Moonshot AI releases 2.8T-parameter Kimi K3.1 Pro weights on Hugging Face
Moonshot AI published the weights and documentation for Kimi K3.1 Pro on Hugging Face, an open-weight, natively multimodal agentic model with 2.8 trillion total parameters and 104 billion activated per token. The Mixture-of-Experts model offers a 1,048,576-token context window, 896 experts, 93 layers and a MoonViT-V2 vision encoder, shipping at 4-bit precision in a roughly 1.56 TB repository. Moonshot claims gains in coding, 3D creation, financial research and agentic tasks, with mixed benchmark results against rival frontier models, under a modified MIT license.
IllustrationEditorial responsibility
- No named human review is recorded for this page.
- Reports are grouped by semantic similarity and deterministic rules. Language models may assist titles, summaries, translation and cross-source analysis; the page reads the event directly, while its address stays stable when the title changes.
- Summary covers the current reports
Cross-source coverage
Reporting timeline
Kimi K3.1 Pro open weights released on HuggingFace under new license
Kimi K3.1 Pro model weights are now available on HuggingFace under a license named after the model, according to a post from the account testingcatalog. The post quotes a description stating that Kimi K3.1 Pro is an open-weight, natively multimodal agentic model and is the developer's most capable 3T-class model to date. The announcement does not specify benchmark results, parameter counts beyond the 3T-class characterization, licensing terms, or intended use restrictions; those details would presumably be found on the HuggingFace repository linked in the post. The release makes the model's weights publicly downloadable, a step beyond API-only access, and positions it as a large-scale multimodal agentic system. No independent verification, performance comparisons, or deployment details are included in the source item.
Read sourceMoonshot AI Releases 2.8T-Parameter Kimi K3.1 Pro Weights on Hugging Face
Moonshot AI released the weights of its 2.8T-parameter Kimi K3.1 Pro on Hugging Face, shipping at 4-bit precision to avoid the 5.6TB a 16-bit copy would require. In Moonshot's own tests the model scores 65.7% on Terminal-Bench 4.0, close to Claude Opus 5.5's 66.4%, but reaches only 39.4% on HLE-Diamond without tools versus GPT-6 Astra's 65.8%. The model activates 104B parameters per token, and 69 of its 93 layers use linear attention, making 1M-token context cheaper to process. Moonshot said it traced rising exploit skill to cyber content in its mid-training corpus and RL environments and removed both before opening the weights. A modified MIT license exempts internal use but requires model-hosting firms earning over $20M a year to negotiate, limiting resellers that could undercut its API.
Read sourceMoonshot AI Releases Kimi-K3.1-Pro With 2.8T Parameters and 1M-Token Context
Moonshot AI released Kimi-K3.1-Pro, a sparse model with 2.8T total parameters but only 104B active per token, making it one of the largest real model releases on the Hub alongside Kimi-K3. It uses 896 experts with 16 activated per token and a 1M-token context handled via linear attention (KDA) in most of its 93 layers, with every fourth layer using DeepSeek-style sparse attention over the top 2048 tokens. It employs a SiTU-GLU activation and a 401M-parameter MoonViT-V2 vision encoder for text and images. Trained in 4-bit MXFP4, the repo is about 1.56 TB versus roughly 5.6 TB in BF16. On the model card benchmarks it trades places with Claude Opus 5.5 and GPT-6 Astra, topping BrowseComp (92.1), HLE with tools (70.0), and SWE Marathon (51.7), while being weaker on pure reasoning without tools and cybersecurity. The modified MIT license requires deals for API businesses over $20M annual revenue and attribution for products over 100M monthly users or $20M monthly revenue.
Read sourceShow 3 older updatesHide older updates
Moonshot AI releases Kimi K3.1 Pro weights with 2.8 trillion parameters
Moonshot AI has released the weights of its Kimi K3.1 Pro model on Hugging Face, according to a post on X. The model reportedly has 2.8 trillion parameters, with the weights alone taking up around 1.56 TB. For local operation, the post suggests a possible setup of a server with 8 NVIDIA B300 GPUs and 2.3 TB of GPU memory, with an estimated hardware cost of $550,000–$650,000. The post frames local deployment as demanding and closes with a wry "Good luck."
Read sourceMoonshot AI Releases Kimi K3.1 Pro as Open-Weight Model on Hugging Face
Moonshot AI has officially uploaded its Kimi K3.1 Pro model to Hugging Face as an open-weight release, according to a post on X. The model is described as having 2.8T total parameters with 104B activated parameters and a 1M token context window. The post frames the upload as making Kimi K3.1 Pro available as open weights, though no license terms, benchmarks, or performance comparisons are provided in the source text. The announcement centers on the release itself and the model's stated scale and context capacity.
Read sourceMoonshot AI Releases Open-Weight Kimi K3.1 Pro With 2.8T Parameters
Moonshot AI published the model weights and documentation for Kimi K3.1 Pro, described as an open-weight, natively multimodal agentic model and its most capable 3T-class model to date. Post-trained from Kimi K3 with large-scale joint reinforcement learning across hundreds of downstream domains, it is the first model in the K3.1 series and offers a 1,048,576-token context window. The Mixture-of-Experts architecture totals 2.8T parameters with 104B activated, 93 layers, 896 experts, and a MoonViT-V2 vision encoder; compared with Kimi K3, full-attention Gated MLA layers are replaced by Gated DeepSeek Sparse Attention layers for more efficient long-context modeling. Moonshot claims substantial gains in coding, 3D creation, financial research, factual accuracy, and agentic capabilities. Benchmark tables compare it against Claude Fable 5.1, GPT-6 Astra, Claude Opus 5.5, DeepSeek-V4.1-Flash-1017, and GLM-5.3, with mixed results across coding, agentic, reasoning, and cybersecurity evaluations.