Perfect score on math competition. No tools enabled.

Perfect score on AIME 2025. Matches GPT-5.2 ceiling — AIME may be saturated for frontier models.

AIME 2025 pass@1 with Python interpreter access. 100% consensus@8. Retired Feb 2026. Tool-assisted score — not directly comparable to models without tool access.

Up from GLM-5.1 (95.3). Near-perfect math competition score.

AIME 2025 with tool use. Edges out gpt-oss-120b (97.9%). Remarkable for a 3.6B active param model.

Near-perfect AIME score. ByteDance reports gold medals on major math olympiads.

AIME 2024+2025 combined (60 questions). New AIME SOTA.

AIME 2025 with tool use. Also scores 96.6% on AIME 2024 (tools). Apache 2.0 open-weight.

Open-weight MoE (196B total, 11B active). Reaches 99.9 with PaCoRe parallel thinking. Best open model on AIME.

Near-frontier math; trails GPT-5.6 Sol (99.9) but matches DeepSeek V4 Pro. Open-weights 975B MoE.

Strong but below GLM-5.2 (99.2) and GPT-5.5 (98.3).

MathArena AIME 2026. 95% without tools, 100% with code execution. Top math performance.

AIME 2026. Slightly up from K2.5 (96.1%). Close to Opus 4.6 (96.7%) but below GPT-5.4 (99.2%).

AIME 2025. Strong math reasoning for a 13B-active MoE model.

1T MoE, 32B active. Open-source. Averaged over 32 runs on AIME 2025.

AIME 2025. Deep reasoning variant. SOTA-level math, surpasses GPT-5 (94.6).

AIME 2026 competition. MathArena independent evaluation, 3rd place behind Step-3.5-Flash (96.67).

Open-weight 400B model. Highest AIME score among open models, competitive with GPT-5.4.

Lower than GLM-5.2 (99.2) and GPT-5.5 (98.3).

Highest AIME score among Qwen models. Strong math reasoning, competitive with Gemini 3 Pro (96.7) and Kimi K2.5 (96.1).

AIME 2026 competition. MathArena independent evaluation, 4th place.

AIME 2026. Strong math reasoning at 12B active params.

Remarkable math for 13B activated params. Near Pro-level.

AIME 2025 without tools. Strong but not perfect — Claude Opus 4.6 and GPT-5.2 both hit 100%.

Multi-model orchestration for verifier-free evolutionary inference. RSA baselines show GPT-5 mini at 85.0% GPQA Diamond and 94.2% AIME. Qwen3-235B at 84.3% GPQA Diamond.

AIME 2026. 685B MoE params with DeepSeek Sparse Attention. Top-tier math reasoning.

15B active MoE. AIME 2025 score, near frontier level.

Beats Gemma 4 31B by 5pp on AIME 2026.

Advanced math reasoning. Anthropic official benchmarks page.

Joint highest AIME 2024 at time of release. Third-party verified.

AIME 2025, Pass@1. Open-weight, 685B MoE. Matches GPT-5 on reasoning.

Mid-tier model in Seed 2.0 family. Handles 95% of enterprise workloads at half the Pro cost.

AIME 2026 I. On par with Kimi K2.5 (92.5) and DeepSeek V3.2 (92.7).

Without tools (closed-book). With tools achieves 99.5%. Best-performing small reasoning model on AIME.

AIME 2026 score. Matches GLM-5 and beats Gemma 4 31B (89.2) as an open MoE.

AIME 2026 variant. Competitive with Qwen3.5 27B (90.8) despite being a VLM. Dense 33B params.

MathArena AIME 2026. A 9B model matching frontier math scores. Remarkable reasoning-per-parameter.

Second open-weight LLM to achieve IMO/IOI/ICPC Gold; 30B MoE with only 3B active params matching frontier models on math reasoning.

30B-A3B MoE with only 3.6B active params. Strong math reasoning for a lightweight model.

AIME 2026. Strong math reasoning from open-weight MoE model.

Ternary variant, AIME25. Conventional IQ2_XXS collapses to 66.67 here; ternary Bonsai retains 90.84 at half the footprint.

AIME 2026. Exceptional math reasoning for a 27B model — near frontier-class.

Multi-model orchestration for verifier-free evolutionary inference. RSA baselines show GPT-5 mini at 85.0% GPQA Diamond and 94.2% AIME. Qwen3-235B at 84.3% GPQA Diamond.

Independent evaluation of 4 frontier LLMs across 6 reasoning paradigms and 10 benchmarks. No single paradigm dominates; learned router recovers 37% of oracle gap.

