Best Open Source LLMs in 2026

Highlights

• +Top open-weights score on the Artificial Analysis Intelligence Index at 46, level with Grok 4.7 (independent)
• +Vals Index 59.47 percent for Pro and 59.58 for Flash, both ahead of DeepSeek V4.1 Flash at 57.86 (independent)
• +DeepSWE v1.1 at 71.9 and Terminal-Bench 2.1 at 89.9, with Flash at 67.9 and 87.6 (vendor-reported)
• +Text, image, video, and audio input in one model, with up to 131,072 output tokens
• +API at $0.435/$0.87 per 1M tokens with cached input at $0.0036; Flash is $0.14/$0.28, both unchanged from V2.5
• +MIT weights on Hugging Face as XiaomiMiMo/MiMo-V2.6-Pro-RL and XiaomiMiMo/MiMo-V2.6-Flash-RL

Best For

The highest-scoring open-weight model on the Artificial Analysis Intelligence Index as of late September 2026, and the pick for omnimodal agents that read video and audio in the same run under plain MIT terms. MiMo-V2.6-Flash, released the same day, scores fractionally higher on the Vals Index (59.58 against 59.47) at about a third of the Pro rate, so test Flash first if cost or cluster size matters.

Highlights

• +A 5.93 GB file against 53.80 GB in FP16, about 9.1x smaller, running on a 16 GB laptop or one 24 GB GPU
• +Every weight is -1, 0, or +1 with one FP16 scale per group of 128, landing at 1.76 effective bits per weight
• +Retains 98.2 percent of the parent model average across 20 benchmarks, 83.9 against 85.4 (vendor-reported)
• +Beats an IQ2_XXS build of the same model by a wide margin: AIME26 95.83 against 78.6, LiveCodeBench v6 90.07 against 70.05
• +142.5 tokens per second on an RTX 5090 and 46.8 on an M5 Max at batch size 1

Best For

Running a genuinely capable multimodal 27B on hardware you already own. This is the clearest local-inference release of the month, and the Apache 2.0 license plus GGUF and MLX builds mean there is no gate between reading about it and running it.

Highlights

• +MIT weights on a 744B agentic model, one of the most permissive licenses at that scale
• +Built for long-horizon research agents: continuous environmental understanding, tool use, and multistep completion
• +Targets carrying scientific work from a paper to executable experiments, reproducible metrics, and an inspectable report
• +Published on Hugging Face and ModelScope with no blog post and no paper on release day
• +The technical report followed on arXiv three days later with more than 140 authors

Best For

Research groups building autonomous experiment loops who need the weights rather than an endpoint, and anyone studying how a large agentic MoE is trained, since the report is unusually detailed for a Chinese lab release.

Highlights

• +First natively multimodal GLM-5: text, image, and video in one stack rather than a bolted-on vision head
• +Plain MIT weights on Hugging Face at zai-org/GLM-5.3-Flash, published on release day
• +DeepSWE 63.4 against GLM-5.2 at 46.2 and AutomationBench 48.8 against 26.2 (vendor-reported)
• +Roughly 3.0x less attention compute and a 4.4x smaller KV cache than text GLM-5.3
• +List pricing of $0.15 input and $0.50 output per 1M tokens, about a tenth of GLM-5.3 on input

Best For

Multimodal coding and computer-use loops at flash pricing, and any team that wants a capable vision-language agent on unambiguous MIT terms at 18B active parameters.

Highlights

• +Only 6B parameters activated per token, among the lowest active counts of any capable open model this year
• +The 51B n-gram table indexes 20 million bigrams and trigrams and can sit in system RAM instead of GPU memory
• +Explicit preview of the architecture that will underpin Qwen4
• +DeepSWE 1.1 at 58.7 and SWE-bench Pro at 62.5, ahead of DeepSeek-V4-Flash-0731 at 54.4 and 56.0 (vendor-reported)
• +Alibaba claims it beats the 397B Qwen3.7-Plus at roughly one ninth the training cost

Best For

Self-hosting shops that want a multimodal agent with a very low active-parameter count, and teams that want an early read on the Qwen4 architecture before the full release.

Highlights

• +Largest open-weight model ever shipped at 2.8 trillion parameters
• +Strongest open-weight GPQA Diamond result on record at 93.5 percent (vendor-reported)
• +Scored 1,679 points on Frontend Code Arena, ahead of Claude Fable 5
• +Native vision alongside text, with a 1M token context window
• +API pricing at $3.00 input (or $0.30 on a cache hit) and $15.00 output per 1M tokens

Best For

Teams that want the closest open-weight model to the proprietary frontier for general knowledge work, science reasoning, and visual tasks, and are content to reach it through a hosted API.

