Best local LLM for agents · 2026
The best local LLM for agents and tool use, ranked for your VRAM
The agents shortlist is ordered by what the vendors and the community report for function calling, multi-step tool loops and long-horizon tasks — Terminal-Bench, OSWorld, Agents' Last Exam — as of the catalogue date. For each memory class below we list every model on that list that actually fits at Q4_K_M with an 8K context, so the top row is the pick and the rest are what you give up by choosing it. Numbers are computed, not benchmarked; how the shortlist works.
An agent re-reads its whole history every step: the tool schemas, every earlier call and its output. 8K runs out within a few turns of real work, so plan for 32K or more and check the "max context" column before you wire a model into a loop.
Best agent LLM for 6 GB VRAM
5.3 GB usable · speeds for a GeForce RTX 3060 Laptop everything that fits 6 GB →Qwen3.5 4B is the pick: Function calling in 8 GB. 2 other models on the agents shortlist also fit, in order of capability:
| # | Model | Why | Params | Quant · total | Max context | Tok/s est. | |
|---|---|---|---|---|---|---|---|
| 1 | Qwen3.5 4B Alibaba · Apache 2.0 |
Function calling in 8 GB. | 4.66B | Q4_K_M · 3.5 GB | 66K | 78 | Runs great |
| 2 | Ling 3.0 Tiny 7.9B-A1.3B inclusionAI · MIT |
Tool use and reasoning at 1.3B active, sized for edge boxes. | 7.9B MoE | Q4_K_M · 5.1 GB | 39K | 107 | Tight fit |
| 3 | Granite 4.1 3B IBM · Apache 2.0 |
Small, dense, follows tool schemas. | 3.4B | Q4_K_M · 3.1 GB | 35K | 106 | Runs great |
Best agent LLM for 8 GB VRAM
7.0 GB usable · speeds for a GeForce RTX 4060 everything that fits 8 GB →Granite 4.1 8B is the pick: Enterprise-grade tool calling in a dense 8B, no thinking overhead. 5 other models on the agents shortlist also fit, in order of capability:
| # | Model | Why | Params | Quant · total | Max context | Tok/s est. | |
|---|---|---|---|---|---|---|---|
| 1 | Granite 4.1 8B IBM · Apache 2.0 |
Enterprise-grade tool calling in a dense 8B, no thinking overhead. | 8.79B | Q4_K_M · 6.8 GB | 9K | 33 | Tight fit |
| 2 | Fara 7B Microsoft · MIT |
Web computer-use only — clicks and fills forms from screenshots, and stops to ask before anything irreversible. | 8.29B | Q4_K_M · 5.7 GB | 31K | 35 | Runs great |
| 3 | Qwen3.5 9B Alibaba · Apache 2.0 |
Tool use on 8–12 GB. | 9.65B | Q4_K_M · 6.3 GB | 31K | 30 | Tight fit |
| 4 | Qwen3.5 4B Alibaba · Apache 2.0 |
Function calling in 8 GB. | 4.66B | Q4_K_M · 3.5 GB | 120K | 63 | Runs great |
| 5 | Ling 3.0 Tiny 7.9B-A1.3B inclusionAI · MIT |
Tool use and reasoning at 1.3B active, sized for edge boxes. | 7.9B MoE | Q4_K_M · 5.1 GB | 128K | 87 | Runs great |
| 6 | Granite 4.1 3B IBM · Apache 2.0 |
Small, dense, follows tool schemas. | 3.4B | Q4_K_M · 3.1 GB | 57K | 86 | Runs great |
Best agent LLM for 12 GB VRAM
10.6 GB usable · speeds for a GeForce RTX 3060 12 GB everything that fits 12 GB →Granite 4.1 8B is the pick: Enterprise-grade tool calling in a dense 8B, no thinking overhead. 5 other models on the agents shortlist also fit, in order of capability:
