399B sparse MoE (13B active), Apache 2.0. Impressive score but one user reports 60% failure rate on production constraint-following tasks. Self-reported by Arcee.

Max reasoning. New SOTA, 2pt over Gemini 3.1 Pro (91.7).

Current #1 on LiveCodeBench leaderboard. Ahead of DeepSeek V3.2 Speciale (88.7).

Max reasoning, 13B activated. Near V4 Pro (93.5) SOTA.

Pass@1 with chain-of-thought. Temporary API-only reasoning variant. Gold-level results in IMO, ICPC World Finals, IOI 2025.

Leads LiveCodeBench v6 at time of release. ByteDance frontier model with ICPC/IMO/CMO gold medals.

Strong coding performance from 3B active params; LiveCodeBench v6 (2408-2505). With TIR: 88.4.

Top-4 on LiveCodeBench leaderboard. Strong agentic coding model with 1M context window.

Reaches 88.9 with PaCoRe. Open-weight model competitive with frontier closed models on live coding.

15B active MoE. v6 score. Strong coding performance for size class.

LiveCodeBench v6, 64k token budget. Highest open-source score on this benchmark. Averaged over 5 runs.

LiveCodeBench v6. Highest open-weight LiveCodeBench score, beats most proprietary models.

LiveCodeBench v6. Outperforms GPT-5.2 and Claude Opus 4.5 on 80% of evaluated categories.

Pass@1 with chain-of-thought. Open-weight.

Up 7pp from Opus 4.6 (76.0%). Strong coding competition performance.

Ternary variant. Beats Gemma-4-31B Q2_K_XL (74.40) at one-third the footprint — long-form code generation survives ternary compression while conventional sub-4-bit collapses here.

Strong competitive coding despite being the mid-tier variant.

First open-weight VLM from LG. Beats Qwen3-VL 235B Thinking (70.1) on LiveCodeBench v6. 33B dense model.

Highest among open-weight models. +28% over non-thinking variant via Error-driven Chain-of-Thought reasoning.

Dense 27B model, all params active. Matches GPT-5-mini on SWE-bench. Best coding score in Qwen 3.5 medium lineup.

Leading open-weight score on LiveCodeBench. 120B MoE with 12B active params.

LiveCodeBench v6. Matches Gemma 4 31B (80.0) with 10x fewer active params.

Official Google model card. 31B dense model on LiveCodeBench v6.

From xAI official blog. No-tools evaluation. LiveCodeBench Jan-May 2025.

MoE variant. v6. Dense 31B scores 80.0%.

1-bit variant. Still beats Qwen3.6-27B IQ2_XXS (56.40) at 2.5x smaller footprint — the key long-form code-gen metric where sub-4-bit collapses.

Competitive coding on real-world problems. Behind Gemini 3 Pro Preview (91.7%) and DeepSeek V3.2 Speciale (88.7%).

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.

Standard (non-thinking) mode. Thinking mode reaches 83.3%. Speciale variant hits 88.7%.

Budget-tier model at $0.25/1M input tokens. Competitive coding despite being Google fastest/cheapest model.

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

15B open-weight. Matches Nemotron-Cascade while generating less than half the output tokens (7.4K vs 16.0K).

4B model. v5 score. Competitive with models 8x larger.

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.

v6. Outperforms Qwen3 (66.0%) and GPT-OSS (61.0%) at same size class.

9B params. Competitive coding for a small model, though lags behind larger models (GPT-OSS-120B: 82.7).

48B MoE, 2.7B active params. Surpasses GLM-4.7-Flash-Thinking by 2.4% with 85% fewer tokens. Uses FiberPO RL algorithm.

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.

v6 benchmark. Decent coding for a 3.6B active param model.

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.

LiveCodeBench v6; beats Qwen3-Coder-480B-A35B (44.93) and Qwen3-Next (51.79) despite 3B active params.

Competitive coding ability for 4B params. Open-weight, Apache 2.0.

8B open-source model. pass@5 score. Beats GPT-5.1 (48.03) and Claude Sonnet 4.5 (50.78) on this benchmark.

Competitive coding for a 4B model. Runs on consumer edge hardware.

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.

v6. Effective 2B params. Surpasses Gemma 3 27B on coding despite 10x fewer active parameters.

pass@1 on LiveCodeBench (10/01/2024-02/01/2025). Far behind frontier models like Gemma 4 31B (80%) on coding.

Simple self-distillation (sample + fine-tune) improves Qwen3-30B from 42.4% to 55.3% on LiveCodeBench v6. Works across 4B, 8B, and 30B scales. No verifier, teacher model, or RL needed.

pass@1. Significantly behind frontier coding models.

RLVR code training without ground-truth tests. ZeroCoder+DyB4 on Qwen2.5-Coder-7B achieves 42.2% avg on code generation benchmarks.

RLVR code training without ground-truth tests. ZeroCoder+DyB4 on Qwen2.5-Coder-7B achieves 42.2% avg on code generation benchmarks.

RLVR code training without ground-truth tests. ZeroCoder+DyB4 on Qwen2.5-Coder-7B achieves 42.2% avg on code generation benchmarks.

RLVR code training without ground-truth tests. ZeroCoder+DyB4 on Qwen2.5-Coder-7B achieves 42.2% avg on code generation benchmarks.