mixed / LB-2026-07.2
Global comparison
Every ranked model is scored on the same frozen suite across Agentic, Knowledge, Instruction, Coding, Math. This is the initial measured ladder — more models land as runs are submitted. The Local Intelligence Index (LB-2026-07.2 | 25/22.5/22.5/22.5/7.5) appears only after a complete run publishes. Profile: Agentic / Knowledge / Instruction / Coding / Math.
Every complete project and community run shares this ranking and the same composite. The global view is a cross-family reference; browse model families to choose among related variants. Note: the Agentic axis is near-floor for every current local entrant, so it compresses headline gaps — read the composite alongside the per-axis columns and unweighted diagnostics.
Local Intelligence Index | LB-2026-07.2
Headline = 0.25 Agentic + 0.225 Knowledge + 0.225 Instruction + 0.225 Coding + 0.075 Math. Agentic is AppWorld task-goal completion on the fixed 96-task subset; ranks require the complete headline profile.
5 complete ranked runs
Swipe horizontally for scores and axes →
| Rank | VRAM @8k | |||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 |  Gemma 4 31B ITLB-2026-07.2Q4_K_M Gemma 4 | run by local-bench | 51.7±3.1 Agentic 10.4 / Knowledge 87.4 / Instruction 79.3 / Coding 35.5 / Math 48.2 | 10/96 — 10.4%±8.5 diagnostics
| 87.4±3.6 | 79.3±4.8 | 35.5±7.8 | 48.2±8.6 | 20.9 GB | llama.cppb9852/fd1a05791 | RTX 5090 · 32 GB | 1,918 | ~40 s | 27.8 h |
| 2 |  Qwen3.6 27BLB-2026-07.2Q4_K_M Qwen3.6 | run by local-bench | 43.2±2.9 Agentic 8.3 / Knowledge 83.4 / Instruction 65.0 / Coding 25.5 / Math 26.6 | 8/96 — 8.3%±7.4 diagnostics
| 83.4±3.9 | 65.0±5.4 | 25.5±7.1 | 26.6±7.2 | 19.5 GB | llama.cppb9852/fd1a05791 | RTX 5090 · 32 GB | 3,287 | ~47 s | 23.7 h |
| 3 | Qwopus 3.6 27B v2 MTPLB-2026-07.2 Q4_K_M Qwen3.6 Fine-tune of Qwen3.6 27B | run by local-bench | 42.1±3.0 Agentic 9.4 / Knowledge 80.4 / Instruction 57.8 / Coding 29.8 / Math 25.9 | 9/96 — 9.4%±8.0 diagnostics
| 80.4±4.2 | 57.8±5.8 | 29.8±7.8 | 25.9±7.2 | 18.1 GB | llama.cppb9852/fd1a05791 | RTX 5090 · 32 GB | 1,623 | ~25 s | 23.6 h |
| 4 | Qwen3.6 35B A3BLB-2026-07.2 UD-Q4_K_M Qwen3.6 | run by local-bench | 41.0±2.7 Agentic 6.3 / Knowledge 82.0 / Instruction 57.5 / Coding 28.4 / Math 22.3 | 6/96 — 6.3%±5.4 diagnostics
| 82.0±4.2 | 57.5±5.8 | 28.4±7.8 | 22.3±7.2 | 24.3 GB | llama.cppb9852/fd1a05791 | RTX 5090 · 32 GB | 3,639 | ~19 s | 9.2 h |
| 5 | Gemma 4 12B ITLB-2026-07.2 QAT Q4_K_XL Gemma 4 | run by local-bench | 40.3±2.7 Agentic 4.2 / Knowledge 75.6 / Instruction 66.7 / Coding 28.4 / Math 11.5 | 4/96 — 4.2%±5.8 diagnostics
| 75.6±4.8 | 66.7±5.4 | 28.4±7.8 | 11.5±5.8 | 8.4 GB | llama.cppb9852/fd1a05791 | RTX 5090 · 32 GB | 8,259 | ~60 s | 17.2 h |
Agentic = AppWorld task-goal completion; 25% weight; near-floor scores compress gaps.