| # | Model | Why | Params | Quant · total | Max context | Tok/s est. | |
|---|---|---|---|---|---|---|---|
| 1 | Granite 4.1 8B IBM · Apache 2.0 |
Enterprise-grade tool calling in a dense 8B, no thinking overhead. | 8.79B | Q4_K_M · 6.8 GB | 32K | 44 | Runs great |
| 2 | Fara 7B Microsoft · MIT |
Web computer-use only — clicks and fills forms from screenshots, and stops to ask before anything irreversible. | 8.29B | Q4_K_M · 5.7 GB | 97K | 47 | Runs great |
| 3 | Qwen3.5 9B Alibaba · Apache 2.0 |
Tool use on 8–12 GB. | 9.65B | Q4_K_M · 6.3 GB | 146K | 40 | Runs great |
| 4 | Qwen3.5 4B Alibaba · Apache 2.0 |
Function calling in 8 GB. | 4.66B | Q4_K_M · 3.5 GB | 236K | 83 | Runs great |
| 5 | Ling 3.0 Tiny 7.9B-A1.3B inclusionAI · MIT |
Tool use and reasoning at 1.3B active, sized for edge boxes. | 7.9B MoE | Q4_K_M · 5.1 GB | 128K | 115 | Runs great |
| 6 | Granite 4.1 3B IBM · Apache 2.0 |
Small, dense, follows tool schemas. | 3.4B | Q4_K_M · 3.1 GB | 103K | 114 | Runs great |
Best agent LLM for 16 GB VRAM
14.4 GB usable · speeds for a GeForce RTX 5080 everything that fits 16 GB →gpt-oss 20B is the pick: Harmony-format tool calling, native 4-bit. 6 other models on the agents shortlist also fit, in order of capability:
| # | Model | Why | Params | Quant · total | Max context | Tok/s est. | |
|---|---|---|---|---|---|---|---|
| 1 | gpt-oss 20B OpenAI · Apache 2.0 |
Harmony-format tool calling, native 4-bit. | 20.9B MoE | MXFP4 · 11.6 GB | 126K | 120 | Runs great |
| 2 | Granite 4.1 8B IBM · Apache 2.0 |
Enterprise-grade tool calling in a dense 8B, no thinking overhead. | 8.79B | Q4_K_M · 6.8 GB | 56K | 118 | Runs great |
| 3 | Fara 7B Microsoft · MIT |
Web computer-use only — clicks and fills forms from screenshots, and stops to ask before anything irreversible. | 8.29B | Q4_K_M · 5.7 GB | 125K | 125 | Runs great |
| 4 | Qwen3.5 9B Alibaba · Apache 2.0 |
Tool use on 8–12 GB. | 9.65B | Q4_K_M · 6.3 GB | 256K | 107 | Runs great |
| 5 | Qwen3.5 4B Alibaba · Apache 2.0 |
Function calling in 8 GB. | 4.66B | Q4_K_M · 3.5 GB | 256K | 222 | Runs great |
| 6 | Ling 3.0 Tiny 7.9B-A1.3B inclusionAI · MIT |
Tool use and reasoning at 1.3B active, sized for edge boxes. | 7.9B MoE | Q4_K_M · 5.1 GB | 128K | 306 | Runs great |
| 7 | Granite 4.1 3B IBM · Apache 2.0 |
Small, dense, follows tool schemas. | 3.4B | Q4_K_M · 3.1 GB | 128K | 304 | Runs great |
Best agent LLM for 24 GB VRAM
22.4 GB usable · speeds for a GeForce RTX 4090 everything that fits 24 GB →Qwen3.8 27B is the pick: Computer use and tool calling are what 3.8 was trained for (OSWorld 84). 11 other models on the agents shortlist also fit, in order of capability:
| # | Model | Why | Params | Quant · total | Max context | Tok/s est. | |
|---|---|---|---|---|---|---|---|
| 1 | Qwen3.8 27B Alibaba · Apache 2.0 |
Computer use and tool calling are what 3.8 was trained for (OSWorld 84). | 27.8B | Q4_K_M · 16.7 GB | 98K | 39 | Runs great |
| 2 | Muse Glimmer 30B Meta · Apache 2.0 |
Distilled specifically for always-on local agent workflows. | 29.8B | Q4_K_M · 17.5 GB | 128K | 36 | Runs great |
