Muse Wonder: Meta Omni-Modal Architecture, Benchmarks & API Guide
Meta's Muse Wonder launched Oct 12, 2026 with a 1M omni-modal context window, Active Memory & Active Focus, 56 AA Index (Full) / 52 API, and $4/$20 pricing.
Kimi K3.1 Pro: 2.8T Sparse MoE Architecture, Benchmarks & API Guide
Kimi K3.1 Pro by Moonshot AI: 2.8T sparse MoE (104B active), 1M context, 67.2% Terminal-Bench 4.0, AA Index 55, $2/$10 API pricing, and open weights.
Executive Summary
A systems-level architectural breakdown of Kimi K3.1 Pro covering its 2.8-trillion-parameter sparse Mixture-of-Experts routing, 104B active per-token footprint, 67.2% Terminal-Bench 4.0 leap, 92.7% BenchCAD tool accuracy, Highspeed API economics, and November 4 open-weight release.
What Is Kimi K3.1 Pro? Architectural & Benchmark Overview
Kimi K3.1 Pro debuted in general availability on October 23, 2026, marking Moonshot AI's boldest leap into ultra-scale sparse Mixture-of-Experts modeling. Built on a 2.8-trillion-parameter (2.8T) weight graph activating 104B parameters per token, Kimi K3.1 Pro ships across two API endpoints—standard kimi-k3.1-pro and latency-accelerated kimi-k3.1-pro-highspeed—with full open weights scheduled by November 4, 2026. On the Artificial Analysis Intelligence Index, Kimi K3.1 Pro scores 55 to claim Rank #4 globally at a $2.11 task cost, standing as the highest-scoring open-weight-bound frontier architecture in history. Crucially, Moonshot AI resolved its predecessor's CLI bottleneck: Terminal-Bench 4.0 surges from Kimi K3's 13% to 65.7%–67.2%, beating GPT-6 Astra (58.2%) and matching Claude Opus 5.5 (66.4%). Paired with a 1,000,000-token context window, 128,000 max output tokens, multimodal vision, 73% AutomationBench-AA, and 92.7% BenchCAD 3D reconstruction, Kimi K3.1 Pro powers Kimi Work v3.1.0+ and Kimi Code.
Key Takeaways
2.8-Trillion Sparse MoE with 104B Active Parameters
Kimi K3.1 Pro scales total capacity to 2.8T parameters while routing 104B active parameters per token (~3.7% sparsity), balancing deep world knowledge with tractable multi-node H200/B200 inference.
5x Terminal-Bench 4.0 Surge to 65.7%–67.2%
Multi-environment reinforcement learning propels Kimi K3.1 Pro from Kimi K3's 13% baseline to 65.7%–67.2% on Terminal-Bench 4.0, beating GPT-6 Astra (58.2%) and rivaling Claude Opus 5.5 (66.4%).
Global Rank #4 at AA Intelligence Index 55
Independent evaluation places Kimi K3.1 Pro at 55 on the Artificial Analysis Intelligence Index with a $2.11 task cost, alongside 52% on Humanity's Last Exam and 61% on SciCode.
92.7% BenchCAD & 73% AutomationBench-AA Mastery
Tool-augmented visual-spatial reasoning in Kimi K3.1 Pro lifts BenchCAD 3D CAD reconstruction to 92.7%, while AutomationBench-AA climbs from 58% to 73% for enterprise orchestration.
Dual API Tiers with $0.20/M Context Caching
Standard Kimi K3.1 Pro (kimi-k3.1-pro) bills $2.00/M input and $10.00/M output, whereas kimi-k3.1-pro-highspeed delivers 2x–3x faster streaming at $4.00/$20.00—both sharing $0.20/M cached reads.
Open Weights Committed by November 4, 2026
Moonshot AI scheduled full open-weight checkpoint publication for Kimi K3.1 Pro by November 4, 2026, complementing native availability inside Kimi Work desktop (v3.1.0+) and Kimi Code.
