
DeepSeek V4 Drops Next Week - 1 Trillion Parameters on Chinese Chips
DeepSeek will release V4, a natively multimodal trillion-parameter model with a 1M token context window, in the first week of March - optimized for Huawei Ascend chips, not Nvidia.
They summarize our coverage. We write it.
Newsletters like this one rebroadcast our headlines - often without the full review, the source reading, or the analysis underneath. Our weekly briefing sends the work they paraphrase, straight from the desk, before they get to it.
Free, weekly, no spam. One email every Tuesday. Unsubscribe anytime.

DeepSeek will release V4, a natively multimodal trillion-parameter model with a 1M token context window, in the first week of March - optimized for Huawei Ascend chips, not Nvidia.

Google DeepMind's natively multimodal image generation and editing model built on Gemini 3.1 Flash - Pro-level quality at Flash speed, free for all Gemini users.

DeepSeek's V4 Lite model has leaked through inference provider testing under strict NDAs, revealing a 1M token context window, native multimodal capabilities, and the internal codename sealion-lite.

Moonshot AI's Kimi K2.5 is a 1T-parameter MoE model activating 32B per token with native multimodal vision via MoonViT-3D, Agent Swarm coordination of up to 100 sub-agents via PARL, and top-tier math and coding benchmarks under a modified MIT license.

Comparing Moonshot AI's 1T-parameter Kimi K2.5 with Google DeepMind's Gemma 3 27B - two multimodal open-weight models separated by 37x in parameter count but sharing a vision-first design philosophy.

Google's cheapest Gemini model pairs a 1M-token context window with $0.10/$0.40 per million token pricing, multimodal input, and 359 tokens/second throughput for high-volume production workloads.

Google Gemma 3 27B is a 27B dense multimodal model supporting text and vision with a 128K context window, 140+ languages, and single-GPU deployment - the most capable open model at its size class.

OpenAI's budget API workhorse pairs 128K context with $0.15/$0.60 per million token pricing, solid coding benchmarks, and the broadest third-party ecosystem of any small model.

Meta's Llama 4 Maverick packs 400B total parameters into a 128-expert MoE architecture with only 17B active per token, beating GPT-4o on Chatbot Arena while matching DeepSeek V3 on reasoning at half the active parameters.

Meta's Llama 4 Scout is a 109B-total, 17B-active MoE model with 16 experts and a 10M-token context window - the longest of any open-weight model - with native multimodal support for text and images.

Mistral Large 3 is a 675B-parameter MoE model activating 41B per token with native multimodal support, a 256K context window, and Apache 2.0 licensing - Europe's first frontier-class open-weight model.

Mistral Small 3.2 is a 24B dense model with strong function calling, multimodal vision, and 128K context under Apache 2.0 - optimized for production tool-use pipelines and EU-compliant deployments.