Not yet benchmarked — 159 catalog models on the roadmap
- Command A Plus 05-2026 BF16Command A
- Command RCommand R
- Command R+Command R
- Command R7B 12-2024Command R
- DeepSeek R1DeepSeek R1
- DeepSeek R1 0528DeepSeek R1
- DeepScaleR 1.5B PreviewDeepSeek R1 Distill
- DeepSeek R1 Distill Llama 70BDeepSeek R1 Distill
- DeepSeek R1 Distill Llama 8BDeepSeek R1 Distill
- DeepSeek R1 Distill Qwen 1.5BDeepSeek R1 Distill
- DeepSeek R1 Distill Qwen 14BDeepSeek R1 Distill
- DeepSeek R1 Distill Qwen 32BDeepSeek R1 Distill
- DeepSeek R1 Distill Qwen 7BDeepSeek R1 Distill
- Nemotron Research Reasoning Qwen 1.5BDeepSeek R1 Distill
- DeepSeek V3DeepSeek V3
- DeepSeek V3 0324DeepSeek V3
- DeepSeek V3.1DeepSeek V3
- DeepSeek V3.1 TerminusDeepSeek V3
- DeepSeek V3.2DeepSeek V3
- DeepSeek V3.2 ExpDeepSeek V3
- DeepSeek V4 FlashDeepSeek V4
- DeepSeek V4 ProDeepSeek V4
- Gemma 3 12B ITGemma 3
- Gemma 3 1B ITGemma 3
- Gemma 3 27B ITGemma 3
- Gemma 3 4B ITGemma 3
- Gemma 3n E2B ITGemma 3n
- Gemma 3n E4B ITGemma 3n
- Dark Scarlett v0.3 26B A4BGemma 4
- Gemma 4 12B Coder Fable ComposerGemma 4
- Gemma 4 12B it AEON Abliterated K4 BF16Gemma 4
- Gemma 4 12B OBLITERATEDGemma 4
- Gemma 4 26B A4B ITGemma 4
- gemma 4 26B A4B it uncensoredGemma 4
- gemma 4 31B it uncensoredGemma 4
- Gemma 4 E2B ITGemma 4
- Gemma 4 E4B ITGemma 4
- Gemma 4 E4B IT OBLITERATEDGemma 4
- gemma 4 E4B it uncensoredGemma 4
- Gemma 4 12B Coder Fable5Gemma 4 Coder
- GLM 5GLM 5
- GLM 5.1GLM 5
- GLM 5.2GLM 5
- GPT OSS 120BGPT OSS
- GPT OSS 20BGPT OSS
- Llama 3.1 405B InstructLlama 3.1
- Llama 3.1 70B InstructLlama 3.1
- Llama 3.1 8B InstructLlama 3.1
- Llama 3.1 Nemotron 70B Instruct HFLlama 3.1
- ToolACE 8BLlama 3.1
- Vikhr Llama3.1 8B Instruct R 21 09 24Llama 3.1
- Llama 3.2 1B InstructLlama 3.2
- Llama 3.2 3B InstructLlama 3.2
- Llama 3.3 70B InstructLlama 3.3
- Llama 3.3 70B Instruct abliteratedLlama 3.3
- Llama 4 Maverick 17B 128E InstructLlama 4
- Llama 4 Scout 17B 16E InstructLlama 4
- Ministral 3 14B InstructMinistral
- Ministral 3 3B InstructMinistral
- Ministral 3 8B InstructMinistral
- bigstral 12b 32kMistral
- Mistral 7B Instruct v0.2Mistral
- Mistral 7B Instruct v0.3Mistral
- Mistral MOE 4X7B Dark MultiVerse Uncensored Enhanced32 24BMistral
- Mistral Medium 3.5 128BMistral Medium
- Mistral Nemo 12B InstructMistral Nemo
- Vikhr Nemo 12B Instruct R 21 09 24Mistral Nemo
- Devstral Small 2 24B InstructMistral Small
- Devstral Small 2505Mistral Small
- Magistral Small 2506Mistral Small
- Mistral Small 3.1 24B InstructMistral Small
- Mistral Small 3.2 24B InstructMistral Small
- Mistral Small Instruct 2409Mistral Small