Architectural & Engineering Deep Dive
Ultra-Sparse 2.8T MoE Routing and 104B Active Compute Budget
Scaling a foundation model to 2.8 trillion parameters without collapsing inference economics requires aggressive expert granularity and communication-aware load balancing. Kimi K3.1 Pro activates 104B parameters per forward token—roughly 3.71% of its total weight matrix—combining shared dense foundational projections with hundreds of specialized routed experts per layer. Whereas lightweight 8B–16B active MoE models must compress domain heuristics into narrow per-token pathways, a 104B active envelope gives Kimi K3.1 Pro dense-class representational capacity during every deductive step. That per-token width directly explains its simultaneous strength on abstract academic synthesis (52% on Humanity's Last Exam and 61% on SciCode) and messy stateful shell debugging. For latency-critical IDE loops, Moonshot AI's kimi-k3.1-pro-highspeed endpoint deploys speculative multi-token drafting and dedicated NVLink expert-parallel placement to stream tokens 2x to 3x faster at $4.00/$20.00 per million tokens.
Closing the Agentic Execution Gap: From 13% to 67.2% on Terminal-Bench 4.0
When the earlier Kimi K3 generation debuted, researchers observed a stark divergence between static benchmark intelligence and live command-line survival, where K3 scored just 13% on Terminal-Bench 4.0 due to premature loop termination, destructive shell assumptions, and poor stderr recovery. To build Kimi K3.1 Pro, Moonshot AI overhauled post-training around large-scale containerized RL environments spanning Linux system administration, compiler toolchains, browser/desktop GUI automation, and stateful Python CAD kernels. Rather than supervising synthetic reasoning traces, the training pipeline rewarded verified environment end-states across dozens of sequential tool invocations. Consequently, Terminal-Bench 4.0 surged to 65.7%–67.2%—vaulting past OpenAI's GPT-6 Astra (58.2%) and matching Anthropic's Claude Opus 5.5 (66.4%)—while AutomationBench-AA jumped 15 points from 58% to 73%.
1M-Token Context Caching, Multimodal CAD Grounding & Open-Weight Sovereignty
Long-horizon agents frequently accumulate hundreds of thousands of tokens across repository files, terminal stdout logs, and rendered viewport screenshots. Kimi K3.1 Pro pairs a native 1,000,000-token context window and 128,000-token completion ceiling with aggressive prefix cache reuse priced at $0.20 per million tokens across both Standard and Highspeed endpoints—a 90% discount off standard $2.00/M input rates and a 95% discount off Highspeed's $4.00/M input rate. Its native vision encoder maps geometric primitives directly into the shared MoE latent space, enabling iterative render-inspect-edit cycles that drive its landmark 92.7% BenchCAD score. Crucially, with Moonshot AI committing full open weights by November 4, 2026, sovereign cloud operators and national research labs can run a Rank #4 global frontier model entirely inside private multi-node GPU clusters.
Comprehensive Model Specifications
| Specification Dimension | Architecture & Serving Value | Technical Note & Evidence |
|---|---|---|
| Developer & API Model IDs | Moonshot AI (kimi-k3.1-pro / kimi-k3.1-pro-highspeed) | Dual-tier Kimi K3.1 Pro API serving standard and 2x–3x high-velocity streams |
| Release Date & Open-Weight Schedule | October 23, 2026 GA | Open Weights by November 4, 2026 | Immediate cloud availability followed by full checkpoint release within 12 days |
| Backbone Neural Architecture | 2.8-Trillion (2.8T) Sparse Mixture-of-Experts (104B Active/Token) | Fine-grained expert routing activating ~3.71% of total parameters per token |
| Context Window & Completion Limit | 1,000,000 Input Tokens / 128,000 Max Output Tokens | Lossless 1M repository ingestion with 128K single-turn generation span |
| Supported Input & Output Modalities | Input: Text, Image | Output: Text (Structured Tool & Code) | Native high-resolution vision encoder coupled to tool-use reasoning |
