One-line summary
Among three open-source MoE flagships, Kimi K3.1 Pro has the largest total parameters (about 2.8T, 104B active) and the highest intelligence score (55); DeepSeek V4.1 Flash has the lowest activation cost (8B input / 16B output) and the lowest off-peak price (¥1 per million input tokens); MiniMax M3.1 Flash has the smallest total parameters (428B) but already released weights, with official claims of about 3x speed and 70% lower cost versus its predecessor.
Model Introductions
- Kimi K3.1 Pro (Moonshot AI): launched early morning October 24, 2026; MoE with about 2.8T total parameters (CCTV Finance: "3-trillion-class") and 104B active; 1M-token context; 55 on the Artificial Analysis Intelligence Index (4th overall); open weights expected in November 2026.
- DeepSeek V4.1 Flash (DeepSeek): released September 10, 2026; 552B-parameter MoE with an asymmetric Causal-Encoder-Decoder design (8B input / 16B output activation); on October 24, 2026 it topped Artificial Analysis's Cyber Index at 57%, first among publicly available models.
- MiniMax M3.1 Flash (MiniMax): first launched October 19, 2026; weights officially open-sourced October 23, 2026; MoE with full-model sparse attention, 428B total / 23B active; official claims of about 3x inference speed and 70% lower usage cost versus the previous M3.
Core Parameter Comparison
| Dimension | Kimi K3.1 Pro | DeepSeek V4.1 Flash | MiniMax M3.1 Flash |
|---|---|---|---|
| Vendor | Moonshot AI | DeepSeek | MiniMax |
| Release date | 2026-10-24 | 2026-09-10 | 2026-10-19 (weights 10-23) |
| Architecture | MoE | MoE (asymmetric Causal-Encoder-Decoder) | MoE (full-model sparse attention) |
| Total parameters | About 2.8T | 552B | 428B |
| Active parameters | 104B | 8B input / 16B output | 23B |
| Context window | 1M tokens | Not disclosed in official announcement (V4 series claims million-level) | 1M (MiniMax platform overview) |
| Intelligence benchmark | AA Intelligence Index 55 (4th) | AA Cyber Index 57% (1st among public models) | Not disclosed by vendor |
| Coding benchmark | Terminal-Bench 4.0 67% (AA measured) | Not disclosed in announcement | Not disclosed |
| API pricing (per M tokens) | $2 input / $10 output; $0.20 cache hit | Off-peak ¥1 input / ¥4 output (cache hit ¥0.02); peak ¥2 / ¥8 | ¥1 input / ¥4 output; ¥0.1 cache read (as of 2026-10-19) |
| Weight status | Open weights in November 2026 | Deployable locally (vLLM/SGLang); V4 family open-sourced | Weights on Hugging Face since 2026-10-23 |
| Key strengths | Long-horizon coding, 3D creation, end-to-end finance | Low activation cost, cyber-defense benchmark top | ~3x speed and 70% lower cost (official) |
Key Differences
- Total vs active parameters: Kimi K3.1 Pro sets the open-source record at 2.8T total with 104B active; DeepSeek V4.1 Flash compresses input activation to 8B via its asymmetric design; MiniMax M3.1 Flash is smallest (428B total, 23B active). Active parameters drive per-inference compute cost; total parameters drive knowledge capacity.
- Pricing tiers: At official rates, DeepSeek V4.1 Flash's off-peak ¥1/M input is the lowest; MiniMax M3.1 Flash matches at ¥1/M input; Kimi K3.1 Pro lists $2/M input (about one-third below its K3) targeting a higher intelligence tier.
- Benchmark focus: Kimi K3.1 Pro leads on general intelligence (55, 4th); DeepSeek V4.1 Flash tops the Cyber Index among public models (57%); MiniMax M3.1 Flash emphasizes speed and cost (official: about 3x speed, 70% lower cost).
- Open-source cadence: MiniMax M3.1 Flash weights are out; DeepSeek V4.1 Flash is locally deployable; Kimi K3.1 Pro weights come in November 2026 (not published at launch; the open release stripped cyber-related training data).
Buying Advice
- Highest general intelligence and largest parameters: choose Kimi K3.1 Pro (AA Intelligence Index 55), accepting weights arriving in November.
- Lowest inference cost and mature local deployment: choose DeepSeek V4.1 Flash (8B input activation, off-peak ¥1/M) or MiniMax M3.1 Flash (weights open, about 3x speed).
- Cyber-defense / security agent use cases: DeepSeek-V4.1-Flash-1017 leads publicly available models on the Cyber Index (57%), a useful reference for security-agent selection.
- Pricing freshness: figures reflect official announcements (DeepSeek effective 2026-09-10, MiniMax 2026-10-19, Kimi 2026-10-24); always check the vendor API docs for current prices before integration.
Data Sources and References
- CCTV Finance: Moonshot AI releases Kimi K3.1 Pro, a "3-trillion-class" open-source model (2026-10-24) link
- AIToolsRecap: Kimi K3.1 Pro details (2026-10-24) link
- DeepSeek official: Introducing DeepSeek-V4.1-Flash (2026-09-10) link
- DeepSeek API Docs: Models & Pricing link
- BAAI Hub: DeepSeek-V4.1-Flash first among public models on Cyber Index at 57% (2026-10-24) link
- 36Kr: MiniMax open-sources M3.1 Flash; 100+ trillion tokens in 3 days (2026-10-23) link
- 36Kr: MiniMax M3.1 Flash launch: 200+ TPS output, ¥1/M input (2026-10-20) link
- MiniMax Open Platform: model overview (M3.1-Flash, native multimodal, 1M context) link