Reasoning / Math Models
Data as of June 18, 2026| Model | FrMathA | AIMEB | MATH L5C | Notes▼ | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1.□ | 2025/12 | ? | 128K | $2/$14 | 73 | 130.34 | 📝👁️🔊 | ○ | 240.7 | 296.1 | 198.1 | 1.7 | - | |
| 2.□ | 2026/4 | 1T (32B active, MoE) | 256K | - | - | - | 📝👁️ | ● | - | 196.4 | - | 2.3 | - | |
| 3.□ | 2025/4 | ? | 128K | - | 161 | 21.94 | 📝 | ○ | - | - | 297.8 | 3 | - | |
| 4.□ | 2026/3 | ? | 128K | - | - | - | 📝👁️🔊 | ○ | 150.0 | - | - | 3.3 | - | |
| 5.□ | 2025/4 | ? | 200K | - | 118 | 5.38 | 📝 | ○ | - | - | 397.8 | 3.3 | - | |
| 6.□ | 2025/7 | 754B MoE | 200K | - | - | - | 📝 | ● | - | 495.3 | - | 3.3 | Open-weight MoE; basis for Big Pickle; strong reasoning + tool use | |
| 7.□ | 2026/5 | 74B (4B active, MoE) | ? | - | - | - | 📝 | ● | - | - | - | 3.7 | Pre-RL base checkpoint; trained on AMD hardware. Marks Zyphra's move beyond small-MoE experimentation. Companion ZAYA1-VL-8B (700M active) released alongside. | |
| 8.□ | 2026/4 | 120B (12B active, MoE) | 1M | - | - | - | 📝 | ● | - | - | - | 3.7 | Hybrid Mamba-Attention MoE w/ LatentMoE; trained on 25T tokens; 2.2x throughput vs GPT-OSS-120B, 7.5x vs Qwen3.5-122B; native MTP speculative decoding | |
| 9.□ | 2026/4 | 295B (21B active, MoE) | 256K | - | - | - | 📝 | ● | - | - | - | 3.7 | Open-weight Hunyuan 3 preview; 192 experts, top-8 routing; strong STEM/reasoning release | |
| 10.□ | 2025/9 | ? | 200K | $3/$15 | - | - | 📝👁️ | ○ | - | - | 497.7 | 3.7 | - | |
| 11.□ | 2026/2 | ? | 1M | $2/$12 | 109 | 29.71 | 📝👁️🔊 | ○ | 436.9 | 395.6 | - | 3.7 | - | |
| 12.□ | 2025/10 | 754B MoE | 200K | free/free | - | - | 📝 | ○ | - | - | - | 3.7 | Stealth model (GLM 4.6); free on OpenCode Zen; reasoning + tool calling; text-only | |
| 13.□ | 2026/4 | 754B MoE | 200K | $0.95/$3 | 74 | 1.64 | 📝👁️ | ● | - | 595.3 | - | 3.7 | - | |
| 14.□ | 2026/5 | ? | 200K | $15/$75 | - | - | 📝👁️ | ○ | - | - | - | 3.7 | Incremental over 4.7; high CoT faithfulness; no steganographic reasoning found in white-box SAE analysis | |
| 15.□ | 2026/4 | ? | 200K | $15/$75 | - | - | 📝👁️ | ○ | - | - | - | 3.7 | - | |
| 16.□ | 2026/4 | ? | 400K | - | - | - | 📝👁️🔊 | ○ | - | - | - | 3.7 | - | |
| 17.□ | 2026/4 | 1.6T (49B active, MoE) | 1M | $2/$3 | - | - | 📝 | ● | - | - | - | 3.7 | NIST/CAISI eval (May 2026): most capable PRC model tested, ~8mo behind US frontier; IRT-Elo 800 vs GPT-5.5 1260, Opus 4.6 999. CAISI scores: GPQA-Diamond 90%, OTIS-AIME-2025 97%, SWE-Bench Verified 74%, ARC-AGI-2 46%. Pricing is DeepSeek V4 Pro (developer-reported). | |
| 18.□ | 2026/4 | 48B (3B active, MoE) | 1M | - | - | - | 📝 | ● | - | - | - | 3.7 | Novel architecture: 3:1 KDA-to-MLA ratio (Kimi Delta Attention + Multi-head Latent Attention). 75% smaller KV cache, 6x decoding throughput at 1M context. Paired with FlashKDA CUTLASS kernels. Efficiency-focused, not flagship-capability. | |
| 19.□ | 2026/2 | ? | 200K | - | - | - | 📝👁️ | ○ | - | - | - | 3.7 | - | |
| 20.□ | 2025/10 | ? | 200K | - | - | - | 📝👁️ | ○ | - | - | - | 3.7 | - | |
| 21.□ | 2026/5 | 1T (63B active, MoE) | 262K | - | - | - | 📝 | ● | - | - | - | 3.7 | Trillion-parameter thinking model w/ adaptive reasoning; tuned for coding agents, tool use, long-horizon tasks; high/xhigh reasoning modes | |
| 22.□ | 2026/4 | - | - | - | - | - | 📝 | ○ | - | - | - | 3.7 | xAI reasoning/agent flagship. Scores pending public benchmark publication. | |
| 23.□ | 2026/3 | 230B (10B active, MoE) | 200K | - | - | - | 📝 | ● | - | - | - | 3.7 | MiniMax reasoning model (text-only). Open weights, non-commercial license. Scores: AA Intelligence Index 50. | |