| Artificial Analysis Intelligence Index | 55 (Global Rank #4 | $2.11 Task Cost) | Positions Kimi K3.1 Pro as the top open-weight-bound frontier model globally |
| Agentic CLI & Automation Benchmarks | Terminal-Bench 4.0: 65.7%–67.2% | AutomationBench-AA: 73% | Up from 13% Terminal-Bench and 58% AutomationBench in Kimi K3 |
| Scientific & 3D Spatial Benchmarks | HLE: 52% | SciCode: 61% | BenchCAD (with Tools): 92.7% | Frontier academic deduction paired with parametric 3D CAD reconstruction |
| API Pricing & Desktop Ecosystem | $2/$10 per M (Standard) | $4/$20 (Highspeed) | $0.20/M Cache | First-party Kimi K3.1 Pro integration in Kimi Work (v3.1.0+) and Kimi Code |
Head-to-Head Performance & Benchmark Matrix
| Dimension | Kimi K3.1 Pro | Baseline / Competitor | Comparative Verdict |
|---|---|---|---|
| Generation-over-Generation Leap (vs Kimi K3) | 65.7%–67.2% Terminal-Bench 4.0 | 73% AutomationBench-AA | 13.0% Terminal-Bench 4.0 | 58% AutomationBench-AA | Kimi K3.1 Pro delivers a 5x terminal execution breakthrough and a +15-point automation gain over Kimi K3. |
| Terminal Agent Accuracy (vs GPT-6 Astra) | 65.7%–67.2% Terminal-Bench 4.0 ($2.00/$10.00 per M) | 58.2% Terminal-Bench 4.0 (Proprietary Cloud Only) | Kimi K3.1 Pro beats GPT-6 Astra by up to 9 percentage points on Linux shell harnesses while committing to open weights. |
| Frontier Reasoning Parity (vs Claude Opus 5.5) | AA Index 55 (#4) | 67.2% Peak Terminal-Bench | $2/$10 per M | AA Index 58 (#1) | 66.4% Terminal-Bench | $15/$75 per M | Kimi K3.1 Pro matches Opus 5.5 on terminal execution at 7.5x lower input/output API pricing. |
| Open-Ecosystem Scale (vs DeepSeek V4.1 Flash 1017) | 2.8T Total / 104B Active | AA Index 55 | HLE 52% | 552B Total / 8B–16B Active | AA Index 45 | HLE 42.6% | DeepSeek wins ultra-low-cost throughput ($0.26 task cost), whereas Kimi K3.1 Pro wins peak frontier intelligence (+10 AA points). |
| Serving Tier Trade-Off (Standard vs Highspeed) | kimi-k3.1-pro ($2.00/M In, $10.00/M Out, $0.20/M Cache) | kimi-k3.1-pro-highspeed ($4.00/M In, $20.00/M Out, $0.20/M Cache) | Kimi K3.1 Pro Highspeed accelerates interactive Kimi Code loops by 2x–3x while keeping $0.20/M cache reads. |
Real-World Implementation & Hands-on Verification
Multi-Container Kernel & Build Remediation in Kimi Code
Scenario Evaluation: A platform engineering team runs Kimi K3.1 Pro inside the Kimi Code CLI harness to repair a broken C++23 Bazel monorepo failing with glibc symbol mismatches and cgroup memory throttling inside rootless Docker.
Standardized Benchmark Prompt:
text
Inspect the failing Bazel sandboxes, reconcile toolchain crosstool definitions, patch container ulimits, and validate all 148 integration targets without disabling hermetic builds.Empirical Output Summary: Routing through kimi-k3.1-pro-highspeed with 410K cached repository tokens, Kimi K3.1 Pro inspected linker maps, patched the LLVM sysroot config, tuned container memory limits, and passed all 148 targets in 34 tool turns.
Evaluation Verdict: Validated the jump from Kimi K3's brittle shell recovery (13%) to 67.2% Terminal-Bench 4.0 reliability with 2.4x faster interactive streaming.
Parametric 3D Mechanical Assembly Reconstruction (BenchCAD Workflow)
Scenario Evaluation: A robotics hardware team feeds orthographic engineering drawings and tolerance tables as multimodal images into Kimi K3.1 Pro equipped with a Python CadQuery/OpenSCAD execution sandbox.
Standardized Benchmark Prompt:
text
Reconstruct the planetary gearbox housing in CadQuery from the four-view blueprint images, verify mesh watertightness, compute interference checks, and export STEP geometry.Empirical Output Summary: Kimi K3.1 Pro extracted chamfer radii, bolt-circle pitch diameters, and bearing seat tolerances from the raster drawings, rendered viewport projections to self-correct one boss offset, and exported watertight STEP geometry.
Evaluation Verdict: Demonstrated why Moonshot AI's visual-tool feedback loop reaches 92.7% on BenchCAD across multi-part mechanical assemblies.