Multi-model orchestration for verifier-free evolutionary inference. RSA baselines show GPT-5 mini at 85.0% GPQA Diamond and 94.2% AIME. Qwen3-235B at 84.3% GPQA Diamond.

Independent user corroborates Google's AIME score; 0 engagement

Multi-model orchestration for verifier-free evolutionary inference. RSA baselines show GPT-5 mini at 85.0% GPQA Diamond and 94.2% AIME. Qwen3-235B at 84.3% GPQA Diamond.

Without tools. Hybrid Mamba-Transformer MoE, 3.3x throughput vs Qwen3-30B on H200.

1-bit variant, AIME25. Binary weights retain 88.75 vs 93.29 FP16 — reasoning is the most compression-robust capability, per PrismML.

MoE with only 3.8B active params outperforms many larger models.

India's first domestically-trained 105B model. MoE with 10.3B active params. Apache 2.0.

Ternary variant, AIME26. The benchmark where conventional sub-4-bit breaks hardest (IQ2_XXS: 57.50). Bonsai holds 87.5 at 1.71 bits/weight.

1-bit variant, AIME26. Both Bonsai variants hold AIME above 87 at sub-2-bit; the exact capability conventional sub-4-bit loses.

Competitive math reasoning. Behind Claude Opus 4.5 (91.0) but ahead of Sonnet 4.5 (88.0).

Selective latent reasoning that activates continuous reasoning only at uncertain tokens via entropy threshold, achieving higher accuracy than standard CoT with fewer tokens. Qwen3-8B + SeLaR matches Q

Selective latent reasoning that activates continuous reasoning only at uncertain tokens via entropy threshold, achieving higher accuracy than standard CoT with fewer tokens. Qwen3-8B + SeLaR matches Q

4B model beating 32B and 253B models on AIME 2025. Trained on 448K samples with Distribution-Aligned Sequence Distillation.

Multi-model orchestration for verifier-free evolutionary inference. RSA baselines show GPT-5 mini at 85.0% GPQA Diamond and 94.2% AIME. Qwen3-235B at 84.3% GPQA Diamond.

SFT+RL variant of Phi-4-reasoning. 14B model hitting 81%+ on AIME 2024 — remarkable for size.

Strong math reasoning for a Haiku-tier model. Extended thinking with 128K budget.

15B open-weight model. RL post-training achieves strong accuracy with 30-50% shorter reasoning traces. Trained on 5 domains with adaptive sampling.

Multi-model orchestration for verifier-free evolutionary inference. RSA baselines show GPT-5 mini at 85.0% GPQA Diamond and 94.2% AIME. Qwen3-235B at 84.3% GPQA Diamond.

AIME 2026 no tools. Beats Gemma 3 27B (20.8) decisively.

Domain Steering MoE (DSMoE): training-free framework that activates domain-specific experts in sparse MoE LLMs at zero additional inference cost. Key finding: GPT-OSS-20B (3.6B active params) with DSM

SliderQuant preserves accuracy in 4-bit quantized LLMs. DeepSeek-R1-Distill-Qwen-14B at W4A16 retains 94.6% MATH and 91.35% GSM8K vs full-precision. Even W2A16 keeps 29.4% MATH vs 0% for OmniQuant. Qw

SliderQuant preserves accuracy in 4-bit quantized LLMs. DeepSeek-R1-Distill-Qwen-14B at W4A16 retains 94.6% MATH and 91.35% GSM8K vs full-precision. Even W2A16 keeps 29.4% MATH vs 0% for OmniQuant. Qw

14B dense model. Matches DeepSeek-R1 (671B) on AIME despite 47x fewer params.

SliderQuant preserves accuracy in 4-bit quantized LLMs. DeepSeek-R1-Distill-Qwen-14B at W4A16 retains 94.6% MATH and 91.35% GSM8K vs full-precision. Even W2A16 keeps 29.4% MATH vs 0% for OmniQuant. Qw

New 48B MoE model (2.7B active) trained on 20T tokens with novel FiberPO RL algorithm. Open-sourced on HuggingFace. Competitive with much larger models on SWE-bench Verified (62.6%), LiveCodeBench v6

Domain Steering MoE (DSMoE): training-free framework that activates domain-specific experts in sparse MoE LLMs at zero additional inference cost. Key finding: GPT-OSS-20B (3.6B active params) with DSM

SliderQuant preserves accuracy in 4-bit quantized LLMs. DeepSeek-R1-Distill-Qwen-14B at W4A16 retains 94.6% MATH and 91.35% GSM8K vs full-precision. Even W2A16 keeps 29.4% MATH vs 0% for OmniQuant. Qw

MCTS with metacognitive reflection improves reasoning. On Qwen3-30B-A3B, PRISM-MCTS achieves 65.15% on GPQA Diamond and 93.28% on MATH500.

