Kimi K3.1 Pro’s Mixture‑of‑Experts backbone combines 896 experts with a one‑million token context window.
KIMI K3.1 PRO WEIGHTS RELEASED: 2.8T MODEL SHIFTS AI BENCHMARKS
*Moonshot AI has opened the weights of its Kimi K3.1 Pro, a 2.8‑trillion‑parameter multimodal model with a one‑million token context. The release lets anyone test a model that claims top scores on coding, 3D creation and finance benchmarks.*
Moonshot AI dropped the full weight files for Kimi K3.1 Pro today, turning a previously closed‑research asset into an open‑weight model. The model is positioned as the company’s most capable 3‑trillion‑class system, built on a Mixture‑of‑Experts (MoE) backbone and trained with large‑scale joint reinforcement learning across hundreds of downstream domains. Its one‑million‑token context window and vision encoder promise a leap in long‑form reasoning, coding assistance and multimodal creation. By publishing the weights, Moonshot invites the community to verify performance claims and to probe the security implications of a model that can operate at agentic scale.
Model Architecture and Scale
Kimi K3.1 Pro uses a Mixture‑of‑Experts design with 2.8 trillion total parameters but only 104 billion activated at inference. The network stacks 93 layers, of which 69 are KDA attention blocks and 24 are gated DeepSeek Sparse Attention (DSA) layers. It runs 96 attention heads, each 128 dimensions, and selects 16 of 896 experts per token, with two experts shared across all tokens. The context length is 1,048,576 tokens, enabling ultra‑long documents. A dedicated vision encoder, MoonViT‑V2, adds 401 million parameters and runs with MXFP4 weight quantization and MXFP8 activation quantization. The architecture replaces full‑attention Gated MLA layers with Gated DSA for more efficient long‑context modeling.
Benchmark Performance Across Domains
The released scores show Kimi K3.1 Pro leading on a range of public benchmarks. In coding, Terminal‑Bench 4.0 records a 65.7 max score, beating DeepSeek‑V4.1‑Flash‑1017’s 58.2. On BenchCAD with tools, Kimi hits 92.7, outpacing DeepSeek’s 81.6. Vision‑heavy tasks like BabyVision with Python reach 95.4, narrowly ahead of DeepSeek’s 95.1. Financial and scientific suites also show gains: Frontier Finance scores 58.6 versus DeepSeek’s 45.0, and Terminal‑Bench‑Science climbs to 68.1, eclipsing the 40.0 of Claude Opus. Security‑focused ExploitBench registers 74.7, higher than DeepSeek’s 83.0 but still competitive. Across the board, Kimi’s internal evaluations use Claude Code as the judge, ensuring a consistent scoring methodology.
How does Kimi K3.1 Pro compare to leading rivals?
Against the listed competitors—Claude Fable 5.1, Claude Opus 5.5, GPT‑6 Astra, DeepSeek‑V4.1‑Flash‑1017 and GLM‑5.3—Kimi K3.1 Pro posts higher max scores on most listed benchmarks. For example, BenchCAD w/ tools: Kimi 92.7 vs DeepSeek 81.6 and Claude Opus 58.8. Coding Terminal‑Bench: Kimi 65.7 vs Claude Opus 58.2 and GPT‑6 Astra 57.9. In the Agentic HLE‑Full suite with tools, Kimi scores 70.0, surpassing Claude Opus 57.2 and DeepSeek 67.7. The model also leads in multimodal tasks such as CharXiv with Python (93.4 vs DeepSeek 92.8) and MMMU‑Pro with Python (90.0 vs DeepSeek 87.8). These gaps suggest Kimi’s MoE scaling and long‑context design translate into measurable advantage across coding, reasoning and vision benchmarks.
Implications for AI Development and Security
Open‑weight release lowers the barrier for researchers to experiment with trillion‑parameter models, accelerating innovation in agentic AI. The one‑million token context could enable new applications in legal analysis, financial modeling and long‑form code generation. However, the same capabilities raise security concerns: a model that can reason over massive text and images may be repurposed for sophisticated phishing, disinformation or automated vulnerability discovery. The public availability of the weights also invites adversaries to fine‑tune the model for malicious ends. Moonshot’s decision to publish quantized weights (MXFP4/MXFP8) suggests an awareness of compute constraints, but does not mitigate the risk of misuse once the model is widely accessible.
Frequently asked questions
What is the size and context window of Kimi K3.1 Pro?
Kimi K3.1 Pro has 2.8 trillion total parameters, 104 billion activated parameters, and a 1‑million token context window.
How does Kimi K3.1 Pro perform on coding benchmarks compared to rivals?
On Terminal‑Bench 4.0 it scores 65.7, higher than DeepSeek‑V4.1‑Flash‑1017’s 58.2 and Claude Opus’s 58.2.
What vision encoder does Kimi K3.1 Pro use?
The model includes the MoonViT‑V2 vision encoder with 401 million parameters, quantized with MXFP4 weights and MXFP8 activations.
The Kimi K3.1 Pro release marks a watershed moment: a truly open, multimodal, trillion‑scale model now sits in the public domain. Its benchmark dominance forces competitors to rethink scaling strategies, while its accessibility forces policymakers and security teams to confront the dual‑use dilemma of powerful AI. The next weeks will reveal whether the community can harness its strengths responsibly or whether the model becomes a catalyst for new threat vectors.
Sources: Hacker News (Kimi K3.1 Pro model weights are out)