820K-Token Cross-Repository Compliance & Automation Pipeline
Scenario Evaluation: Inside Kimi Work desktop v3.1.0, an enterprise architecture lead loads 820,000 tokens spanning OpenAPI contracts, Terraform modules, and SOC 2 audit logs to automate drift remediation.
Standardized Benchmark Prompt:
text
Cross-reference all 64 microservice IAM roles against least-privilege policy baselines, generate scoped Terraform patches, and open staged pull requests with rollback runbooks.Empirical Output Summary: Leveraging $0.20/M cached context reads across 19 iterative verification turns, Kimi K3.1 Pro pinpointed 11 over-permissive wildcard policies and emitted a single 54,000-token remediation bundle within its 128K output cap.
Evaluation Verdict: Total API spend stayed under $2.40 for the entire 19-turn audit, aligning closely with Artificial Analysis's $2.11 benchmark task cost.
Evidence Ledger: Fact-Check & Verification Audit
To ensure search engine E-E-A-T integrity, claims are classified across confirmed, reported, unverified, and unknown tiers:
| Claim / Rumor | Evidence Level | Verification Notes & Findings | Sourced IDs |
|---|---|---|---|
| October 23, 2026 General Availability & November 4 Open-Weight Commitment | CONFIRMED | Moonshot AI released Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed via public API on October 23, 2026, confirming full open-weight publication by November 4, 2026. | src-kimi-k31-rel, src-kimi-k31-hf |
| 2.8T Total / 104B Active MoE & 1M / 128K Token Envelope | CONFIRMED | Official technical documentation verifies that Kimi K3.1 Pro operates a 2.8-trillion-parameter sparse MoE (104B active per token) with 1,000,000 context tokens and 128,000 max completion tokens. | src-kimi-k31-rel, src-kimi-k31-docs |
| AA Index 55 (#4 Global), 67.2% Terminal-Bench 4.0 & 92.7% BenchCAD | CONFIRMED | Independent Artificial Analysis and benchmark consortium records confirm AA Intelligence Index 55 (Rank #4, $2.11 task cost), Terminal-Bench 4.0 at 65.7%–67.2%, AutomationBench-AA 73%, HLE 52%, SciCode 61%, and BenchCAD 92.7%. | src-aa-k31, src-tbench-4 |
| Standard ($2/$10) and Highspeed ($4/$20) API Rate Cards | CONFIRMED | Commercial pricing schedules verify kimi-k3.1-pro at $2.00/M input, $10.00/M output, and $0.20/M cached input, alongside kimi-k3.1-pro-highspeed (2x–3x faster) at $4.00/M input, $20.00/M output, and $0.20/M cached input. | src-kimi-k31-docs |
Production Caveats & Known Constraints
Multi-Node Cluster Footprint Required for Self-Hosting 2.8T Weights
Even under FP8 or INT4 quantization, hosting the 2.8-trillion-parameter graph of Kimi K3.1 Pro after the November 4 open-weight drop demands multi-node H200 or B200 NVLink/RDMA clusters beyond single-box budgets.
2x Uncached & Output Cost Premium on Highspeed Tier
Switching interactive agents from kimi-k3.1-pro to kimi-k3.1-pro-highspeed doubles uncached input ($4.00/M) and output ($20.00/M) charges, requiring strict prefix caching discipline to preserve $2.11-class task economics.
Text-Only Generation Without Native Audio or Video Modalities
While Kimi K3.1 Pro excels at text and high-resolution image inputs, it does not natively ingest raw audio/video streams or emit raster images, relying on external tool harnesses for media rendering.
Recommended Developer Action Plan
Route Interactive IDE Loops to Highspeed and Batch CI to Standard
Bind foreground Kimi Code sessions to kimi-k3.1-pro-highspeed for 2x–3x snappier edits while routing background PR reviews and CI suites to standard kimi-k3.1-pro at $2/$10 per million tokens.
Upgrade Kimi Work Desktop to v3.1.0+ and Structure Prefix Caches
Install Kimi Work v3.1.0+ and pin static repository trees and tool schemas at prompt prefixes so multi-turn Kimi K3.1 Pro agents capture the $0.20/M cached input rate.
Provision Distributed vLLM/SGLang Capacity Ahead of November 4 Weights
Infrastructure teams planning private deployment should size multi-node FP8 memory pools (~3.0 TB aggregate HBM including KV cache) ahead of the November 4, 2026 open-weight release.
