{"slug": "deepseek-v4-1-flash-hits-848-tokens-per-second-on-nvidias-gb300-nvl72", "title": "DeepSeek V4.1 Flash hits 848 tokens per second on Nvidia’s GB300 NVL72", "summary": "DeepSeek V4.1 Flash now generates 848 tokens per second per agent on Nvidia's GB300 NVL72 system, up from 596 tokens per second roughly a month earlier — a gain of 252 tokens per second on unchanged hardware. The 552-billion-parameter multimodal Mixture-of-Experts model, which launched around September 10, 2026 and received a version update labeled 1017 in mid-to-late October 2026, activates 8 billion parameters during prefill and 16 billion during decoding, supports a 1-million-token context window, and compresses its KV cache to roughly 890 bytes per token. The reported gains point to software-layer improvements in serving stacks, kernels and model tweaks rather than new silicon, while the article notes that aggregate throughput above 1 million tokens per second per GPU and about $0.002 per million tokens apply only to specialized configurations and large-hyperscaler pricing.", "body_md": "# DeepSeek V4.1 Flash hits 848 tokens per second on Nvidia’s GB300 NVL72\n\nPhoto: Merlin Lightpainting / Pexels\n\nPer-agent decode speed climbed from 596 tokens per second in roughly a month, showing how fast software and hardware tuning is moving.\n\n[DeepSeek](https://cryptobriefing.com/markets/deepseek/) V4.1 Flash now generates 848 tokens per second per agent on [Nvidia](https://cryptobriefing.com/markets/nvidia/)’s GB300 NVL72 system. A month ago, that number was 596.\n\nThat is an extra 252 tokens per second without swapping out the hardware.\n\n## What changed in a month\n\nThe figure that matters here is per-agent decode speed. Think of it as how fast a single AI session can type out its answer once it starts responding.\n\nGoing from 596 to 848 tokens per second on the same GB300 NVL72 setup points to gains in the software layer. That covers serving stacks, kernels, and model-level tweaks, not new silicon.\n\nThe research findings frame this as a joint effort across model design, serving software, and hardware. No single breakthrough did the heavy lifting.\n\nThere are bigger numbers floating around too. In specialized configurations, aggregate throughput reportedly topped 1 million tokens per second per GPU.\n\nThe reported cost in those setups is about $0.002 per million tokens. That estimate applies under large-hyperscaler pricing scenarios, so your average startup should not expect that invoice.\n\n## Inside the model and the machine\n\nDeepSeek V4.1 Flash is a multimodal Mixture-of-Experts model with 552 billion total parameters. It launched around September 10, 2026.\n\n### AI, tech, and the markets they move—in one daily briefing.\n\nDaily. Free. Join 34,000+ readers across crypto, finance, and policy.\n\nIn this model, only 8 billion parameters are active during prefill, the stage where it reads your prompt. During decoding, when it writes the response, that rises to 16 billion active parameters.\n\nThe model uses a causal encoder-decoder architecture. It handles both text and image inputs and supports a context window of 1 million tokens.\n\nDeepSeek compresses the KV cache down to roughly 890 bytes per token. Smaller memory per token means more sessions can share the same hardware before things get crowded.\n\nOn the hardware side, the GB300 NVL72 packs 72 GPUs into a single rack. The findings credit it with improvements over its predecessor, the GB200 NVL72, particularly for workloads with high concurrency and long contexts.\n\n## A model that keeps getting updates\n\nDeepSeek has not treated V4.1 Flash as a finished product. The model received a version update labeled 1017 in mid-to-late October 2026.\n\nThe findings describe DeepSeek-AI’s strategy as focused on fast inference. The 848 tokens per second figure is a fairly direct expression of that priority.\n\n## What this means\n\nFor companies building AI products, the headline number is less important than the trajectory. If per-agent speed can climb by 252 tokens per second in a month on fixed hardware, capacity planning gets a lot harder to pin down.\n\nPer-agent speed and aggregate throughput also measure different things. One tracks how fast a single session responds, while the other tracks total output across the system, so they should not be read as interchangeable.\n\nThere are caveats worth keeping in view. The most eye-catching throughput and cost numbers come from specialized configurations and large-hyperscaler pricing, not typical deployments.\n\n**Disclosure:** This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our [Editorial Policy](https://cryptobriefing.com/editorial-policy/).", "url": "https://wpnews.pro/news/deepseek-v4-1-flash-hits-848-tokens-per-second-on-nvidias-gb300-nvl72", "canonical_source": "https://cryptobriefing.com/deepseek-v4-1-flash-gb300-nvl72-speed/", "published_at": "2026-11-06 20:37:32+00:00", "updated_at": "2026-11-06 20:51:02.506390+00:00", "lang": "en", "topics": ["large-language-models", "ai-infrastructure", "ai-chips", "machine-learning"], "entities": ["DeepSeek", "DeepSeek V4.1 Flash", "Nvidia", "GB300 NVL72", "GB200 NVL72", "DeepSeek-AI"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/deepseek-v4-1-flash-hits-848-tokens-per-second-on-nvidias-gb300-nvl72", "markdown": "https://wpnews.pro/news/deepseek-v4-1-flash-hits-848-tokens-per-second-on-nvidias-gb300-nvl72.md", "text": "https://wpnews.pro/news/deepseek-v4-1-flash-hits-848-tokens-per-second-on-nvidias-gb300-nvl72.txt", "jsonld": "https://wpnews.pro/news/deepseek-v4-1-flash-hits-848-tokens-per-second-on-nvidias-gb300-nvl72.jsonld"}}