AIME24 baseline; ETR boosts to 73.3. Strong AIME24 performance for 8B dense model.

Two-phase GRPO framework: first solve, then refine. Qwen3-4B ThinkTwice achieves 60.43% AIME (vs 29.18% base) and 95.70% MATH500 after self-refinement, beating DAPO and DrGRPO.

Independent evaluation of 4 frontier LLMs across 6 reasoning paradigms and 10 benchmarks. No single paradigm dominates; learned router recovers 37% of oracle gap.

Domain Steering MoE (DSMoE): training-free framework that activates domain-specific experts in sparse MoE LLMs at zero additional inference cost. Key finding: GPT-OSS-20B (3.6B active params) with DSM

MCTS with metacognitive reflection improves reasoning. On Qwen3-30B-A3B, PRISM-MCTS achieves 65.15% on GPQA Diamond and 93.28% on MATH500.

MCTS with metacognitive reflection improves reasoning. On Qwen3-30B-A3B, PRISM-MCTS achieves 65.15% on GPQA Diamond and 93.28% on MATH500.

Selective latent reasoning that activates continuous reasoning only at uncertain tokens via entropy threshold, achieving higher accuracy than standard CoT with fewer tokens. Qwen3-8B + SeLaR matches Q

AIME24 baseline; ETR boosts to 73.3. Compact 4B model with strong math reasoning via efficient CoT optimization.

Test-time variational synthesis improves RLVR for small models without labeled data. Qwen3-4B + TTVS achieves 90.3% MATH500 and 48.9% GPQA, outperforming baseline by +26.3 and +21.7 points respectivel

MCTS with metacognitive reflection improves reasoning. On Qwen3-30B-A3B, PRISM-MCTS achieves 65.15% on GPQA Diamond and 93.28% on MATH500.

Evolutionary model merging produces Pareto-optimal reasoning models that reduce output length by 50%+ while matching or improving accuracy on math benchmarks.

Selective latent reasoning that activates continuous reasoning only at uncertain tokens via entropy threshold, achieving higher accuracy than standard CoT with fewer tokens. Qwen3-8B + SeLaR matches Q

Label-free RL via reasoning structure rewards on Qwen3-4B matches or beats ground-truth RL: AIME25 45.83% vs 46.67% oracle, MATH500 93.30%, and Minerva 61.99%.

Two-phase GRPO framework: first solve, then refine. Qwen3-4B ThinkTwice achieves 60.43% AIME (vs 29.18% base) and 95.70% MATH500 after self-refinement, beating DAPO and DrGRPO.

Evolutionary model merging produces Pareto-optimal reasoning models that reduce output length by 50%+ while matching or improving accuracy on math benchmarks.

Impressive for a 4B model running on T4 GPUs. MoE architecture.

Independent evaluation of 4 frontier LLMs across 6 reasoning paradigms and 10 benchmarks. No single paradigm dominates; learned router recovers 37% of oracle gap.

Two-phase GRPO framework: first solve, then refine. Qwen3-4B ThinkTwice achieves 60.43% AIME (vs 29.18% base) and 95.70% MATH500 after self-refinement, beating DAPO and DrGRPO.

Matches Gemma 3 27B despite being a fraction of the size. Built-in reasoning capability.

Evolutionary model merging produces Pareto-optimal reasoning models that reduce output length by 50%+ while matching or improving accuracy on math benchmarks.

Test-time variational synthesis improves RLVR for small models without labeled data. Qwen3-4B + TTVS achieves 90.3% MATH500 and 48.9% GPQA, outperforming baseline by +26.3 and +21.7 points respectivel

Evolutionary model merging produces Pareto-optimal reasoning models that reduce output length by 50%+ while matching or improving accuracy on math benchmarks.

Evolutionary model merging produces Pareto-optimal reasoning models that reduce output length by 50%+ while matching or improving accuracy on math benchmarks.

Two-phase GRPO framework: first solve, then refine. Qwen3-4B ThinkTwice achieves 60.43% AIME (vs 29.18% base) and 95.70% MATH500 after self-refinement, beating DAPO and DrGRPO.

Label-free RL via reasoning structure rewards on Qwen3-4B matches or beats ground-truth RL: AIME25 45.83% vs 46.67% oracle, MATH500 93.30%, and Minerva 61.99%.