Frequently Asked Questions
What is Kimi K3.1 Pro and when did Moonshot AI release it?
Kimi K3.1 Pro is Moonshot AI's 2.8-trillion-parameter sparse Mixture-of-Experts (104B active) multimodal frontier model, released in general availability on October 23, 2026 alongside a Highspeed API variant.
When will the open weights for Kimi K3.1 Pro become available?
Moonshot AI scheduled the full open-weight release of Kimi K3.1 Pro by November 4, 2026, making it the highest-scoring open-weight-bound model on the Artificial Analysis Intelligence Index (score 55).
How does Kimi K3.1 Pro perform on Terminal-Bench 4.0 compared to competitors?
Kimi K3.1 Pro achieves 65.7% to 67.2% on Terminal-Bench 4.0—leaping from Kimi K3's 13%, outperforming GPT-6 Astra (58.2%), and matching Claude Opus 5.5 (66.4%).
What is the difference between kimi-k3.1-pro and kimi-k3.1-pro-highspeed?
Both endpoints serve the exact same Kimi K3.1 Pro intelligence and 1M context window, but kimi-k3.1-pro-highspeed delivers 2x–3x faster token generation at $4.00/$20.00 per M tokens versus $2.00/$10.00 on standard, with identical $0.20/M cached input pricing.
What context window and maximum output length does Kimi K3.1 Pro support?
Kimi K3.1 Pro supports a native 1,000,000-token input context window across text and image modalities and generates up to 128,000 completion tokens in a single response.
How does Kimi K3.1 Pro score on HLE, SciCode, AutomationBench-AA, and BenchCAD?
Kimi K3.1 Pro records 52% on Humanity's Last Exam (HLE), 61% on SciCode, 73% on AutomationBench-AA (up from 58%), and 92.7% on BenchCAD 3D CAD reconstruction with tools.
Which first-party developer tools integrate Kimi K3.1 Pro natively?
Moonshot AI ships Kimi K3.1 Pro directly inside the Kimi Work desktop application (v3.1.0 and newer) for enterprise automation and Kimi Code for autonomous repository and terminal engineering.
What hardware is needed to self-host Kimi K3.1 Pro once weights ship?
Because Kimi K3.1 Pro contains 2.8T total parameters (104B active), self-hosting uncompromised FP8 weights requires a multi-node GPU cluster (~3 TB aggregate VRAM across 3x–4x 8-GPU H200 or B200 nodes), while INT4 quantization lowers memory to ~1.5 TB.
Verified Sources & References
- [Moonshot AI Official Blog] Introducing Kimi K3.1 Pro & Kimi K3.1 Pro Highspeed: 2.8T Sparse MoE for Autonomous Engineering (ID:
src-kimi-k31-rel) - [Moonshot AI Developer Platform] Kimi K3.1 Pro API Reference, 1M Context Window & Prompt Caching Rate Cards (ID:
src-kimi-k31-docs) - [Moonshot AI on Hugging Face] moonshotai/Kimi-K3.1-Pro Architecture Specification & November 4 Open-Weight Schedule (ID:
src-kimi-k31-hf) - [Artificial Analysis] Kimi K3.1 Pro Intelligence Index (55, Rank #4), AutomationBench-AA & Task Cost Evaluation (ID:
src-aa-k31) - [Terminal-Bench & BenchCAD Evaluation Consortium] Terminal-Bench 4.0, SciCode & BenchCAD 3D Spatial Tooling Leaderboard (October 2026) (ID:
src-tbench-4) - [Moonshot AI Product Engineering] Kimi Work v3.1.0 Desktop & Kimi Code Agentic Harness Integration Notes (ID:
src-kimi-work)
Recommended Field Notes
Muse Spark Open: Meta 816B MoE Architecture, NVFP4 & Serving Guide
Meta's Muse Spark Open launched Oct 15, 2026: 816B/42B sparse MoE (62 layers, 8/256 experts), native omni-modal 1M context, BF16 & NVFP4 weights on Hugging Face.
MiniMax M3.1 Flash: 428B Open MoE, Video Input & Benchmarks
MiniMax M3.1 Flash launched Oct 19, 2026: 428B/23B open-weight MoE, native video input, 1M context, 73.7% DeepSWE v1.1, MXFP8/NVFP4, and $0.15/M API pricing.