| 3 | Qwen3.6 35B-A3B Alibaba · Apache 2.0 |
Reliable function calling at MoE speed. | 35.9B MoE | Q4_K_M · 20.9 GB | 82K | 126 | Tight fit |
| 4 | Nemotron 3.5 Lightning 30B-A3B NVIDIA · OpenMDW-1.1 |
Made for the execution layer: high-volume tool calls at up to 4× the speed. | 31.6B MoE | Q4_K_M · 18.4 GB | 256K | 130 | Runs great |
| 5 | GLM-4.7-Flash 30B-A3B Z.ai · MIT |
Agentic coding and tool use in a 30B-A3B, MIT. | 31.2B MoE | Q4_K_M · 18.6 GB | 82K | 139 | Runs great |
| 6 | gpt-oss 20B OpenAI · Apache 2.0 |
Harmony-format tool calling, native 4-bit. | 20.9B MoE | MXFP4 · 11.6 GB | 128K | 126 | Runs great |
| 7 | Granite 4.1 8B IBM · Apache 2.0 |
Enterprise-grade tool calling in a dense 8B, no thinking overhead. | 8.79B | Q4_K_M · 6.8 GB | 107K | 123 | Runs great |
| 8 | Fara 7B Microsoft · MIT |
Web computer-use only — clicks and fills forms from screenshots, and stops to ask before anything irreversible. | 8.29B | Q4_K_M · 5.7 GB | 125K | 131 | Runs great |
| 9 | Qwen3.5 9B Alibaba · Apache 2.0 |
Tool use on 8–12 GB. | 9.65B | Q4_K_M · 6.3 GB | 256K | 112 | Runs great |
| 10 | Qwen3.5 4B Alibaba · Apache 2.0 |
Function calling in 8 GB. | 4.66B | Q4_K_M · 3.5 GB | 256K | 233 | Runs great |
| 11 | Ling 3.0 Tiny 7.9B-A1.3B inclusionAI · MIT |
Tool use and reasoning at 1.3B active, sized for edge boxes. | 7.9B MoE | Q4_K_M · 5.1 GB | 128K | 321 | Runs great |
| 12 | Granite 4.1 3B IBM · Apache 2.0 |
Small, dense, follows tool schemas. | 3.4B | Q4_K_M · 3.1 GB | 128K | 319 | Runs great |
Best agent LLM for 32 GB VRAM
30.4 GB usable · speeds for a GeForce RTX 5090 everything that fits 32 GB →Qwen3.8 27B is the pick: Computer use and tool calling are what 3.8 was trained for (OSWorld 84). 11 other models on the agents shortlist also fit, in order of capability:
| # | Model | Why | Params | Quant · total | Max context | Tok/s est. | |
|---|---|---|---|---|---|---|---|
| 1 | Qwen3.8 27B Alibaba · Apache 2.0 |
Computer use and tool calling are what 3.8 was trained for (OSWorld 84). | 27.8B | Q4_K_M · 16.7 GB | 226K | 69 | Runs great |
| 2 | Muse Glimmer 30B Meta · Apache 2.0 |
Distilled specifically for always-on local agent workflows. | 29.8B | Q4_K_M · 17.5 GB | 128K | 65 | Runs great |
| 3 | Qwen3.6 35B-A3B Alibaba · Apache 2.0 |
Reliable function calling at MoE speed. | 35.9B MoE | Q4_K_M · 20.9 GB | 256K | 225 | Runs great |
| 4 | Nemotron 3.5 Lightning 30B-A3B NVIDIA · OpenMDW-1.1 |
Made for the execution layer: high-volume tool calls at up to 4× the speed. | 31.6B MoE | Q4_K_M · 18.4 GB | 256K | 232 | Runs great |
| 5 | GLM-4.7-Flash 30B-A3B Z.ai · MIT |
Agentic coding and tool use in a 30B-A3B, MIT. | 31.2B MoE | Q4_K_M · 18.6 GB | 198K | 247 | Runs great |
| 6 | gpt-oss 20B OpenAI · Apache 2.0 |
Harmony-format tool calling, native 4-bit. | 20.9B MoE | MXFP4 · 11.6 GB | 128K | 224 | Runs great |
| 7 | Granite 4.1 8B IBM · Apache 2.0 |
Enterprise-grade tool calling in a dense 8B, no thinking overhead. | 8.79B | Q4_K_M · 6.8 GB | 128K | 219 | Runs great |