Highlights

• +Only about 8B active parameters, so a small team can genuinely self-host it
• +70.2 percent on Terminal-Bench 2.1 in thinking mode and 78.5 percent on SWE-Bench Multilingual
• +Weights on Hugging Face day one, no waitlist and no delayed drop
• +Hosted pricing of $0.10 input and $0.20 output per 1M tokens, among the cheapest anywhere
• +poolside published every unedited benchmark trajectory and disclosed its own harness

Best For

Cost-sensitive agentic coding at volume, terminal and shell automation, and any team that actually intends to run the weights rather than call someone else's endpoint.

Highlights

• +Plain Apache 2.0, with none of the scale-based carve-outs in the Llama 4 Community License
• +Distilled from Muse Spark and built for agents that run locally without cloud access
• +76.0 on SWE-Bench Verified, 51.2 on SWE-Bench Pro, and 83.5 on GPQA Diamond at high reasoning (Meta-reported)
• +Text and image input, so agents can read screenshots, charts, and documents
• +Meta says 4-bit weights shrink the language model under 20 GB, fitting a 24 GB or 32 GB card with room for the KV cache

Best For

Always-on local agents and private multimodal workflows on a single GPU. It is the most capable open-weight model Meta has published, and the natural upgrade for teams running Llama 4 today.

Highlights

• +Enormous 10M token context window
• +Competitive with GPT-4o on many benchmarks
• +Efficient MoE architecture keeps inference costs low
• +Supports 12 languages natively
• +Multimodal: handles text and images

Best For

Long-context applications, multilingual tasks, and general-purpose use where you need a strong all-around model with exceptional context length.

Highlights

• +Meta's largest open-weight model by total parameters
• +Native 1M token context window
• +Good for complex multi-step tasks
• +Multimodal with strong image understanding

Best For

Teams already standardized on the Llama ecosystem that want a large multimodal MoE with a 1M context window.

Highlights

• +DeepSeek's own comparison puts it ahead of V4 Pro on most agentic and coding rows, though V4 Pro keeps the higher GPQA Diamond score (92.4 to 90.9)
• +GPQA Diamond 90.9, Terminal-Bench 2.1 at 90.6, DeepSWE v1.1 at 74.2 (all vendor-reported)
• +Causal Encoder-Decoder design activates only 8B parameters per token during prefill
• +Native multimodal vision in a 552B open-weight model
• +MIT license with weights on Hugging Face
• +API at $0.30/$1.20 per 1M peak, halving to $0.15/$0.60 off-peak

Best For

The strongest open-weight release of September 2026 and the clearest case that the cheap tier has caught the flagship tier. At 8B active parameters in prefill it is far more deployable than V4 Pro or Kimi K3, and DeepSeek's own table shows it beating the 1.6T V4 Pro on Terminal-Bench 2.1, DeepSWE, and CyberGym. Best pick for agentic coding and terminal work on MIT terms.

Highlights

• +Near-frontier performance: 80.6% on SWE-bench Verified
• +MIT license allows unrestricted commercial use
• +Native 1M token context window
• +Hybrid Attention architecture for better long-context recall
• +Open weights stay downloadable under MIT; the DeepSeek API retired the model on September 14, 2026 and routes deepseek-v4-pro calls to V4.1 Flash

Best For

The largest MIT-licensed model in the DeepSeek line, and still its top GPQA Diamond scorer. Near-frontier coding and reasoning at a fraction of proprietary pricing, for teams that want unrestricted license freedom, which Kimi K3 and Laguna S 2.1 do not offer.

Highlights

• +Open weights on Hugging Face, including a July 2026 refresh (DeepSeek-V4-Flash-0731)
• +Native 1M token context window
• +Strong performance for its active parameter count
• +MIT license, same as V4 Pro
• +Efficient enough to run on smaller GPU setups

Best For

Self-hosted, cost-sensitive workloads where you need 1M context on a budget and already have V4 Flash validated in production. Classification, summarization, and batch processing tasks where V4 Pro is overkill.

Highlights

• +MiniMax Sparse Attention cuts per-token compute at 1M context to roughly 1/20th of M2
• +Claimed 59% on SWE-Bench Pro and 83.5 on BrowseComp
• +Multimodal input: text, image, and video
• +API pricing at $0.30/$1.20 per 1M tokens, among the cheapest at this tier
• +Weights now published on Hugging Face (MiniMaxAI/MiniMax-M3)

Best For

Cost-sensitive agentic coding and browser-agent workloads that need very long context, through the API or on a small cluster now that the weights are downloadable.

