The Best Open Source LLMs (2026): Ranked by Benchmark, Size, and Use Case

The Best Open Source LLMs, Ranked

This page ranks the best open-weight LLMs, models with downloadable weights you can self-host. What the open source label does and does not mean for these licenses is covered in open weight vs open source below. If you want the overall coding ranking including proprietary frontier models (Claude, GPT, Gemini), see best LLM for coding. Below is the ranked open-weight shortlist for general and coding use, each with the verified benchmark that justifies the pick. Sources are each model's Hugging Face card or maker announcement.

1. Kimi K3 (Moonshot AI). Best overall. 88.3 on Terminal-Bench 2.1, 81.2 on FrontierSWE, 67.5 on DeepSWE, 93.5 on GPQA Diamond, 94.5 on MCPMark-Verified. Mixture-of-Experts, 2.8T total with 104B active (16 of 896 experts routed per token), 1M context, native text and image input. Weights published July 27, 2026 under the custom Kimi K3 License.
2. DeepSeek-V4.1-Flash (DeepSeek-AI). Best MIT model on agentic benchmarks. DeepSeek's own table reports 90.6 on Terminal-Bench 2.1, 74.2 on DeepSWE v1.1, 30.0 on Terminal-Bench 3.0, and a 3471 Codeforces rating. 552B total parameters in a causal encoder-decoder layout that activates 8B on prefill and 16B on decode, 1M context, native image input, MIT. Weights and technical report published September 10, 2026. The scores are vendor-reported and were measured in DeepSeek's own harness, so treat the ordering against Kimi K3 as provisional until an independent run lands.
3. GLM-5.3 (Z.ai). Best coding upgrade on an existing base. Same 753B base as GLM-5.2 with scaled post-training: DeepSWE v1.1 up from 46.2 to 66.9, Terminal-Bench 3.0 from 4.6 to 28.3, 88.2 on Terminal-Bench 2.1 in DeepSeek's cross-vendor table. 1M context. Weights published August 28, 2026 under the GLM-5.3 License, which is MIT-like except that hosts above $10B in trailing revenue must pass a Z.ai security review before commercial use. GLM-5.3 details.
4. GLM-5.3-Flash (Z.ai). Best plain-MIT release of the season. 320B total / 18B active, natively multimodal, hybrid linear-plus-sparse attention, 1M context, unmodified MIT License, published August 26, 2026. Z.ai reports it beats GLM-5.2 across its benchmark set at about a tenth of the serving price. GLM 5.3 Flash on Morph.
5. GLM-5.2 (Z.ai). Best MIT-licensed model on SWE-bench Pro. 62.1% on SWE-bench Pro, 81.0 on Terminal-Bench 2.1 under Terminus-2, 46.2 on DeepSWE, 99.2 on AIME 2026. 753B parameters, 1M context, standard MIT License. NIST's CAISI assessment called it probably the most capable open-weight model at its release; GLM-5.3 supersedes it on agentic coding but not on license terms.
6. DeepSeek-V4-Pro (DeepSeek-AI). Best MIT generalist for long context. 80.6% on SWE-bench Verified, 67.9 on Terminal-Bench 2.0, 51.8 on Toolathlon. 1.6T total / 49B active, 1M context, MIT. The 284B / 13B DeepSeek-V4-Flash sibling shares the license and the 1M window at a fraction of the serving cost.
7. MiniMax M3 (MiniMax AI). Best cheap agentic MoE. 59.0% SWE-bench Pro, 66.0 Terminal-Bench 2.1, 74.2 MCP Atlas, 34.8 SWE-fficiency. 428B total / 23B active, 1M context through MiniMax Sparse Attention, native image and video input. Ships under the MiniMax Community License, which is non-commercial by default.
8. Qwen3.8-2.4T-A95B (Alibaba). Largest Qwen release yet. 2.4T total with 95B active, 262K native context extensible to about 1.01M, weights published August 8, 2026 under the custom Qwen3.8-Max license rather than Apache-2.0. Qwen's launch table scores the hosted Qwen3.8-Max build at 86.6 on Terminal-Bench 2.1, 67.7% on SWE-bench Pro and 73.5 on FrontierSWE. Model card.
9. Kimi K2.7 Code (Moonshot AI). Coding-tuned K2.6. 1T total / 32B active (8 of 384 experts per token), 256K context, Modified MIT. Moonshot reports 62.0 on its in-house Kimi Code Bench v2 (K2.6 scored 50.9) with about 30% fewer thinking tokens than K2.6. Moonshot's own table has it behind GPT-5.5 and Opus 4.8 on Kimi Code Bench v2 and Program Bench, yet ahead of Opus 4.8 on MCPMark-Verified, 81.1 to 76.4. Model card.
10. Nemotron 3 Ultra (NVIDIA). Best-documented frontier-scale release. 550B total / 55B active in a hybrid LatentMoE design that interleaves Mamba-2 and MoE layers with a few attention layers, plus multi-token prediction. Context defaults to 256K and reaches 1M with a config change. Weights are under the OpenMDW-1.1 license. Its coding scores sit mid-pack (70.7% SWE-bench Verified, 56.4 Terminal-Bench 2.1), but the model card itemizes the training data, down to a purchased set of Terminal-Bench-style tasks from Mercor. Model card.
11. Qwen3.8-Flash-Next (Alibaba). Best active-parameter efficiency. 125B total with 6B active, plus 51B of n-gram embedding parameters and a 4B multi-token-prediction head, and a Qwen Sparse Attention layer that selects micro-blocks of tokens. Published August 24, 2026 under the Qwen Community License 1.0, it reports 62.5% SWE-bench Pro, 58.7 DeepSWE 1.1 and 81.0 SWE-bench Multilingual, ahead of the dense Qwen3.8-27B on all three. Model card.
12. Qwen3.8-27B (Alibaba). Best single-GPU model. 61.7% on SWE-bench Pro and 73.0 on Terminal-Bench 2.1 from a dense 27B under Apache-2.0, weights published August 5, 2026, with 262K native context extensible to 1M. Qwen3.6-27B scored 53.5% SWE-bench Pro in the same table, so the gain at equal size is 8.2 points. Qwen also lists Muse Glimmer-30B at 51.2% SWE-bench Pro and 51.7 Terminal-Bench 2.1 for comparison. Model card.
13. Qwen3.5-397B-A17B (Alibaba). Best large Apache-2.0 model. 76.2% SWE-bench Verified, 397B total / 17B active, released February 2026. The permissive-license choice when you want a big MoE without custom terms.
14. Gemma 4 (Google DeepMind). Best model family for edge and laptops. Five sizes (E2B, E4B, 12B, 26B-A4B, 31B), Apache-2.0, 128K context on the edge models and 256K on the rest, multimodal with audio on E2B, E4B, and 12B. Google reports the 31B ranking third among open models on the Arena text leaderboard.
15. gpt-oss-120b (OpenAI). Best competition coder. 2622 Elo on Codeforces, matching or exceeding o4-mini; 16.2 SWE-bench Pro. 117B total / 5.1B active, 128K context, Apache-2.0.