| 8 | Fara 7B Microsoft · MIT |
Web computer-use only — clicks and fills forms from screenshots, and stops to ask before anything irreversible. | 8.29B | Q4_K_M · 5.7 GB | 125K | 233 | Runs great |
| 9 | Qwen3.5 9B Alibaba · Apache 2.0 |
Tool use on 8–12 GB. | 9.65B | Q4_K_M · 6.3 GB | 256K | 200 | Runs great |
| 10 | Qwen3.5 4B Alibaba · Apache 2.0 |
Function calling in 8 GB. | 4.66B | Q4_K_M · 3.5 GB | 256K | 414 | Runs great |
| 11 | Ling 3.0 Tiny 7.9B-A1.3B inclusionAI · MIT |
Tool use and reasoning at 1.3B active, sized for edge boxes. | 7.9B MoE | Q4_K_M · 5.1 GB | 128K | 571 | Runs great |
| 12 | Granite 4.1 3B IBM · Apache 2.0 |
Small, dense, follows tool schemas. | 3.4B | Q4_K_M · 3.1 GB | 128K | 567 | Runs great |
Best agent LLM for 48 GB VRAM
46.4 GB usable · speeds for an RTX 6000 Ada everything that fits 48 GB →Qwen3.8 27B is the pick: Computer use and tool calling are what 3.8 was trained for (OSWorld 84). 11 other models on the agents shortlist also fit, in order of capability:
| # | Model | Why | Params | Quant · total | Max context | Tok/s est. | |
|---|---|---|---|---|---|---|---|
| 1 | Qwen3.8 27B Alibaba · Apache 2.0 |
Computer use and tool calling are what 3.8 was trained for (OSWorld 84). | 27.8B | Q4_K_M · 16.7 GB | 256K | 37 | Runs great |
| 2 | Muse Glimmer 30B Meta · Apache 2.0 |
Distilled specifically for always-on local agent workflows. | 29.8B | Q4_K_M · 17.5 GB | 128K | 35 | Runs great |
| 3 | Qwen3.6 35B-A3B Alibaba · Apache 2.0 |
Reliable function calling at MoE speed. | 35.9B MoE | Q4_K_M · 20.9 GB | 256K | 120 | Runs great |
| 4 | Nemotron 3.5 Lightning 30B-A3B NVIDIA · OpenMDW-1.1 |
Made for the execution layer: high-volume tool calls at up to 4× the speed. | 31.6B MoE | Q4_K_M · 18.4 GB | 256K | 124 | Runs great |
| 5 | GLM-4.7-Flash 30B-A3B Z.ai · MIT |
Agentic coding and tool use in a 30B-A3B, MIT. | 31.2B MoE | Q4_K_M · 18.6 GB | 198K | 133 | Runs great |
| 6 | gpt-oss 20B OpenAI · Apache 2.0 |
Harmony-format tool calling, native 4-bit. | 20.9B MoE | MXFP4 · 11.6 GB | 128K | 120 | Runs great |
| 7 | Granite 4.1 8B IBM · Apache 2.0 |
Enterprise-grade tool calling in a dense 8B, no thinking overhead. | 8.79B | Q4_K_M · 6.8 GB | 128K | 118 | Runs great |
| 8 | Fara 7B Microsoft · MIT |
Web computer-use only — clicks and fills forms from screenshots, and stops to ask before anything irreversible. | 8.29B | Q4_K_M · 5.7 GB | 125K | 125 | Runs great |
| 9 | Qwen3.5 9B Alibaba · Apache 2.0 |
Tool use on 8–12 GB. | 9.65B | Q4_K_M · 6.3 GB | 256K | 107 | Runs great |
| 10 | Qwen3.5 4B Alibaba · Apache 2.0 |
Function calling in 8 GB. | 4.66B | Q4_K_M · 3.5 GB | 256K | 222 | Runs great |
| 11 | Ling 3.0 Tiny 7.9B-A1.3B inclusionAI · MIT |
Tool use and reasoning at 1.3B active, sized for edge boxes. | 7.9B MoE | Q4_K_M · 5.1 GB | 128K | 306 | Runs great |
| 12 | Granite 4.1 3B IBM · Apache 2.0 |
Small, dense, follows tool schemas. | 3.4B | Q4_K_M · 3.1 GB | 128K | 304 | Runs great |
Best agent LLM for 16 GB unified memory
