November 4 agent guide: Kimi K3.1 Pro weights and Tencent Hy4

November 4, 2026 is a Chinese open-frontier day for Bacancy Technology readers. Moonshot published the open weights it promised with the October 23 Kimi K3.1 Pro launch. Tencent shipped the full Hy4 API with vision while keeping open weights on a near-term schedule. Neither release is a reason to wipe a working planner overnight. Both are reasons to re-run cost-per-completed-task and long-horizon agent benches against the current default.

What shipped on November 4

Kimi K3.1 Pro: the weight drop that October 23 promised

Bacancy Technology already covered the Kimi K3.1 Pro product and API launch in the October 22–23 routing guide. November 4 is the weight drop. The Hugging Face card for moonshotai/Kimi-K3.1-Pro describes an open-weight, natively multimodal agentic model post-trained from Kimi K3 with joint reinforcement learning across hundreds of domains. Architecture highlights stay consistent with the K3 family: Mixture-of-Experts, 2.8T total parameters, 104B activated, 1,048,576-token context, MoonViT-V2 vision, and MXFP4 weights with MXFP8 activations from quantization-aware training.

What is technically significant

• Gated DeepSeek Sparse Attention replaces full-attention Gated MLA layers for long-context efficiency versus Kimi K3.
• Reasoning effort is request-level (low, high, max; default max). Thinking cannot be treated as optional glue.
• Multi-turn tool calls require the full assistant message, including reasoning_content and tool_calls, passed back as-is.
• Moonshot recommends vLLM, SGLang, and TokenSpeed for inference, with an OpenAI/Anthropic-compatible API on platform.kimi.ai as kimi-k3.1-pro.

Vendor-reported scores to re-verify

Moonshot reports Terminal-Bench 4.0 at 65.7, SWE Marathon v1.1 at 51.7, HLE-Full with tools at 70.0, BrowseComp at 92.1, and OSWorld V2.1 at 79.6 under max reasoning effort. Several coding scores use a Claude Code harness. Treat every figure as vendor-reported until an independent harness lands. The useful builder move is to re-run the same tasks used for the October 23 API review, now comparing hosted kimi-k3.1-pro against a self-hosted MXFP4 lane.

Tencent Hy4: full API with vision, weights next

Tencent released Hy4 on November 4 as the full successor to Hy4 preview. Compared with preview, Hy4 scales pre-training and post-training, adds vision for the first time, and targets instruction following, long-context understanding, long-horizon tasks, and deep research. The model is live on Tencent Cloud TokenHub (model id hy4), OpenRouter (tencent/hy4), and Tencent products including CodeBuddy, WorkBuddy, Yuanbao, and ima.

API shape and pricing

• 1M-token context window with up to about 960k input and 64k output on TokenHub docs.
• Compatible with OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages protocols.
• Tencent lists the same API price as Hy4 preview (about $0.834 input / $0.042 cached input / $2.501 output per million tokens on the Hy research page).
• Weights are not on Hugging Face yet. Tencent says open weights and a system card are coming, with vLLM and SGLang ecosystem support planned.

For Bacancy Technology clients, Hy4 is an API routing candidate today and an open-weight candidate only after the weight drop clears legal review. Do not write “open source” into an architecture diagram until the weights and license are published.

Secondary November 4 notes

OpenAI introduced Images 2.6 across ChatGPT, Codex, and the API, with GIF generation, image markup in the API, SVG export, and a complexity control alongside quality. API model ids are gpt-image-2.6-flare (fast iteration) and gpt-image-2.6-sunburst (detail and edit consistency). Claude Code 2.1.318 added cloud sessions without a repository, expanded deniedMcpServers URL matching to Anthropic-provided servers and connectors, and tightened several gateway and network-home reliability paths. Full detail for those items lives in the AI Updates ticker.

What this means for developers and businesses

• Developers: stand up a Kimi K3.1 Pro self-host lane if GPU budget exists; otherwise keep using the Kimi API and add Hy4 as a second Chinese long-context option.
• Businesses: refresh model-policy allowlists for kimi-k3.1-pro and hy4, and keep image pipelines ready for GIF and markup if marketing or design agents use OpenAI Images.
• Security and legal: review the Kimi K3.1 Pro License before self-host, and wait for Hy4’s system card before claiming open-weight compliance.

Builder checklist

1. 01
   Diff October 23 API results against November 4 weights

   Run the same Terminal-Bench and internal SWE set on hosted kimi-k3.1-pro and on a self-hosted MXFP4 build.
2. 02
   Add Hy4 to the API matrix only

   Call hy4 through TokenHub or OpenRouter with vision and tool tests. Park weight assumptions until Tencent publishes them.
3. 03
   Update image clients for Images 2.6

   Map Flare vs Sunburst, set complexity for edit loops, and decide whether GIF output belongs in any production workflow.

Conclusion

November 4 delivers the Kimi K3.1 Pro weight drop and a full Hy4 API with vision. Bacancy Technology reads both as open-frontier routing moves for agentic coding and long-horizon work. Keep the October 23 product context; do not re-announce the launch. Re-bench, then decide.

Related reading

• October 22–23 Kimi K3.1 Pro product launch: /blog/kimi-k31-pro-decisions-api-routing-october-2026
• November 3 Claude Code 2.1.317: /blog/november-03-claude-code-2-1-317-mcp-security-2026
• November 1 Mage-VL-Cowork: /blog/november-01-mage-vl-cowork-computer-use-2026
• Model routing note: /notes/stop-paying-frontier-prices-model-routing
• Agentic AI topic map: /topics/agentic-ai

Need help routing open-weight and Chinese API models?

Bacancy Technology designs model matrices, self-host plans, and agent evaluation harnesses for production teams. Consulting: /consulting/ . Contact: /contact/ .

Sources: Hugging Face ; hy.tencent.ai — hy4 ; tencent.com ; OpenAI ; Claude blog .