New SOTA. 13.1pp above Opus 4.6 (80.8%). Largest single-generation coding jump in Anthropic history.

Up 6.8pp from Opus 4.6. New SOTA on SWE-bench Verified, surpassing Claude Mythos Preview.

Updated official score with prompt modification, averaged over 25 trials. Previous entry was 80.8%.

First model to exceed 80% on SWE-bench Verified. 25 trials avg.

Agentic coding benchmark. Verified from official Gemini page.

Max reasoning. Matches Gemini 3.1 Pro, below Opus 4.7 (87.6).

MoE 230B params, 10B active. Matches Claude Opus 4.6 speed on SWE-Bench eval.

Open-weight model matching Opus 4.6 (80.8%) and Gemini 3.1 Pro (80.6%) on SWE-bench Verified.

Multi-language bug fixing. SWE-bench Pro score: 55.6% (new SOTA).

OpenAI coding-focused model. Also scores 56.8% on SWE-bench Pro (harder multilingual variant).

From TML Inkling blog comparison table. GLM-5.2 official is 80.0.

79.6% on SWE-bench Verified, 1.2 points behind Opus 4.6, surpasses Sonnet 4.5 by 2.4 points

Near Claude Opus 4.6 (80.8%) using Claude Code scaffold. Specialize-then-Unify paradigm with MCLA for MoE RL training.

Max reasoning, 13B activated. Impressive for small model, near V4 Pro (80.6).

Qwen 3.6 Plus scores 78.8% on SWE-bench Verified

Distilled model outperforms Gemini 3 Pro (76.2%) on agentic coding. Less than a quarter the cost of Pro.

Self-improving MoE model activating only 10B params. Scores below M2.5 (80.2%) on this benchmark — may use different eval config or scaffold.

Open-weight 295B MoE (21B active). Official Tencent score, not the 74.4 rumour. Beats DeepSeek V4 Pro (80.6 is higher actually - check). Strong agentic model.

Leading open-model score. Only 3 points behind Opus 4.6. Terminal-Bench 2.0: 56.2%.

Grok bot relayed Zhipu GLM-5.1 score; no direct link to paper or official source

Self-reported via bash-only harness. Competitive with Kimi K2.5 (76.8) but below GLM-5.2 (80.0).

First model from Meta Superintelligence Labs. Competitive with Gemini 3.1 Pro on agentic coding.

Matches Inkling on SWE-bench Verified at a third the active params.

77.2% standard, 78.2% with 1M context, 82.0% with parallel compute. SOTA at time of release.

Independent Vals.ai evaluation. OpenAI dropped SWE-bench Verified in favor of SWE-bench Pro (57.7%).

Best open 27B dense model on SWE-bench. Outperforms Gemma 4 31B by 25pp.

1T MoE open-source. Competitive with Claude Sonnet 4.6 (79.6). Uses internal eval framework with bash/createfile tools.

Competitive with Claude Sonnet 4.5 and DeepSeek V3.2 on coding tasks.

SWE-bench Verified. Beats Step 3.5 Flash (74.4).

Strong open-weight coding performance. 17B active params, MoE architecture.

Below GPT-5.4 (77.2%) and Claude Opus 4.6 (81.4%). Coding was not GPT-5 main strength.

Third-party harness engineering by Fujitsu. Test-time scaling with 8 candidate patches, no fine-tuning. Highest score among <229B param models.

Open-weight MoE with only 11B active params. Competitive with proprietary mid-tier models on coding.

Top tier with Claude/GPT. Fast inference + huge context.

Open-weight, 400B params. Strong coding for an open model but behind GLM-5/5.1 successors.

Near Pro level (76.5) at roughly half the cost.

15B active MoE model. Competitive with models 10-20x larger. Uses Multi-Teacher On-Policy Distillation.

Open-weight MoE with only 3B active params. Beats Gemma 4 31B (52.0) by +21.4pts on same benchmark.

Anthropic first Haiku with extended thinking. Matches Sonnet 4 on coding at 1/3 cost.

Primary score from DeepSeek technical report. Robustness tests across frameworks range 72-74%. Trails Claude Opus 4.6 (80.8%) and GPT-5.3 Codex (80.0%).

Grok bot stated ~70-75% range; midpoint used; no official xAI source linked

Open-weight 27B dense model matching GPT-5 mini. Released Feb 2026.

123B dense transformer, 256k context. Best open-weight SWE-bench score. 7x more cost-efficient than Claude Sonnet.

Solid coding ability. Behind frontier models (Claude Opus 4.6 at 81.4%, Gemini 3.1 Pro at 80.6%).

OpenHands agent score; SWE-Agent gets 70.6. 3B active params competitive with 35B+ models.

SWE-Hero-32B achieves 62.2% on SWE-bench Verified and 44.1% on SWE-bench Lite with execution-based fine-tuning. Open-weight 32B model competitive with MiniMax-M2.5 (80.2% Verified).

From TML Inkling blog. NVIDIA's Nemotron 3 Ultra.

SWE-Agent score. 80B MoE, only 3B active params. Coding specialist matching models 10-20x larger.

32B open-weight model with ECoT reasoning traces and industrial code world model. Top open-weight LiveCodeBench result.

17B MoE with 128 experts. Strong coding performance for open-weight model, competitive with GPT-4o class.