Self-curriculum test-time RL method that improves small models substantially. Qwen3-4B-Base with DiSCTT reaches 81.3% MMLU and 75.2% MATH-500, surpassing Qwen2.5-7B-Instruct base (76.2% MMLU, 58.8% MA

Two-phase GRPO framework: first solve, then refine. Qwen3-4B ThinkTwice achieves 60.43% AIME (vs 29.18% base) and 95.70% MATH500 after self-refinement, beating DAPO and DrGRPO.

Evolutionary model merging produces Pareto-optimal reasoning models that reduce output length by 50%+ while matching or improving accuracy on math benchmarks.

Self-curriculum test-time RL method that improves small models substantially. Qwen3-4B-Base with DiSCTT reaches 81.3% MMLU and 75.2% MATH-500, surpassing Qwen2.5-7B-Instruct base (76.2% MMLU, 58.8% MA

Stable-OPD stabilizes on-policy distillation for small math models, boosting Qwen2.5-Math-1.5B from 28.9% to 36.1% avg accuracy and Qwen2.5-Math-7B to 47.6%, surpassing GRPO and SFT baselines on math

Stable-OPD stabilizes on-policy distillation for small math models, boosting Qwen2.5-Math-1.5B from 28.9% to 36.1% avg accuracy and Qwen2.5-Math-7B to 47.6%, surpassing GRPO and SFT baselines on math

Self-curriculum test-time RL method that improves small models substantially. Qwen3-4B-Base with DiSCTT reaches 81.3% MMLU and 75.2% MATH-500, surpassing Qwen2.5-7B-Instruct base (76.2% MMLU, 58.8% MA

Test-time variational synthesis improves RLVR for small models without labeled data. Qwen3-4B + TTVS achieves 90.3% MATH500 and 48.9% GPQA, outperforming baseline by +26.3 and +21.7 points respectivel

Stable-OPD stabilizes on-policy distillation for small math models, boosting Qwen2.5-Math-1.5B from 28.9% to 36.1% avg accuracy and Qwen2.5-Math-7B to 47.6%, surpassing GRPO and SFT baselines on math

Test-time variational synthesis improves RLVR for small models without labeled data. Qwen3-4B + TTVS achieves 90.3% MATH500 and 48.9% GPQA, outperforming baseline by +26.3 and +21.7 points respectivel

Self-curriculum test-time RL method that improves small models substantially. Qwen3-4B-Base with DiSCTT reaches 81.3% MMLU and 75.2% MATH-500, surpassing Qwen2.5-7B-Instruct base (76.2% MMLU, 58.8% MA

Stable-OPD stabilizes on-policy distillation for small math models, boosting Qwen2.5-Math-1.5B from 28.9% to 36.1% avg accuracy and Qwen2.5-Math-7B to 47.6%, surpassing GRPO and SFT baselines on math

Stable-OPD stabilizes on-policy distillation for small math models, boosting Qwen2.5-Math-1.5B from 28.9% to 36.1% avg accuracy and Qwen2.5-Math-7B to 47.6%, surpassing GRPO and SFT baselines on math

Stable-OPD stabilizes on-policy distillation for small math models, boosting Qwen2.5-Math-1.5B from 28.9% to 36.1% avg accuracy and Qwen2.5-Math-7B to 47.6%, surpassing GRPO and SFT baselines on math

SliderQuant preserves accuracy in 4-bit quantized LLMs. DeepSeek-R1-Distill-Qwen-14B at W4A16 retains 94.6% MATH and 91.35% GSM8K vs full-precision. Even W2A16 keeps 29.4% MATH vs 0% for OmniQuant. Qw

Independent evaluation of 4 frontier LLMs across 6 reasoning paradigms and 10 benchmarks. No single paradigm dominates; learned router recovers 37% of oracle gap.

TTC-Net adds test-time compute layers to Llama-3-Instruct-7B via hardware-efficient optimal control. Acc@8 scores: MATH-500 improves from 25.0% to 52.8%, AIME25 from 0% to 5.0%. The TTC layer runs in

Self-curriculum test-time RL method that improves small models substantially. Qwen3-4B-Base with DiSCTT reaches 81.3% MMLU and 75.2% MATH-500, surpassing Qwen2.5-7B-Instruct base (76.2% MMLU, 58.8% MA

TTC-Net adds test-time compute layers to Llama-3-Instruct-7B via hardware-efficient optimal control. Acc@8 scores: MATH-500 improves from 25.0% to 52.8%, AIME25 from 0% to 5.0%. The TTC layer runs in

Evolutionary model merging produces Pareto-optimal reasoning models that reduce output length by 50%+ while matching or improving accuracy on math benchmarks.