Comparison Table: Size, License, Benchmark

Every model side by side: total parameters, license, context window, the single benchmark it leads on, and the category it wins. Active-parameter counts (the driver of inference cost) are noted where the model is Mixture-of-Experts.

Context windows converged at the top. Kimi K3, GLM-5.3 and GLM-5.3-Flash, GLM-5.2, all three DeepSeek-V4 variants including V4.1-Flash, and MiniMax M3 ship 1M-token windows. Gemma 4 spans 128K on E2B and E4B up to 256K on the 12B, 26B-A4B, and 31B sizes. Qwen3.6-27B is 131K, Qwen3.5-397B-A17B is 262K, and gpt-oss-120b remains at 128K. Longer context costs more memory per request, so pick the smallest window that fits your repository.

Best for Coding

For agentic coding, rank by SWE-bench Verified and its harder successors. Verified is close to saturated among frontier open models: DeepSeek-V4-Pro resolves 80.6% and Qwen3.6-27B resolves 77.2%, so the discriminating benchmarks are now SWE-bench Pro, FrontierSWE, DeepSWE v1.1, and Terminal-Bench 2.1 and 3.0. Kimi K3 leads the independently reported set at 81.2 FrontierSWE and 88.3 Terminal-Bench 2.1. DeepSeek-V4.1-Flash reports 90.6 Terminal-Bench 2.1 and 74.2 DeepSWE v1.1 in its September 10 model card, which would put it first if reproduced. GLM-5.3 moved the 753B base from 46.2 to 66.9 on DeepSWE v1.1, and GLM-5.2 is still the best MIT-licensed option at 62.1% SWE-bench Pro.

The size story flipped this year. Qwen3.8-27B, a dense 27B, resolves 61.7% of SWE-bench Pro, ahead of every 400B-plus open model from 2025, and Qwen3.8-Flash-Next reaches 62.5% with only 6B parameters active per token. If your agent loop is latency-bound rather than accuracy-bound, a 27B-class model is now a serious primary, not a fallback.

For the coding-specific deep dive, including how these models pair with a fast-apply edit model, see Best Open Source Coding Model 2026 and Best AI Model for Coding.