Highlights

• +Fully MIT licensed on GitHub and Hugging Face
• +Dynamic activation of 33 to 56 billion parameters per token, adjusting to task difficulty
• +Purpose-built for agentic coding: code understanding, generation, and execution in agent workflows
• +Self-reported 59.5 on SWE-Bench Pro, ahead of Gemini 3.1 Pro, GPT-5.5, and Claude Opus 4.6 by Meituan's measure
• +First trillion-parameter release trained end-to-end on a 50,000-card domestic Chinese compute cluster
• +Pretrained on 30T+ tokens spanning Chinese, English, multilingual, and code data

Best For

Sovereign or geopolitically sensitive deployments where domestic-chip provenance matters, and agentic coding workloads with variable per-token difficulty that benefit from dynamic activation.

Highlights

• +Apache 2.0 weights for a frontier-scale European model
• +Multilingual across dozens of languages, including major European and Asian languages
• +Image input alongside text
• +Native function calling and JSON output
• +Mistral says the FP8 build runs on a single node of B200s or H200s

Best For

European language applications and enterprise use cases where a permissive license and a European vendor both matter. Also strong for tool-using and function-calling applications.

Highlights

• +Only 6.5B parameters active per token, so it serves quickly
• +One model for instruct, reasoning, and coding work, with reasoning effort set per request
• +Text and image input
• +Native function calling and JSON output
• +Apache 2.0 license

Best For

Applications where speed and cost matter more than absolute capability. Great for tool-using agents, classification tasks, and high-throughput workloads.

Highlights

• +Apache 2.0 with no size or usage carve-outs, unusual at this capability level
• +Natively multimodal out of the box: text, images, and hour-scale video
• +Terminal-Bench 2.1 rises from 63.4 to 73.0 against Qwen3.6-27B (vendor-reported)
• +OSWorld-Verified jumps from 63.9 to 84.3 and DeepSWE 1.1 from 13.3 to 42.2
• +Fits on a single high-memory card at 4-bit, so one workstation runs it

Best For

The strongest locally deployable multimodal model near 30 billion parameters. Best pick for teams that want vision and video understanding on their own hardware under a permissive license, without the multi-GPU bill that DeepSeek V4 Pro or Kimi K3 demand.

Highlights

• +Outstanding performance for its small size
• +Strong math and reasoning capabilities
• +Runs on consumer hardware (even laptops)
• +MIT license allows unrestricted use
• +Trained on high-quality synthetic data

Best For

On-device applications, edge computing, and scenarios where you need good reasoning in a small package. Excellent for math-heavy tasks and as a component in larger systems.

Highlights

• +Apache 2.0, replacing the custom Gemma Terms of Use that covered earlier generations
• +Five sizes, from on-device E2B up to a 31B dense model
• +Text and image input on every size, with audio on E2B, E4B, and 12B
• +84.3 percent on GPQA Diamond for the 31B, against 42.4 for Gemma 3 27B (Google-reported)
• +The 26B A4B MoE activates about 4B parameters per token and runs almost as fast as a 4B model

Best For

Fine-tuning experiments, mobile and edge applications, and single-GPU deployments. E2B and E4B target phones and laptops, while the 12B, 26B A4B, and 31B fit consumer GPUs and workstations.

Highlights

• +Fully Apache 2.0 licensed (Cohere's first), unrestricted commercial use
• +Multimodal reasoning: text and image input, native tool use
• +Runs on a single NVIDIA B200 or two H100s at W4A4 quantization
• +Lossless quantization across BF16, FP8, and W4A4 on Hugging Face
• +48 language coverage, with new tokenizer that cuts tokens 16 to 20% in Arabic, Korean, and Japanese
• +Artificial Analysis Intelligence Index of 37; MMMU 75.1%, MathVista 80.6%

Best For

Enterprises that need sovereign, self-hostable AI with native citations and RAG, plus a permissive license for production. Strong fit for agentic workflows across regulated industries (financial services, healthcare, public sector).

Highlights

• +Purpose-built for RAG (Retrieval-Augmented Generation)
• +Excellent at grounding responses in provided documents
• +Strong citation and attribution capabilities
• +Good multilingual support (10+ languages)
• +Reliable tool use and function calling

Best For

RAG applications where you need the model to carefully reference and cite source documents. Enterprise search, knowledge bases, and document Q&A.