10.7 GB usable · speeds for an M4 · 16 GB everything that fits 16 GB →Granite 4.1 8B is the pick: Enterprise-grade tool calling in a dense 8B, no thinking overhead. 5 other models on the agents shortlist also fit, in order of capability:
| # | Model | Why | Params | Quant · total | Max context | Tok/s est. | |
|---|---|---|---|---|---|---|---|
| 1 | Granite 4.1 8B IBM · Apache 2.0 |
Enterprise-grade tool calling in a dense 8B, no thinking overhead. | 8.79B | Q4_K_M · 6.8 GB | 33K | 14 | Runs great |
| 2 | Fara 7B Microsoft · MIT |
Web computer-use only — clicks and fills forms from screenshots, and stops to ask before anything irreversible. | 8.29B | Q4_K_M · 5.7 GB | 99K | 14 | Runs great |
| 3 | Qwen3.5 9B Alibaba · Apache 2.0 |
Tool use on 8–12 GB. | 9.65B | Q4_K_M · 6.3 GB | 149K | 12 | Runs great |
| 4 | Qwen3.5 4B Alibaba · Apache 2.0 |
Function calling in 8 GB. | 4.66B | Q4_K_M · 3.5 GB | 239K | 26 | Runs great |
| 5 | Ling 3.0 Tiny 7.9B-A1.3B inclusionAI · MIT |
Tool use and reasoning at 1.3B active, sized for edge boxes. | 7.9B MoE | Q4_K_M · 5.1 GB | 128K | 43 | Runs great |
| 6 | Granite 4.1 3B IBM · Apache 2.0 |
Small, dense, follows tool schemas. | 3.4B | Q4_K_M · 3.1 GB | 104K | 35 | Runs great |
Best agent LLM for 64 GB unified memory
48.0 GB usable · speeds for an M4 Max · 64 GB everything that fits 64 GB →Qwen3.8 27B is the pick: Computer use and tool calling are what 3.8 was trained for (OSWorld 84). 11 other models on the agents shortlist also fit, in order of capability:
| # | Model | Why | Params | Quant · total | Max context | Tok/s est. | |
|---|---|---|---|---|---|---|---|
| 1 | Qwen3.8 27B Alibaba · Apache 2.0 |
Computer use and tool calling are what 3.8 was trained for (OSWorld 84). | 27.8B | Q4_K_M · 16.7 GB | 256K | 20 | Runs great |
| 2 | Muse Glimmer 30B Meta · Apache 2.0 |
Distilled specifically for always-on local agent workflows. | 29.8B | Q4_K_M · 17.5 GB | 128K | 18 | Runs great |
| 3 | Qwen3.6 35B-A3B Alibaba · Apache 2.0 |
Reliable function calling at MoE speed. | 35.9B MoE | Q4_K_M · 20.9 GB | 256K | 77 | Runs great |
| 4 | Nemotron 3.5 Lightning 30B-A3B NVIDIA · OpenMDW-1.1 |
Made for the execution layer: high-volume tool calls at up to 4× the speed. | 31.6B MoE | Q4_K_M · 18.4 GB | 256K | 79 | Runs great |
| 5 | GLM-4.7-Flash 30B-A3B Z.ai · MIT |
Agentic coding and tool use in a 30B-A3B, MIT. | 31.2B MoE | Q4_K_M · 18.6 GB | 198K | 84 | Runs great |
| 6 | gpt-oss 20B OpenAI · Apache 2.0 |
Harmony-format tool calling, native 4-bit. | 20.9B MoE | MXFP4 · 11.6 GB | 128K | 76 | Runs great |
| 7 | Granite 4.1 8B IBM · Apache 2.0 |
Enterprise-grade tool calling in a dense 8B, no thinking overhead. | 8.79B | Q4_K_M · 6.8 GB | 128K | 62 | Runs great |
| 8 | Fara 7B Microsoft · MIT |
Web computer-use only — clicks and fills forms from screenshots, and stops to ask before anything irreversible. | 8.29B | Q4_K_M · 5.7 GB | 125K | 65 | Runs great |
| 9 | Qwen3.5 9B Alibaba · Apache 2.0 |
Tool use on 8–12 GB. | 9.65B | Q4_K_M · 6.3 GB | 256K | 56 | Runs great |
| 10 | Qwen3.5 4B Alibaba · Apache 2.0 |
