable for programming assistance such as implementing Python functions based on specifications. In addition, from the BFCL benchmark results, it has high accuracy in tool use like function selection and argument assembly, making it suitable for use as an AI agent executing complex instructions. Significant improvements exceeding the base model have been reported particularly in tool use involving multi-turn (multiple dialogues).<\/p>\n<p>Also, as indicated by the <code>image-text-to-text<\/code> tag, this model is a multimodal model capable of accepting both images and texts as inputs. It can understand the contents of charts and photos and perform reasoning based on them. Because it supports a very long context length natively expandable up to 262,144 tokens and up to 1,000,000 tokens maximum, it is said to handle large document analysis and processing contexts containing massive numbers of images.<\/p>\n<p>On the other hand, since performance degradation is recognized in knowledge-intensive tasks (such as MMLU and C-Eval), utilization as an &#8220;inference engine&#8221; that solves problems by leveraging provided context and tools is recommended over use cases asking broad general knowledge or directly outputting specific expert knowledge. The expert reduction by NAS can be interpreted as choosing a tradeoff that enhances the efficiency of specific logical tasks in exchange for retained knowledge volume.<\/p>\n<h2>How It Differs from Similar Models<\/h2>\n<p>On our site, the original model of the architecture underlying this model has been introduced as <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/2026\/10\/05\/qwen-38-flash-next\/\">&#8220;Qwen3.8-Flash-Next&#8221; Multimodal MoE Model: Approx. 402GB Required Memory<\/a>.<\/p>\n<p>The biggest difference between this &#8220;Qwen3.8-Flash-Next-E384-FP4-Local&#8221; and the original model lies in resource efficiency during inference and execution environment optimization. The original model was a massive configuration with 125B total parameters in BF16 precision, requiring immense memory to run. In contrast, this model is based on the NVFP4 quantized version by NVIDIA, and further reduces MoE experts from 512 to 384 by 25% via NAS (Neural Architecture Search). It is reported that this reduces the model weight file size by about 16%, enabling operation on more limited hardware resources.<\/p>\n<p>In addition, unique optimizations are also applied to the MTP (Multi-Token Prediction) layer, which is a speculative decoding technology. Restricting the draft vocabulary and quantizing the MTP layer itself drastically cuts data read amounts per step, aiming to improve inference speed, which is said to be a major difference from the standard Qwen3.8-Flash-Next series.<\/p>\n<p><!-- lmw:hardware --><\/p>\n<h2>Hardware Requirements<\/h2>\n<p><strong>Estimated requirements (calculated by Local Model Watch)<\/strong><\/p>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Your VRAM<\/th>\n<th>Quantization<\/th>\n<th>File size<\/th>\n<th>Est. memory needed<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>More than 96GB of VRAM (multi-GPU or CPU offload required)<\/td>\n<td>NVFP4(tensors)<\/td>\n<td>79.8GB<\/td>\n<td>95.8GB<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><strong>Inference engine support<\/strong> (architecture name matched against each project&#8217;s own model registry in its source code, checked 2026-11-17): llama.cpp: registered, vLLM: registered, MLX (mlx-lm): not registered. &#8220;Not registered&#8221; means the name is absent from that registry today, not that the model cannot run.<\/p>\n<p><em>Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model&#8217;s authors. Compare with other models in our <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/vram-guide-en\/\">VRAM quick reference<\/a>. What the quantization names mean: <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/glossary-quantization-en\/\">glossary<\/a>.<\/em><\/p>\n<p><!-- \/lmw:hardware --><\/p>\n<p><!-- lmw:runnability --><\/p>\n<h2>Can You Run It Locally?<\/h2>\n<p><strong>Not usable in Ollama, LM Studio and llama.cpp yet \u2014 we have found no GGUF build.<\/strong><\/p>\n<p>The publisher ships safetensors only. However, llama.cpp&#8217;s registry does list this architecture, so <strong>conversion to GGUF is possible<\/strong> and the model will run once someone publishes a converted build. Today it can be run with transformers or vLLM, using the memory figures in the table above.<\/p>\n<p><strong>License \u2014 <code>nvidia-open-model-agreement<\/code>:<\/strong> A custom license from the publisher. Check the original terms directly, including whether commercial use is permitted.<\/p>\n<p><em>Compiled by this site&#8217;s code from the published formats, converted builds we have found, and each engine&#8217;s own model registry. &#8220;Not found&#8221; means we have not seen such a build, not that none exists. License summaries are not legal advice \u2014 check the publisher&#8217;s original terms before relying on them.<\/em><\/p>\n<p><!-- \/lmw:runnability --><\/p>\n<p><!-- lmw:lab --><\/p>\n<h2>Our Own Measurements<\/h2>\n<h3>Measurements We Did Not Take<\/h3>\n<p>We have not confirmed that this site meets the commercial-use terms of this model&#8217;s license (<code>nvidia-open-model-agreement<\/code>). Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).<\/p>\n<p><!-- \/lmw:lab --><\/p>\n<h2>How to Get It<\/h2>\n<p>This model is distributed in <code>safetensors<\/code> format from the Hugging Face repository. vLLM (verified version 0.31.0) is recommended for execution, and it is reported that utilizing acceleration by MTP particularly requires unique plugins and environment variable settings included in the repository. Please note that this information has not been officially confirmed and is based on unverified community reports.