- Mixtral 8x22B InstructMixtral
- Mixtral 8x7B InstructMixtral
- Nex N2 MiniNex N2
- Nex N2 ProNex N2
- Ornith 1.0 35BOrnith
- Phi 3 Medium 4K InstructPhi 3
- Phi 3 Mini 4K InstructPhi 3
- Phi 3 Small 8K InstructPhi 3
- Phi 4Phi 4
- Phi 4 Mini InstructPhi 4
- Phi 4 ReasoningPhi 4
- Athene V2 ChatQwen2.5
- OpenThinker2 32BQwen2.5
- OpenThinker2 7BQwen2.5
- OpenThinker3 7BQwen2.5
- Qwen2.5 0.5B InstructQwen2.5
- Qwen2.5 1.5B InstructQwen2.5
- Qwen2.5 14B InstructQwen2.5
- Qwen2.5 32B InstructQwen2.5
- Qwen2.5 32B Instruct abliteratedQwen2.5
- Qwen2.5 3B InstructQwen2.5
- Qwen2.5 72B InstructQwen2.5
- Qwen2.5 72B Instruct abliteratedQwen2.5
- Qwen2.5 7B InstructQwen2.5
- Qwen2.5 7B Instruct abliterated v2Qwen2.5
- Qwen2.5 7B Instruct UncensoredQwen2.5
- RL MemoryAgent 14BQwen2.5
- TinySwallow 1.5BQwen2.5
- VulnLLM R 7BQwen2.5
- Arctic Text2SQL R1 7BQwen2.5 Coder
- Qwen2.5 Coder 14B InstructQwen2.5 Coder
- Qwen2.5 Coder 32B InstructQwen2.5 Coder
- Qwen2.5 Coder 7B InstructQwen2.5 Coder
- VL 1 CoderQwen2.5 Coder
- Qwen2.5 Math 1.5B InstructQwen2.5 Math
- Qwen2.5 Math 7B InstructQwen2.5 Math
- DeepSWE PreviewQwen3
- fable tracesQwen3
- FrogBoss 32B 2510Qwen3
- Goedel Prover V2 32BQwen3
- Goedel Prover V2 8BQwen3
- Huihui Qwen3 14B abliterated v2Qwen3
- Huihui Qwen3 30B A3B Instruct 2507 abliteratedQwen3
- Huihui Qwen3 30B A3B Thinking 2507 abliteratedQwen3
- Huihui Qwen3 4B Instruct 2507 abliteratedQwen3
- Karnak 40B v1.0Qwen3
- lynx instruct 30bQwen3
- Qwen3 0.6BQwen3
- Qwen3 1.7BQwen3
- Qwen3 14BQwen3
- Qwen3 235B A22BQwen3
- Qwen3 235B A22B Instruct 2507Qwen3
- Qwen3 235B A22B Thinking 2507Qwen3
- Qwen3 30B A3BQwen3
- Qwen3 30B A3B abliteratedQwen3
- Qwen3 30B A3B Instruct 2507Qwen3
- Qwen3 30B A3B Thinking 2507Qwen3
- Qwen3 32BQwen3
- Qwen3 32B abliteratedQwen3
- Qwen3 4BQwen3
- Qwen3 4B Instruct 2507Qwen3
- Qwen3 4B SafeRLQwen3
- Qwen3 4B Thinking 2507Qwen3
- Qwen3 8BQwen3
- Rio 3.0 Open MiniQwen3
- T pro it 2.1Qwen3
- Huihui Qwen3 Coder 30B A3B Instruct abliteratedQwen3 Coder
- Qwen3 Coder 30B A3B InstructQwen3 Coder
- Qwen3 Coder NextQwen3 Coder
- Qwen3 Next 80B A3B InstructQwen3 Next
- Qwen3 Next 80B A3B ThinkingQwen3 Next
- Qwen3.5 9BQwen3.5
- Qwythos 9B Claude Mythos 5 1MQwen3.5
- Qwythos 9B v2Qwen3.5
- Bonsai 27B TernaryQwen3.6
- Qwen3.6 27B AEON Ultimate Uncensored BF16Qwen3.6
- Qwen3.6 27B MTP pi tuneQwen3.6
- Qwen3.6 35B A3B Claude 4.7 Opus Reasoning DistilledQwen3.6
- Qwen3.6 35B A3B Uncensored HauhauCS AggressiveQwen3.6
- Qwopus 3.6 27B v1 PreviewQwen3.6
- SIQ 1 35BQwen3.6
- QwQ 32BQwQ
- Yi 1.5 34B ChatYi 1.5
- Yi 1.5 6B ChatYi 1.5
- Yi 1.5 9B ChatYi 1.5
- YiSM 34B 0rnYi 1.5