Function calling in 8 GB. | 4.66B | Q4_K_M · 3.5 GB | 256K | 116 | Runs great |
| 11 | Ling 3.0 Tiny 7.9B-A1.3B inclusionAI · MIT |
Tool use and reasoning at 1.3B active, sized for edge boxes. | 7.9B MoE | Q4_K_M · 5.1 GB | 128K | 195 | Runs great |
| 12 | Granite 4.1 3B IBM · Apache 2.0 |
Small, dense, follows tool schemas. | 3.4B | Q4_K_M · 3.1 GB | 128K | 160 | Runs great |
Best agent LLM for 128 GB unified memory
96.0 GB usable · speeds for an M4 Max · 128 GB everything that fits 128 GB →Ling 3.0 Flash 124B-A5B is the pick: Agentic workflows are what Ling 3.0 was trained for; MIT, 5B active. 12 other models on the agents shortlist also fit, in order of capability:
| # | Model | Why | Params | Quant · total | Max context | Tok/s est. | |
|---|---|---|---|---|---|---|---|
| 1 | Ling 3.0 Flash 124B-A5B inclusionAI · MIT |
Agentic workflows are what Ling 3.0 was trained for; MIT, 5B active. | 124B MoE | Q4_K_M · 70.4 GB | 256K | 50 | Runs great |
| 2 | Qwen3.8 27B Alibaba · Apache 2.0 |
Computer use and tool calling are what 3.8 was trained for (OSWorld 84). | 27.8B | Q4_K_M · 16.7 GB | 256K | 20 | Runs great |
| 3 | Muse Glimmer 30B Meta · Apache 2.0 |
Distilled specifically for always-on local agent workflows. | 29.8B | Q4_K_M · 17.5 GB | 128K | 18 | Runs great |
| 4 | Qwen3.6 35B-A3B Alibaba · Apache 2.0 |
Reliable function calling at MoE speed. | 35.9B MoE | Q4_K_M · 20.9 GB | 256K | 77 | Runs great |
| 5 | Nemotron 3.5 Lightning 30B-A3B NVIDIA · OpenMDW-1.1 |
Made for the execution layer: high-volume tool calls at up to 4× the speed. | 31.6B MoE | Q4_K_M · 18.4 GB | 256K | 79 | Runs great |
| 6 | GLM-4.7-Flash 30B-A3B Z.ai · MIT |
Agentic coding and tool use in a 30B-A3B, MIT. | 31.2B MoE | Q4_K_M · 18.6 GB | 198K | 84 | Runs great |
| 7 | gpt-oss 20B OpenAI · Apache 2.0 |
Harmony-format tool calling, native 4-bit. | 20.9B MoE | MXFP4 · 11.6 GB | 128K | 76 | Runs great |
| 8 | Granite 4.1 8B IBM · Apache 2.0 |
Enterprise-grade tool calling in a dense 8B, no thinking overhead. | 8.79B | Q4_K_M · 6.8 GB | 128K | 62 | Runs great |
| 9 | Fara 7B Microsoft · MIT |
Web computer-use only — clicks and fills forms from screenshots, and stops to ask before anything irreversible. | 8.29B | Q4_K_M · 5.7 GB | 125K | 65 | Runs great |
| 10 | Qwen3.5 9B Alibaba · Apache 2.0 |
Tool use on 8–12 GB. | 9.65B | Q4_K_M · 6.3 GB | 256K | 56 | Runs great |
| 11 | Qwen3.5 4B Alibaba · Apache 2.0 |
Function calling in 8 GB. | 4.66B | Q4_K_M · 3.5 GB | 256K | 116 | Runs great |
| 12 | Ling 3.0 Tiny 7.9B-A1.3B inclusionAI · MIT |
Tool use and reasoning at 1.3B active, sized for edge boxes. | 7.9B MoE | Q4_K_M · 5.1 GB | 128K | 195 | Runs great |
| 13 | Granite 4.1 3B IBM · Apache 2.0 |
Small, dense, follows tool schemas. | 3.4B | Q4_K_M · 3.1 GB | 128K | 160 | Runs great |
The whole agents shortlist, and what each one needs
Every model on the list, best first, with the smallest discrete card and the smallest Mac that run it with 15% headroom at the recommended quantisation and 8K context. Every class has its own page, Macs from 8 to 512 GB included — the full list.
