PulseAugur
EN
LIVE 13:33:35
Deutsch(DE) NVIDIA quantisiert Kimi-K2.7-Code (DeepSeek-V3-Architektur, 1T/32B aktiv) via Model Optimizer auf NVFP4. Text/Bild/Video, 256k Kontext, Serving über vLLM auf B2

NVIDIA releases Kimi-K2.7-Code with DeepSeek-V3 architecture

By PulseAugur Editorial · [2 sources] ·

NVIDIA has released Kimi-K2.7-Code, an open-source model based on the DeepSeek-V3 architecture. This model features 32 billion active parameters and a 256,000 token context window. It utilizes speculative decoding within the vLLM framework and has been quantized for deployment on NVFP4 hardware, supporting text, image, and video modalities. Performance evaluations include SWE-bench Verified and Terminal-Bench 2.1, though the training data is noted to contain toxic content. AI

IMPACT This release offers a large context window and multimodal capabilities, potentially advancing research and applications in complex data processing.

RANK_REASON Frontier-lab model release with system card. [lever_c_demoted from frontier_release: ic=2 ai=1.0]

Read on Mastodon — mastodon.social →

AI-generated summary · Google Gemini · from 2 sources. How we write summaries →

NVIDIA releases Kimi-K2.7-Code with DeepSeek-V3 architecture

COVERAGE [2]

  1. Mastodon — mastodon.social TIER_1 Deutsch(DE) · aisyndicate ·

    NVIDIA Publishes DFlash Draft Head for Kimi-K2.7 Code (DeepSeek-V3 Architecture, 32B Active, 256K Context). It Enables Speculative Decoding in vLLM

    NVIDIA publiziert den DFlash-Draft-Head zu Kimi-K2.7-Code (DeepSeek-V3-Architektur, 32B aktiv, 256K Kontext). Er aktiviert das spekulative Dekodierung in vLLM auf Blackwell B200, wobei SPEED-Bench eine durchschnittliche Akzeptanzrate von 3,13 Tokens pro Draft-Schritt misst. Lizen…

    LINKS huggingface.co/…/Kimi-K2.7-Code-DFlash
  2. Mastodon — mastodon.social TIER_1 Deutsch(DE) · aisyndicate ·

    NVIDIA quantizes Kimi-K2.7 code (DeepSeek-V3 architecture, 1T/32B active) via Model Optimizer on NVFP4. Text/Image/Video, 256k context, serving via vLLM on B2

    NVIDIA quantisiert Kimi-K2.7-Code (DeepSeek-V3-Architektur, 1T/32B aktiv) via Model Optimizer auf NVFP4. Text/Bild/Video, 256k Kontext, Serving über vLLM auf B200. Evaluierung: SWE-bench Verified, Terminal-Bench 2.1. Trainingsdaten enthalten toxische Inhalte. https:// huggingface…

    LINKS huggingface.co/…/Kimi-K2.7-Code-NVFP4

RELATED ENTITIES

RELATED TOPICS