| 24.□ | 2025/12 | - | - | - | - | - | 📝 | ○ | - | - | - | 3.7 | Iterative GPT-5 release. Scores pending. | |
| 25.□ | 2026/3 | - | - | - | - | - | 📝👁️ | ○ | - | - | - | 3.7 | Fast tier of Gemini 3 family. Scores pending. | |
| 26.□ | 2025/3 | - | - | - | - | - | 📝👁️ | ○ | - | - | - | 3.7 | Previous-gen Gemini flagship. Largely superseded by 3.1 Pro. | |
| 27.□ | 2026/1 | - | - | - | - | - | 📝 | ● | - | - | - | 3.7 | Predecessor to K2.6. Scores pending. | |
| 28.□ | 2026/2 | - | - | - | - | - | 📝 | ● | - | - | - | 3.7 | DeepSeek V3.2-Exp release. Scores pending. | |
| 29.□ | 2026/6 | ? | 1M | $0.60/$2 | - | - | 📝👁️ | ● | - | - | - | 3.7 | Open-weight frontier coding + 1M context + native multimodality. SWE-Bench Pro 59.0, Terminal-Bench 2.1 66.0, BrowseComp 83.5. MSA: ~9x faster prefill / 15x decode at 1M. | |
| 30.□ | 2026/5 | ? | 1M | $2/$9 | - | - | 📝👁️🔊 | ○ | - | - | - | 3.7 | ~4x faster than prior frontier; beats Gemini 3.1 Pro on coding/agentic (Terminal-Bench 2.1 76.2, MCP Atlas 83.6, CharXiv 84.2). Dynamic thinking on by default. | |
| 31.□ | 2026/6 | ? | 1M | $0.40/$2 | - | - | 📝👁️🎥 | ○ | - | - | - | 3.7 | AA Intelligence Index 53.3 (coding 46.5). Multimodal (text/image/video in). ScreenSpot Pro 79.0 GUI grounding. | |
| 32.□ | 2026/6 | 550B (55B active, MoE) | 1M | - | 146.3 | - | 📝 | ● | - | - | - | 3.7 | AA Intelligence Index 48; AA-Omniscience 78.7 (top non-hallucination in set). ~20T train tokens, 11 langs + 43 prog langs. | |
| 33.□ | 2025/11 | 27B (Gemma 3) | 128K | - | - | - | 📝👁️ | ● | - | - | - | 3.7 | Southeast Asian multilingual (11 langs incl Malay); Gemma 3 27B base. #4 on SEA-HELM, #1 Tamil/Filipino; runs on a 32GB laptop. | |
| 34.□ | 2024/1 | 5B | 20K | - | - | - | 📝 | ● | - | - | - | 3.7 | Malaysia LLM (Mistral-based); Malay slang/colloquialisms + 16 regional MY languages, ~90B Malay tokens. | |
| 35.□ | 2024/7 | 7B (Qwen2) | 32K | - | - | - | 📝 | ● | - | - | - | 3.7 | Southeast Asian multilingual (Malay/Indonesian/Thai/Vietnamese/…); Qwen2-based, SOTA for its size on SEA tasks. | |
| 36.□ | 2026/5 | ? | 1M | $3/$8 | - | - | 📝👁️ | ○ | - | - | - | 3.7 | Qwen3.7 flagship; agent-centric (long-horizon ~35h). AA Intelligence Index 56.6 (#5, top Chinese at launch); Terminal-Bench 2.0 69.7. | |
| 37.□ | 2026/4 | ? | 128K | - | - | - | 📝👁️ | ○ | - | 695.3 | - | 4 | - | |
| 38.□ | 2025/9 | ? | 128K | - | - | - | 📝👁️ | ○ | - | - | 597.1 | 4 | - | |
| 39.□ | 2025/8 | ? | 128K | - | - | - | 📝👁️🔊 | ○ | - | - | 696.8 | 4.3 | - | |
| 40.□ | 2026/6 | 35B active (SMoE) | 256K | - | - | - | 📝 | ○ | - | 794.5 | - | 4.3 | Microsoft's first in-house reasoning model (Build 2026); no distillation from OpenAI/Anthropic. AIME 2026 94.5; SWE-Bench Pro 53 (~Opus 4.6). Foundry private preview. | |
| 41.□ | 2025/1 | 685B | 128K | $0.55/$2 | 60 | 0.84 | 📝 | ● | - | - | 796.6 | 4.7 | - | |
| 42.□ | 2026/2 | ? | 200K | $5/$25 | 40 | 1.78 | 📝👁️ | ○ | 340.7 | 894.4 | - | 5 | - | |
| 43.□ | 2025/1 | ? | 128K | - | 160 | 7.12 | 📝 | ○ | - | - | 896.5 | 5 | - | |
| 44.□ | - | ? | 128K | - | - | - | 📝 | ○ | - | 994.2 | - | 5 | - | |
| 45.□ | - | 31B | 128K | - | - | - | 📝 | ● | - | 1089.2 | - | 5.3 | - | |
#3.3
Z.ai754B MoE200K📝
AIME 95.3
Open-weight MoE; basis for Big Pickle; strong reasoning + tool use
#3.7
Zyphra74B (4B active, MoE)?📝
Pre-RL base checkpoint; trained on AMD hardware. Marks Zyphra's move beyond small-MoE experimentation. Companion ZAYA1-VL-8B (700M active) released alongside.