| # | Model | Params | Weights | Needs (GPU) | Needs (Mac) | Ollama tag |
|---|---|---|---|---|---|---|
| 1 | DeepSeek V4.1 Flash 552B-A16B New Twice V4 Flash's backbone (552B, 16B active per generated token) plus 196B of Engram lookup tables (189 GiB at FP8) that DwarfStar streams from SSD, so they are not counted here. A 512 GB Mac model, and as of October 2026 no mainline llama.cpp or Ollama build: DwarfStar on a Mac, vLLM across four GPUs. DeepSeek puts the cache at 890 bytes/token; modelled conservatively as 4 latent layers plus a 128-token window. |
552B MoE | 310.4 GB @ Q4_K_M | more than one card | 512 GB Mac | — |
| 2 | Kimi K2.6 1T-A32B The open coding-agent benchmark leader of spring 2026 (80.2 SWE-bench). A 512 GB Mac Studio pair, or a ceiling. |
1027B MoE | 577.5 GB @ Q4_K_M | more than one card | no single Mac | — |
| 3 | GLM-5.3 744B-A40B Same base as GLM-5.2, new post-training: Z.ai's most capable open-weights coder (Terminal-Bench 3.0 28.3, DeepSWE 66.9). Not MIT any more — its own GLM-5.3 licence. Same ceiling as 5.2: 512 GB of unified memory at Q4. |
753B MoE | 423.4 GB @ Q4_K_M | more than one card | no single Mac | — |
| 4 | GLM-5.3-Flash 320B-A18B The first natively multimodal GLM-5 and the first hybrid: 34 linear-attention blocks and 11 sparse-attention blocks with a 512-wide latent cache, so a 1M window stays affordable. Z.ai says it beats GLM-5.2 at 18B active; MIT. |
321B MoE | 180.5 GB @ Q4_K_M | more than one card | 512 GB Mac | — |
| 5 | DeepSeek V4 Flash 284B-A13B The V4 for 192–256 GB machines: Q3 squeezes into 192 GB, Q4 wants 256 GB, and 128 GB falls short even at Q2. Cache is modelled as a 576-wide latent; V4 compresses it further at long context, so this is conservative. |
284B MoE | 159.7 GB @ Q4_K_M | 192 GB card | 256 GB Mac | — |
| 6 | Qwen3.8-Flash-Next 180B-A6B The open preview of the Qwen4 architecture: a 125B-A6B hybrid (Gated DeltaNet + sparse attention, KV cache on 12 of 48 blocks) plus a 51B n-gram embedding and a 4B draft head — 180B on disk, 6B active. The hosted "Qwen3.8-Flash" is this model with a 1M window. |
180B MoE | 101.2 GB @ Q4_K_M | 141 GB card | 192 GB Mac | qwen3.8-flash-next:125b-a6b-nvfp4 |
| 7 | Ling 3.0 Flash 124B-A5B A 124B hybrid (5 linear-attention layers per MLA layer) with 5.1B active: SWE-bench Pro 56.6 and AIME 93 claimed. Built for 96–128 GB machines. |
124B MoE | 69.7 GB @ Q4_K_M | 141 GB card | 128 GB Mac | — |
| 8 | Qwen3.8 27B The current default local Qwen: dense 27B, text + image + video, 262K context. Only 16 of its 64 blocks keep a KV cache, so long context is cheap. |
27.8B | 15.6 GB @ Q4_K_M | 24 GB card | 32 GB Mac | qwen3.8:27b |
