Best open-source models for coding and AI agents
Coding assistants and AI agents need reliable tool calling, long context and the stamina to keep making progress. These open models are built for that kind of work. Try two or three on your own tasks, because benchmark scores in announcements are reported by the developers themselves.
-
Kimi K3.1 Pro
Moonshot tuned this release for coding and agent work, with a 1M-token context and image input. It is the heaviest option here and needs a multi-node cluster.
-
LongCat-2.5
Meituan's MIT-licensed 1.6T multimodal mixture of experts with 48B active parameters, built for long coding and terminal tasks. Meituan's own table reports 88.4 on Terminal Bench v2.1. It needs a data-center cluster, and deployment is documented only for NPU servers.
-
GLM-5.3
A text-only model aimed at coding, with 753B parameters (40B active) and a 1M context. The license is close to MIT, with a security-review condition only for giant hosting businesses.
-
DeepSeek V4
DeepSeek's MIT-licensed pair for long-context agents: a 1.6T Pro with 49B active parameters and a 284B Flash with 13B active, both with a 1M context. By our estimate the Flash needs about 184 GB at 4-bit.
-
Ling-3.1-flash
Ant Group's open text model for agents and coding: 560B total, 25B active, 1M context, MIT license. Mostly linear attention keeps long contexts affordable, but you need the specific SGLang or vLLM builds named on its model card.
-
MiniMax M3.1 Flash
A compact 428B multimodal mixture of experts with 23B active parameters and a 1M context, trained to run in 4-bit. Read the license: commercial use needs a credit line and a notice, and written authorization above US$20 million a year.
-
Laguna M.1
Poolside's 225B mixture of experts with 23B active parameters and a 256K context, under Apache 2.0 since June 2026, with a base checkpoint. In Poolside's own table DeepSeek-V4 Flash and Qwen3.5-397B score higher, and its newer S 2.1 reports better scores at half the size. It needs a multi-GPU node.
-
Laguna S 2.1
Poolside's 118B mixture of experts with 8B active parameters and a 1M context, under the permissive OpenMDW-1.1 license. Its own table shows 70.2% on Terminal-Bench 2.1 and 59.4% on SWE-Bench Pro, though Kimi K3 and Claude Fable 5 lead on terminal tasks. By our estimate the 96 GB Q4_K_M file needs about 106 GB of memory.
-
Motif 3
Motif Technologies' MIT-licensed 314B mixture of experts with only 13.2B active parameters and a 256K context, built for tool use and terminal work. Its own table shows 74.9 on Terminal-Bench 2.1 but weaker knowledge and math scores, and the card's commands use eight H200 or B200 GPUs.
-
Step 5 Preview
StepFun's 600B mixture of experts with 27B active parameters, built for repository-level software engineering, long agent runs and finance. In StepFun's own chart it scores 67.7 on DeepSWE v1.1 at High effort, level with Kimi K3 and GLM-5.3 at Max effort. It needs an 8-GPU H200 node in FP8.
-
Step 3.7 Flash
StepFun's Apache 2.0 vision-language agent model with 198B parameters, 11B active, a 256K context and selectable reasoning levels. It scores 56.3 on SWE-Bench Pro in StepFun's own table, and a 111.5 GB GGUF runs on a 128 GB machine with StepFun's llama.cpp branch.
-
Kimi K2.7 Code
Moonshot's coding agent model built on Kimi K2.6: 1T parameters with 32B active, a 256K context, native INT4 and image and video input. Moonshot says it handles sequences of 4,000 or more tool calls, which is self-reported. The license is modified MIT.
-
MiMo-V2.6
Xiaomi's omnimodal agent models, trained with one large reinforcement-learning run across coding, general agents, visual tasks and cybersecurity. Hugging Face tags them MIT, but the repositories have no LICENSE file, so confirm the terms before building a product.
-
Ripple-65B Preview
Ricursive Intelligence's 65B dense model under Apache 2.0 with a 1M-token context, released as a preview on 4 November 2026. Its own table reports 80.5 on Terminal-Bench 2.1 against 80.1 for Reflection Beam, but Beam is ahead on SWE-bench Verified (81.6 to 76.5). The BF16 weights are about 130 GB.
-
Muse Glimmer
Meta's 30B dense model for local agents, distilled from Muse Spark and quantized to fit a 24 GB GPU, under Apache 2.0.
-
Nanbeige4.3-33B
BOSS Zhipin's 33B dense agent model under Apache 2.0, with a 512K context, released on 24 October 2026. Its own table reports 80.9 on Terminal-Bench 2.1 against 73.0 for Qwen3.8-27B, while Qwen leads on SWE-Bench Pro. An INT4 build is about 23 GB, and the repositories have no LICENSE file.
-
Laguna XS 2.1
Poolside's 33B mixture of experts with only 3B active parameters and a 256K context, under OpenMDW-1.1, and the card says it runs on a Mac with 36 GB. In Poolside's own table Qwen3.6-35B-A3B scores higher on all four coding benchmarks, and the card warned about llama.cpp, vLLM and Ollama rough edges.
-
Raon-1 mini preview
KRAFTON's 28.8B mixture of experts with 4.3B active parameters under Apache 2.0, with a 262K context and built-in thinking. Its own table reports 70.8 on SWE-bench Verified against 63.2 for Qwen3.6-35B-A3B, but 47.0 against 63.3 on Terminal-Bench Lite. The card calls it a preview, and a 17.6 GB NVFP4 build fits one 32 GB GPU.
-
Qwen3.6
Alibaba's Apache 2.0 pair tuned for agentic coding: a 35B mixture of experts with only 3B active and a 27B dense model. By our estimate they need about 25 GB and 19 GB at 4-bit.