MoE (480B total, 35B active) code model with OpenHands agent scaffold. Without execution: 59.4%.

SWE-Hero-32B achieves 62.2% on SWE-bench Verified and 44.1% on SWE-bench Lite with execution-based fine-tuning. Open-weight 32B model competitive with MiniMax-M2.5 (80.2% Verified).

Cost-efficient reasoning model. Behind o3 (69.1%) but far ahead of o1 (48.9%) and o3-mini (49.3%).

24B open-weight coding model. Runs locally on RTX 4090. Apache 2.0 license.

Reframes codebase as structured AST action space. GPT-5-nano + CodeStruct jumps from 19.6% to 40.4% on SWE-bench Verified (+20.8). GPT-5 + CodeStruct hits 67.2%. Reduces errors by up to 87.8%.

Reframes codebase as structured AST action space. GPT-5-nano + CodeStruct jumps from 19.6% to 40.4% on SWE-bench Verified (+20.8). GPT-5 + CodeStruct hits 67.2%. Reduces errors by up to 87.8%.

Reframes codebase as structured AST action space. GPT-5-nano + CodeStruct jumps from 19.6% to 40.4% on SWE-bench Verified (+20.8). GPT-5 + CodeStruct hits 67.2%. Reduces errors by up to 87.8%.

Evaluated with mini-swe-agent-v2. Behind closed-source leaders but competitive for open-weight.

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

OpenAI's first open-weight model at 116.8B params (5.1B active). Apache 2.0 license. High reasoning effort.

32B open-source SWE agent. Top open-source model on SWE-bench Verified (excludes Git hacking results).

Open-weight 32B model distilled from Qwen3-Coder-480B. Two-stage SFT: execution-free then execution-backed. New SOTA for open-source 32B-class models on SWE-bench.

SWE-Hero-32B achieves 62.2% on SWE-bench Verified and 44.1% on SWE-bench Lite with execution-based fine-tuning. Open-weight 32B model competitive with MiniMax-M2.5 (80.2% Verified).

Reframes codebase as structured AST action space. GPT-5-nano + CodeStruct jumps from 19.6% to 40.4% on SWE-bench Verified (+20.8). GPT-5 + CodeStruct hits 67.2%. Reduces errors by up to 87.8%.

32B model trained with SFT+RL via R2E-Gym. Second-best open-source 32B-class model on SWE-bench.

Reframes codebase as structured AST action space. GPT-5-nano + CodeStruct jumps from 19.6% to 40.4% on SWE-bench Verified (+20.8). GPT-5 + CodeStruct hits 67.2%. Reduces errors by up to 87.8%.

Mid-size SWE-Hero variant competitive with much larger models. Distilled from Qwen3-Coder-480B.

20.9B total params (3.6B active). Runs on 16GB consumer hardware. Apache 2.0. Impressive for its size — near gpt-oss-120b SWE-bench score.

Highest open-weight SWE-bench Verified score. Evaluated with OpenHands scaffold.

Mid-size GPT-5 variant. 60.4% with SWE-Agent, improves to 62.0% with structured code actions.

Open-weight 72B coding LLM, RL-trained in Docker containers (reward only when full test suite passes). SOTA among open-source workflow approaches at release.

Lightweight MoE flash variant. Solid for its size class.

Independent vals.ai test. Significant gap from xAI self-reported 72-75%, highlighting scaffold sensitivity.

Smallest SWE-Hero variant. Outperforms same-size open-source agents by significant margin.

Dense 31B model. Lagging behind MoE competitors like Qwen 3.6 35B-A3B (73.4) on agentic coding.

SWE-bench Verified via OpenHands agent; lower than Qwen3.5-35B (69.2) showing SWE tasks still challenge compact MoE models.

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

30B MoE (3B active) fine-tuned with STITCH trajectory filtering. 63% relative improvement over base (26.6%).

+25.4% over prior SLM SOTA on SWE-bench Verified. Expert-protégé collaboration framework enables 7B model to selectively seek guidance while remaining sole decision-maker.

MoE model (106B total, 12B active). Baseline with SWE-Agent; improves to 48.4% with STITCH trajectory filtering.

Reframes codebase as structured AST action space. GPT-5-nano + CodeStruct jumps from 19.6% to 40.4% on SWE-bench Verified (+20.8). GPT-5 + CodeStruct hits 67.2%. Reduces errors by up to 87.8%.

Smallest GPT-5 variant. Struggles with SWE-agent workflows (46.6% empty-patch rate), but improves to 40.4% with structured code actions.

MoE model with 4B active params. Significantly behind Qwen 3.6 35B-A3B on agentic coding.

Reframes codebase as structured AST action space. GPT-5-nano + CodeStruct jumps from 19.6% to 40.4% on SWE-bench Verified (+20.8). GPT-5 + CodeStruct hits 67.2%. Reduces errors by up to 87.8%.

Reframes codebase as structured AST action space. GPT-5-nano + CodeStruct jumps from 19.6% to 40.4% on SWE-bench Verified (+20.8). GPT-5 + CodeStruct hits 67.2%. Reduces errors by up to 87.8%.

Small open-weight model baseline on SWE-Agent. Too small for effective autonomous SWE without specialized fine-tuning.