| 9 | Muse Glimmer 30B Meta's first open weights since Llama 4: a dense 30B distilled from Muse Spark for always-on local agents. Two KV heads keep the cache small. |
29.8B | 16.8 GB @ Q4_K_M | 24 GB card | 32 GB Mac | muse-glimmer:30b |
| 10 | Qwen3.6 35B-A3B Mixture of experts with ~3B active: the fastest serious model a 24 GB card runs, and the best MoE under 40B on agentic coding. |
35.9B MoE | 20.2 GB @ Q4_K_M | 32 GB card | 36 GB Mac | qwen3.6:35b-a3b |
| 11 | Nemotron 3.5 Lightning 30B-A3B Mamba-2 + MoE hybrid built for the execution layer of agents: only 6 attention blocks, so the KV cache is almost free. Weights, data and recipe all open. |
31.6B MoE | 17.8 GB @ Q4_K_M | 24 GB card | 32 GB Mac | nemotron-3.5-lightning:30b-a3b |
| 12 | GLM-4.7-Flash 30B-A3B MIT-licensed 30B-A3B tuned for agentic coding, with a DeepSeek-style latent KV cache. 60–80 tok/s reported on a 4090. |
31.2B MoE | 17.5 GB @ Q4_K_M | 24 GB card | 32 GB Mac | glm-4.7-flash:latest |
| 13 | gpt-oss 20B Ships natively in MXFP4, so the 4-bit weights are the reference weights, not a lossy copy. Fits 16 GB. |
20.9B MoE | 10.8 GB @ MXFP4 | 16 GB card | 24 GB Mac | gpt-oss:20b |
| 14 | Granite 4.1 8B Matches the old Granite 4.0 32B MoE at a quarter of the size. Fast, Apache 2.0, no reasoning traces. |
8.79B | 4.9 GB @ Q4_K_M | 10 GB card | 16 GB Mac | granite4.1:8b |
| 15 | Fara 7B A web computer-use agent on a Qwen2.5-VL base — it clicks, fills forms and stops for permission. Not a chat model; size it like an 8B with vision. |
8.29B | 4.7 GB @ Q4_K_M | 8 GB card | 16 GB Mac | — |
| 16 | Qwen3.5 9B The default for 8–12 GB cards in 2026: beats every older 8B on every published benchmark, with vision. |
9.65B | 5.4 GB @ Q4_K_M | 10 GB card | 16 GB Mac | qwen3.5:9b |
| 17 | Qwen3.5 4B The 8 GB coding agent. Q4 lands near 3.4 GB, leaving room for a real context window. |
4.66B | 2.6 GB @ Q4_K_M | 6 GB card | 8 GB Mac | qwen3.5:4b |
| 18 | Ling 3.0 Tiny 7.9B-A1.3B An 8B MoE with 1.3B active and a latent KV cache on only 6 of 24 layers — reasoning and tool use sized for Apple Silicon and edge boxes. |
7.9B MoE | 4.4 GB @ Q4_K_M | 8 GB card | 16 GB Mac | — |
| 19 | Granite 4.1 3B Dense, small, enterprise-flavoured: tool calling and instruction following, no thinking mode. |
3.4B | 1.9 GB @ Q4_K_M | 6 GB card | 8 GB Mac | granite4.1:3b |
Tool calling depends on the chat template as much as on the weights. Start llama-server with --jinja so the model's own template formats the calls; Ollama and LM Studio do this for you. A model that fits with headroom but fumbles its JSON usually has a template problem, not a capability one.
Other jobs