-
GLM-4.7-Flash
Z.ai's MIT-licensed 30B mixture of experts with 3B active parameters, from January 2026 and still downloaded about a million times a month. Z.ai's own table shows 59.2 on SWE-bench Verified against 34.0 for gpt-oss-20b, but those rivals are older than Qwen3.5 and Qwen3.6. Official weights are BF16 only, so use a community 4-bit build, which needs about 22 GB by our estimate.
-
Ornith 1.5
DeepReinforce's agentic-coding models post-trained from Qwen3.5, in 9B, 35B-A3B, 122B and 397B sizes under the MIT license. All benchmark claims are self-reported. By our estimate the 9B needs about 7 GB at 4-bit.
-
Sage-A1.5
Shanghai AI Laboratory's 35B-A3B agent model, built on Qwen3.6 with text and image input, a 256K context and Apache 2.0. It adds parallel sub-agents through a tool and an opt-in self-review mode. Every score comes from the lab, and the official Q4_K_M GGUF is about 22 GB.
-
Nex-N2.6
Nex-AGI's Apache 2.0 agent models for computer use and coding, led by the new N2.6 on GLM-5.3-Flash (321B, FP8) and the N2.5 mini, Pro and Max line up to 1.6T. Scores come from Nex-AGI's own harnesses, and N2.6 does not beat its base on every row.
-
Flint
Recursive's 39.8B mixture of experts with about 2.5B active per token, trained for coding agents and released under Apache 2.0. Recursive asks for at least 48 GB of GPU or unified memory and ships 4-bit builds of 22.6 to 26.0 GB, and says it generates more than 125 tokens per second on an M4 Pro MacBook Pro.
-
Mellum2.1
JetBrains' 12B mixture of experts with 2.5B active parameters, an Apache 2.0 thinking model built to work inside code repositories as a fast sub-agent. In JetBrains' own table it scores 47.0 on SWE-bench Verified, behind Qwen3.5-9B at 50.0, and its 4-bit files are about 7 to 8 GB.
-
Qwen3.8-27B
A 27B dense model under Apache 2.0 with coding and agent skills that fits on one GPU, which makes it a practical model to run a local coding agent on.
-
K2 Horizon
IFM's fully open family, with a 375B-A23B flagship, a 36B-A4B that needs about 22 GB in 4-bit, and a 7B that reports 72.4 on SWE-bench Verified in IFM's own table. In the same table Qwen3.8-27B is well ahead of the dense 32B on Terminal-Bench 2.1.
-
AREX-2
BAAI's 27B dense agent built on Qwen3.8 and trained to improve its answer over many rounds, under Apache 2.0. Its own table reports 81.8 on MLE-Bench Lite and 70.7 on Frontier-CS. By our estimate it needs about 19 GB at 4-bit.
-
Dory 1.0
NVIDIA's 4B looped model: one checkpoint reruns a shared block one to four times, so you trade speed for quality when you serve it. NVIDIA reports SWE-bench Verified scores of 50.3 to 61.8 across the settings. It is about 8 GB in BF16 under OpenMDW-1.1.
-
gpt-oss
OpenAI's Apache 2.0 reasoning models with function calling and tool use. The 117B version fits one 80 GB GPU and the 21B version runs in 16 GB.
-
Mage-VL-Cowork
Microsoft's 4B computer-use agent under Apache 2.0 that chooses between clicking and running shell commands. It reports 66.0% on OSWorld-Verified, the best 4B result in its own table, though a 9B model scores 68.7%. It acts for real, so run it in a sandbox.
-
Needle 3
Cactus Compute's 121M Apache 2.0 model that does one job: turn a request into tool calls or structured data, offline on phones and Raspberry Pi-class boards. Accuracy drops once the tool list passes about two dozen.
-
Clef
Cloudflare's 27B and 9B decision models do not write text. They return a probability for each option you define, as a fast routing and triage step inside an agent. Apache 2.0, and the 9B has a 6.7 GB 4-bit GGUF.
-
Liquid d1
Liquid AI's 0.6B to 3B decision models answer named questions about a state in one forward pass and write no text, which suits routing and triage inside an agent. Liquid measured d1-3B at 8 ms on an RTX 4090 and 50 ms on a Jetson Orin Nano. The LFM Open License is free only below US$10 million of revenue.
-
AutoTrust Decider
AutoTrust's open decision models on Qwen and Gemma backbones, from 9B to 35B, with typed questions, up to 256 options and calibrated probabilities. Decider Pro 35B had about 1.65 million downloads and the vision 27B about 1.55 million in the 30 days to 11 November 2026. The labels were distilled from Typesafe's closed Jev, and the scores are AutoTrust's own.
-
pplx-decider
Perplexity's Apache 2.0 decision models: a 27B and a 4B Flash that answer several typed questions about one message or image in a single pass, with a probability for every option. Perplexity's own table gives v1.2 a Decision Index of 68.09 against 62.30 for v1.1, and the weights are 52.2 GB and 9.1 GB.
-
Smaug-Decisions
Abacus.AI's Apache 2.0 decision model for audio, images and text, a Gemma 4 E4B fine-tune with a 3.45 GB phone build and a 0.7B audio sibling. Its Earshot audio benchmark is its own, and downloads are still tiny, between 95 and 190 in 30 days per repository.
How we chose
- The developer positions the model for coding, agents or tool use, and the weights are open.
- We list design, size and license rather than benchmark rankings, which are self-reported.
- Where a model needs special serving builds or has unclear license files, the entry says so.