<\/p>\n<p>As a specific startup procedure, it takes the form of first installing <code>mtp_draft\/plugin<\/code>, setting necessary environment variables, and then executing the <code>vllm serve<\/code> command.<\/p>\n<pre><code class=\"language-bash\"># \u30d7\u30e9\u30b0\u30a4\u30f3\u306e\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb\npip install -e mtp_draft\/plugin\n\n# \u74b0\u5883\u5909\u6570\u306e\u8a2d\u5b9a\nexport Q38_DRAFT_HOME=$PWD\/mtp_draft\nexport Q38_DRAFT_VOCAB=$PWD\/mtp_draft\/artifacts\/draft_vocab_64k.pt\nexport Q38_DRAFT_HEAD_Q=marlin\nexport Q38_DRAFT_MTP_Q=marlin\n\n# vLLM\u30b5\u30fc\u30d0\u30fc\u306e\u8d77\u52d5\u4f8b\nvllm serve autotrust\/Qwen3.8-Flash-Next-E384-FP4-Local \\\n  --speculative-config '{&quot;method&quot;:&quot;mtp&quot;,&quot;num_speculative_tokens&quot;:7,&quot;use_local_argmax_reduction&quot;:true}' \\\n  --block-size 48\n<\/code><\/pre>\n<p>Regarding the license, the NVIDIA Open Model Agreement applies, inheriting the Qwen Community License of the base model. Please note that commercial use and distribution require compliance with these license terms. Note that this model is not a gated model, so downloading is said to be possible without individual application or consent procedures on Hugging Face.<\/p>\n<p><!-- lmw:variants --><\/p>\n<h2>Quantized and Converted Variants<\/h2>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<p class=\"lmw-table-hint\" style=\"margin:0 0 4px;font-size:0.85em;opacity:0.7;\">\u2192 Scroll horizontally to see all columns<\/p>\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Added<\/th>\n<th>Publisher<\/th>\n<th>Format<\/th>\n<th>Repository<\/th>\n<th>Smallest VRAM tier (build, est. memory)<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>2026-10-27<\/td>\n<td>autotrust<\/td>\n<td>NVFP4<\/td>\n<td><a href=\"https:\/\/huggingface.co\/autotrust\/Qwen3.8-Flash-Next-E384-FP4-V2\">autotrust\/Qwen3.8-Flash-Next-E384-FP4-V2<\/a><\/td>\n<td>NVFP4 98.0GB (does not fit a single consumer GPU)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>File sizes of each build:<\/p>\n<ul>\n<li>Available builds in autotrust\/Qwen3.8-Flash-Next-E384-FP4-V2: NVFP4 81.7GB<\/li>\n<\/ul>\n<p><em>This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/glossary-quantization-en\/\">glossary<\/a>.<\/em><\/p>\n<p><!-- \/lmw:variants --><\/p>\n<p><!-- lmw:related --><\/p>\n<h2>Related Articles<\/h2>\n<ul>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/10\/05\/qwen-3-8-flash-next-released\/\">Qwen3.8-Flash-Next Multimodal MoE Model: ~402GB Memory<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/qwen3-8-27b-nvfp4-64k-mtp\/\">Qwen3.8-27B-NVFP4-DGX-SPARK-V2 NVFP4 Quantized 27B Model: 24GB+ VRAM<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/autotrust-tts-v5-overview\/\">TTS-V5 Chinese TTS Model: 4GB+ VRAM, File List<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/autotrust-tts-v4-chinese-call-center-tts\/\">TTS-V4 Chinese Text-to-Speech Model: 8GB+ VRAM, File List<\/a><\/li>\n<\/ul>\n<p><!-- \/lmw:related --><\/p>\n<p><!-- lmw:next-steps --><\/p>\n<h2>What to Read Next<\/h2>\n<ul>\n<li><strong>Find models by VRAM<\/strong> (This model does not fit even in 80GB; its smallest build needs about 96GB) \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/vram-guide-en\/\">VRAM quick reference, including models over 80GB<\/a><\/li>\n<li><strong>Engines that run this model<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/engine-vllm-en\/\">vLLM<\/a><\/li>\n<li><strong>What NVFP4 mean and where to get this model<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/format-fp4-en\/\">NVFP4 \/ MXFP4 format guide and models<\/a> \/ <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/glossary-quantization-en\/\">Quantization and model-format glossary<\/a><\/li>\n<li><strong>Learn about the publisher<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/publisher-alibaba-en\/\">Alibaba (Qwen): models, licenses and articles<\/a><\/li>\n<li><strong>Other models for the same task<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/models-by-task-en\/#task-vision\">Other vision-language models<\/a><\/li>\n<\/ul>\n<p><!-- \/lmw:next-steps --><\/p>\n<h2>Sources<\/h2>\n<ul>\n<li><a href=\"https:\/\/huggingface.co\/autotrust\/Qwen3.8-Flash-Next-E384-FP4-Local\">https:\/\/huggingface.co\/autotrust\/Qwen3.8-Flash-Next-E384-FP4-Local<\/a><\/li>\n<\/ul>\n<blockquote>\n<p><strong>This article contains unverified information.<\/strong> We will append an update note once it is confirmed by a primary source.<\/p>\n<\/blockquote>\n","protected":false},"excerpt":{"rendered":"<p>Explore Qwen3.8-Flash-Next-E384-FP4-Local, an unofficial NAS-optimized MoE model with reduced experts for efficient local inference.<\/p>\n","protected":false},"author":1,"featured_media":13137,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[310],"tags":[4024,165,4026,813,2967,4028,1565,592,1952],"class_list":["post-13138","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-new-models","tag-autotrust-en","tag-moe-en","tag-nas-en","tag-nvfp4-en","tag-qwen3-8-flash-next-en","tag-qwen3-8-flash-next-e384-fp4-local-en","tag-unverified","tag-vllm-en","tag-vlm-en"],"lang":"en","translations":{"en":13138,"ja":13136},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/13138","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/comments?post=13138"}],"version-history":[{"count":3,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/13138\/revisions"}],"predecessor-version":[{"id":18180,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/13138\/revisions\/18180"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media\/13137"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=13138"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/categories?post=13138"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/tags?post=13138"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}