#3.7
NVIDIA120B (12B active, MoE)1M📝
Hybrid Mamba-Attention MoE w/ LatentMoE; trained on 25T tokens; 2.2x throughput vs GPT-OSS-120B, 7.5x vs Qwen3.5-122B; native MTP speculative decoding
#3.7
Tencent Hunyuan295B (21B active, MoE)256K📝
Open-weight Hunyuan 3 preview; 192 experts, top-8 routing; strong STEM/reasoning release
#3.7
OpenCode Zen754B MoE200Kfree/free📝
Stealth model (GLM 4.6); free on OpenCode Zen; reasoning + tool calling; text-only
#3.7
Anthropic?200K$15/$75📝👁️
Incremental over 4.7; high CoT faithfulness; no steganographic reasoning found in white-box SAE analysis
#3.7
Anthropic?200K$15/$75📝👁️
#3.7
OpenAI?400K📝👁️🔊
#3.7
DeepSeek1.6T (49B active, MoE)1M$2/$3📝
NIST/CAISI eval (May 2026): most capable PRC model tested, ~8mo behind US frontier; IRT-Elo 800 vs GPT-5.5 1260, Opus 4.6 999. CAISI scores: GPQA-Diamond 90%, OTIS-AIME-2025 97%, SWE-Bench Verified 74%, ARC-AGI-2 46%. Pricing is DeepSeek V4 Pro (developer-reported).
#3.7
Moonshot48B (3B active, MoE)1M📝
Novel architecture: 3:1 KDA-to-MLA ratio (Kimi Delta Attention + Multi-head Latent Attention). 75% smaller KV cache, 6x decoding throughput at 1M context. Paired with FlashKDA CUTLASS kernels. Efficiency-focused, not flagship-capability.
#3.7
Anthropic?200K📝👁️
#3.7
Anthropic?200K📝👁️
#3.7
InclusionAI1T (63B active, MoE)262K📝
Trillion-parameter thinking model w/ adaptive reasoning; tuned for coding agents, tool use, long-horizon tasks; high/xhigh reasoning modes
#3.7
MiniMax230B (10B active, MoE)200K📝
MiniMax reasoning model (text-only). Open weights, non-commercial license. Scores: AA Intelligence Index 50.
#3.7
MiniMax?1M$0.60/$2📝👁️
Open-weight frontier coding + 1M context + native multimodality. SWE-Bench Pro 59.0, Terminal-Bench 2.1 66.0, BrowseComp 83.5. MSA: ~9x faster prefill / 15x decode at 1M.
#3.7
Google?1M$2/$9📝👁️🔊
~4x faster than prior frontier; beats Gemini 3.1 Pro on coding/agentic (Terminal-Bench 2.1 76.2, MCP Atlas 83.6, CharXiv 84.2). Dynamic thinking on by default.
#3.7
Alibaba?1M$0.40/$2📝👁️🎥
AA Intelligence Index 53.3 (coding 46.5). Multimodal (text/image/video in). ScreenSpot Pro 79.0 GUI grounding.
#3.7
NVIDIA550B (55B active, MoE)1M146.3📝
AA Intelligence Index 48; AA-Omniscience 78.7 (top non-hallucination in set). ~20T train tokens, 11 langs + 43 prog langs.
#3.7
AI Singapore27B (Gemma 3)128K📝👁️
Southeast Asian multilingual (11 langs incl Malay); Gemma 3 27B base. #4 on SEA-HELM, #1 Tamil/Filipino; runs on a 32GB laptop.
#3.7
Mesolitica5B20K📝
Malaysia LLM (Mistral-based); Malay slang/colloquialisms + 16 regional MY languages, ~90B Malay tokens.
#3.7
Alibaba DAMO7B (Qwen2)32K📝
Southeast Asian multilingual (Malay/Indonesian/Thai/Vietnamese/…); Qwen2-based, SOTA for its size on SEA tasks.
#3.7
Alibaba?1M$3/$8📝👁️
Qwen3.7 flagship; agent-centric (long-horizon ~35h). AA Intelligence Index 56.6 (#5, top Chinese at launch); Terminal-Bench 2.0 69.7.
#4.3
Microsoft AI35B active (SMoE)256K📝
AIME 94.5
Microsoft's first in-house reasoning model (Build 2026); no distillation from OpenAI/Anthropic. AIME 2026 94.5; SWE-Bench Pro 53 (~Opus 4.6). Foundry private preview.