{"id":1269,"date":"2026-09-10T21:07:21","date_gmt":"2026-09-10T12:07:21","guid":{"rendered":"https:\/\/stratos-research.com\/?p=1269"},"modified":"2026-09-10T21:07:21","modified_gmt":"2026-09-10T12:07:21","slug":"deepseek-v4-flash-vision-exp-%e3%81%ae%e6%a6%82%e8%a6%81%e3%81%a8%e5%88%a9%e7%94%a8%e6%89%8b%e9%a0%86","status":"publish","type":"post","link":"https:\/\/stratos-research.com\/?p=1269","title":{"rendered":"DeepSeek\u2011V4\u2011Flash\u2011Vision\u2011Exp \u306e\u6982\u8981\u3068\u5229\u7528\u624b\u9806"},"content":{"rendered":"<h2>\u306f\u3058\u3081\u306b<\/h2>\n<p>DeepSeek\u2011V4\u2011Flash\u2011Vision\u2011Exp \u306f\u3001DeepSeek\u2011V4 \u7cfb\u5217\u3067\u521d\u3081\u3066\u5b9f\u9a13\u7684\u306b\u30de\u30eb\u30c1\u30e2\u30fc\u30c0\u30eb\u6a5f\u80fd\u3092\u7d44\u307f\u8fbc\u3093\u3060\u30e2\u30c7\u30eb\u3067\u3059\u3002DeepSeek\u2011V4\u2011Flash \u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u306b\u8996\u899a\u30e2\u30b8\u30e5\u30fc\u30eb\u3092\u8ffd\u52a0\u3057\u3001\u7d99\u7d9a\u7684\u306a\u5b66\u7fd2\u306b\u3088\u3063\u3066\u753b\u50cf\u7406\u89e3\u80fd\u529b\u3092\u7372\u5f97\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n<h2>\u5f93\u6765\u30e2\u30c7\u30eb\u3068\u306e\u6bd4\u8f03<\/h2>\n<p>\u540c\u3058\u30c6\u30ad\u30b9\u30c8\u30a8\u30fc\u30b8\u30a7\u30f3\u30c8\u30bf\u30b9\u30af\u3067\u306f DeepSeek\u2011V4\u2011Flash\u20110731 \u3068\u307b\u307c\u540c\u7b49\u306e\u6027\u80fd\u3092\u4fdd\u3061\u3064\u3064\u3001\u30de\u30eb\u30c1\u30e2\u30fc\u30c0\u30eb\u30a8\u30fc\u30b8\u30a7\u30f3\u30c8\u80fd\u529b\u304c\u5927\u5e45\u306b\u5411\u4e0a\u3057\u307e\u3057\u305f\u3002<\/p>\n<h2>\u8a55\u4fa1\u6761\u4ef6<\/h2>\n<p>\u30c6\u30ad\u30b9\u30c8\u30a8\u30fc\u30b8\u30a7\u30f3\u30c8\u30d9\u30f3\u30c1\u30de\u30fc\u30af\u306f\u3001DeepSeek Harness \u306e\u6700\u5c0f\u69cb\u6210\u3067\u3001\u6700\u5927\u63a8\u8ad6\u52aa\u529b\u30ec\u30d9\u30eb\uff08temperature=1.0\u3001top_p=0.95\uff09\u3067\u8a55\u4fa1\u3055\u308c\u3066\u3044\u307e\u3059\u3002ApexBench \u304a\u3088\u3073 Agents&#8217; Last Exam \u3067\u306f\u3001DeepSeek\u2011V4\u2011Flash\u20110731 \u304c\u5165\u529b\u4e2d\u306e\u30de\u30eb\u30c1\u30e2\u30fc\u30c0\u30eb\u8981\u7d20\u3092\u7121\u8996\u3057\u307e\u3059\u3002<\/p>\n<h2>\u30ea\u30dd\u30b8\u30c8\u30ea\u69cb\u6210<\/h2>\n<p>\u672c\u30ea\u30dd\u30b8\u30c8\u30ea\u306f\u30c8\u30fc\u30af\u30ca\u30a4\u30b6\u30fc\u3001\u30d7\u30ed\u30f3\u30d7\u30c8\u30a8\u30f3\u30b3\u30fc\u30c7\u30a3\u30f3\u30b0\u53c2\u7167\u3001\u304a\u3088\u3073 DeepSeek\u2011V4 Flash Vision \u7528\u306e\u6700\u5c0f PyTorch \u63a8\u8ad6\u5b9f\u88c5\u3092\u542b\u307f\u307e\u3059\u3002<\/p>\n<ul>\n<li><code>encoding\/<\/code>: OpenAI \u5f62\u5f0f\u306e\u30e1\u30c3\u30bb\u30fc\u30b8\u3092\u30e2\u30c7\u30eb\u30d7\u30ed\u30f3\u30d7\u30c8\u3078\u5909\u63db\u3059\u308b\u30c7\u30a3\u30ec\u30af\u30c8\u30ea<\/li>\n<li><code>inference\/<\/code>: \u91cd\u307f\u5909\u63db\u3068\u6700\u5c0f\u63a8\u8ad6\u5b9f\u88c5\u3001\u30b5\u30f3\u30d7\u30eb\u30d7\u30ed\u30f3\u30d7\u30c8\u4f8b\u3092\u683c\u7d0d<\/li>\n<li>\u305d\u306e\u4ed6: <code>config.json<\/code>, <code>generation_config.json<\/code>, <code>model.safetensors.index.json<\/code>, <code>tokenizer.json<\/code>, <code>tokenizer_config.json<\/code><\/li>\n<\/ul>\n<p>\u30a8\u30f3\u30b3\u30fc\u30c7\u30a3\u30f3\u30b0\u3068\u63a8\u8ad6\u306f\u610f\u56f3\u7684\u306b\u5206\u96e2\u3055\u308c\u3066\u304a\u308a\u3001\u30d7\u30ed\u30f3\u30d7\u30c8\u30d5\u30a9\u30fc\u30de\u30c3\u30c8\u306f PyTorch \u306b\u4f9d\u5b58\u305b\u305a\u3001\u63a8\u8ad6\u5074\u304c\u660e\u793a\u7684\u306b\u30a8\u30f3\u30b3\u30fc\u30c7\u30a3\u30f3\u30b0\u30e2\u30b8\u30e5\u30fc\u30eb\u3092\u30a4\u30f3\u30dd\u30fc\u30c8\u3057\u307e\u3059\u3002\u30b7\u30f3\u30dc\u30ea\u30c3\u30af\u30ea\u30f3\u30af\u306f\u4e0d\u8981\u3067\u3059\u3002<\/p>\n<h2>\u30d7\u30ed\u30f3\u30d7\u30c8\u30a8\u30f3\u30b3\u30fc\u30c7\u30a3\u30f3\u30b0<\/h2>\n<p>OpenAI \u30b9\u30bf\u30a4\u30eb\u306e JSON \u30b3\u30f3\u30c6\u30f3\u30c4\u30d6\u30ed\u30c3\u30af\u3068\u3001\u30b3\u30f3\u30d1\u30af\u30c8\u306a <code>&lt;image&gt;path&lt;\/image&gt;<\/code> \u30c6\u30ad\u30b9\u30c8\u8868\u8a18\u306e\u4e21\u65b9\u304c\u30b5\u30dd\u30fc\u30c8\u3055\u308c\u3066\u3044\u307e\u3059\u3002<code>inference\/examples\/<\/code> \u914d\u4e0b\u306e\u4f8b\u306f\u3001\u3069\u3061\u3089\u3082\u540c\u4e00\u306e\u30d7\u30ed\u30f3\u30d7\u30c8\u3068\u30c8\u30fc\u30af\u30f3 ID \u3092\u751f\u6210\u3057\u307e\u3059\u3002<\/p>\n<h2>\u6700\u5c0f\u63a8\u8ad6<\/h2>\n<p>\u4f9d\u5b58\u95a2\u4fc2\u306e\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb\u3001\u30c1\u30a7\u30c3\u30af\u30dd\u30a4\u30f3\u30c8\u5909\u63db\u3001TXT\uff0fJSON \u63a8\u8ad6\u30b3\u30de\u30f3\u30c9\u306b\u3064\u3044\u3066\u306f <code>inference\/README.md<\/code> \u3092\u53c2\u7167\u3057\u3066\u304f\u3060\u3055\u3044\u3002<\/p>\n<h2>vLLM \u3067\u306e\u5b9f\u884c\u4f8b<\/h2>\n<p>\u5358\u4e00\u30ce\u30fc\u30c9\uff084\u00d7GB300\uff09\u3067\u30e2\u30c7\u30eb\u3092\u63d0\u4f9b\u3059\u308b\u30b3\u30de\u30f3\u30c9\u4f8b\u3067\u3059\u3002<\/p>\n<pre><code>docker run --gpus all \\\n vllm\/vllm-openai:deepseekv4-flash-vision deepseek-ai\/DeePSeek-V4-Flash-Vision-Exp \\\n --kv-cache-dtype fp8 \\\n --block-size 256 \\\n --tensor-parallel-size 4 \\\n --tool-call-parser deepseek_v4 \\\n --enable-auto-tool-choice \\\n --reasoning-parser deepseek_v4 \\\n --reasoning-config '{\"reasoning_parser\":\"deepseek_v4\",\"reasoning_start_str\":\"\",\"reasoning_end_str\":\"\"}' \\\n --speculative-config '{\"method\":\"dsspark\",\"model\":\"deepseek-ai\/DeePSeek-V4-Flash-Vision-Exp\",\"num_speculative_tokens\":3,\"draft_sample_method\":\"probabilistic\",\"enable_adaptive_verification\":true}'<\/code><\/pre>\n<h2>SGLang \u3067\u306e\u5b9f\u884c\u4f8b<\/h2>\n<p>DSPARK \u3092\u6709\u52b9\u5316\u3057\u3001\u540c\u4e00\u30c1\u30a7\u30c3\u30af\u30dd\u30a4\u30f3\u30c8\u304b\u3089\u30bf\u30fc\u30b2\u30c3\u30c8\u3068\u30c9\u30e9\u30d5\u30c8\u306e\u91cd\u307f\u3092\u53d6\u5f97\u3057\u307e\u3059\u3002<\/p>\n<pre><code>sglng serve \\\n --model-path deepseek-ai\/DeePSeek-V4-Flash-Vision-Exp \\\n --tp 4 \\\n --speculative-algorithm DSPARK \\\n --mem-fraction-static 0.85 \\\n --host 0.0.0.0 \\\n --port 30000<\/code><\/pre>\n<h2>\u30e9\u30a4\u30bb\u30f3\u30b9<\/h2>\n<p>\u672c\u30ea\u30dd\u30b8\u30c8\u30ea\u306f\u30e9\u30a4\u30bb\u30f3\u30b9\u60c5\u5831\u3092\u542b\u3093\u3067\u3044\u307e\u3059\u3002<\/p>\n<p class=\"source-link\">\u5143\u8a18\u4e8b: <a href=\"https:\/\/huggingface.co\/deepseek-ai\/DeepSeek-V4-Flash-Vision-Exp\" rel=\"noopener noreferrer\">https:\/\/huggingface.co\/deepseek-ai\/DeepSeek-V4-Flash-Vision-Exp<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u306f\u3058\u3081\u306b DeepSeek\u2011V4\u2011Flash [&hellip;]<\/p>\n","protected":false},"author":3,"featured_media":1270,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"_spp_pexels_query":"","_jetpack_feature_clip_id":0,"_jetpack_memberships_contains_paid_content":false,"footnotes":"","jetpack_post_was_ever_published":false},"categories":[2],"tags":[],"class_list":["post-1269","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-technology"],"jetpack_sharing_enabled":true,"jetpack_featured_media_url":"https:\/\/stratos-research.com\/wp-content\/uploads\/2026\/09\/painted-1268-P048cj.webp","_links":{"self":[{"href":"https:\/\/stratos-research.com\/index.php?rest_route=\/wp\/v2\/posts\/1269","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/stratos-research.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/stratos-research.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/stratos-research.com\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/stratos-research.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=1269"}],"version-history":[{"count":0,"href":"https:\/\/stratos-research.com\/index.php?rest_route=\/wp\/v2\/posts\/1269\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/stratos-research.com\/index.php?rest_route=\/wp\/v2\/media\/1270"}],"wp:attachment":[{"href":"https:\/\/stratos-research.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=1269"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/stratos-research.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=1269"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/stratos-research.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=1269"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}