Best Small / Local (<=27B)

The leaders above need multiple high-memory GPUs. For a model that runs on a single GPU or a workstation, the pick is now Qwen3.8-27B, a dense 27B released August 5, 2026 under Apache-2.0 with 262K native context (extensible to 1M), 61.7% on SWE-bench Pro and 73.0 on Terminal-Bench 2.1. It replaces Qwen3.6-27B (April 22, 2026; 77.2% SWE-bench Verified, 53.5% SWE-bench Pro), which remains a fallback with a 131K window.

Dense matters here. Qwen3.8-27B activates all 27B parameters per token, which is simpler to serve than a Mixture-of-Experts model and predictable on a single device. If you would rather trade a little accuracy for much lower per-token cost, Qwen3.6-35B-A3B activates 3B parameters and still resolves 73.4% of SWE-bench Verified.

Gemma 4 is the alternative family, and it is the one that scales down furthest: E2B and E4B run on phones, 12B on a laptop, 26B-A4B and 31B on a consumer GPU or workstation. Gemma 4 moved to Apache-2.0, so the license caveat that applied to Gemma 3 is gone. Qwen's own comparison table puts Gemma4-31B at 52.0% SWE-bench Verified, well behind Qwen3.6-27B on coding, so pick Gemma 4 for multimodal and edge deployment rather than for agentic coding.

On Morph's fleet, qwen38-27b and gemma4-31b run alongside the large MoE models, so a router can drop routine turns to a 27B-class model and reserve the frontier models for hard work.

Best Reasoning

For step-by-step reasoning, Kimi K3 and GLM-5.2 lead the open weights. Kimi K3 scores 93.5 on GPQA Diamond and 91.2 on BrowseComp. GLM-5.2 reaches 99.2 on AIME 2026 and 91.2 on GPQA Diamond, and ships under the standard MIT License, which permits commercial use and distillation for training other models.

The MIT permission on distillation is why so many small reasoning models are trained on outputs from DeepSeek and Z.ai models: the license allows it. If you need an open reasoning model whose outputs you can legally build on, GLM-5.2 or DeepSeek-V4 is the answer. Kimi K3's license is more restrictive, so read it before you train on its outputs.

For competition coding specifically, gpt-oss-120b reaches 2622 Elo on Codeforces, matching or exceeding OpenAI o4-mini, under Apache-2.0. It is smaller (117B total / 5.1B active) and cheaper to serve than R1 while leading on algorithmic contests.

Best for Agentic / Tool Use

Agentic workloads (multi-file edits, run-fix loops, terminal commands, MCP tools) reward models tuned for tool calling and long action sequences. Kimi K3 leads the tool-use benchmarks outright: 94.5 on MCPMark-Verified, 84.8 on OSWorld-Verified, 88.3 on Terminal-Bench 2.1. MiniMax M3 is the value pick at 74.2 MCP Atlas and 66.0 Terminal-Bench 2.1 while activating only 23B of 428B parameters per token.

Throughput is the practical bottleneck for agentic loops, where a single task fires dozens of tool calls. A faster open model that resolves the issue in fewer, faster turns often beats a marginally more accurate but slower one. Morph serves the low-active-parameter models (minimax3-428b, dsv4flash, qwen38-27b) precisely for that loop, and offers a separate latency SKU of Kimi K3 (kimik3-fast) for cases where decode speed matters more than price.

Open Weight vs Open Source, in Short

Every model ranked above is open-weight, not OSI open-source: the parameters ship, the training data does not. The models that meet the strict standard (OLMo, Pythia) are not the ones topping these benchmarks.

The OSI Open Source AI Definition 1.0 asks for the weights, the training code, and enough information about the training data to rebuild the model. Most checkpoints on this list ship the weights under Apache 2.0, MIT, or a custom community license, so read the license file on the Hugging Face card before you put one in a commercial product.

Tradeoffs: License and Hardware

Several of the strongest models carry licenses that restrict deployment. State the downside plainly before you build on them.

The MiniMax Community License is the sharpest edge on this list. M3 is free for non-commercial use, but commercial use requires displaying "Built with MiniMax M3" and notifying MiniMax, and above $20M in yearly revenue it requires prior written authorization. Cohere Command A is CC-BY-NC, so it is a research model rather than a production one. Kimi K3 permits commercial deployment but adds obligations aimed at large companies and AI service providers, which is the category most inference platforms fall into.

Hardware is the other tradeoff. Self-hosting Kimi K3 (2.8T total) or DeepSeek-V4-Pro (1.6T) is a real infrastructure project: multiple high-memory GPUs plus a serving stack such as vLLM or SGLang, and a 1M-token window multiplies the KV-cache memory per request. Running the best open model is not free even when the weights are. The cheaper path is a hosted inference API that amortizes the hardware across many users.

How to Run the Best Open Source LLM

Three deployment paths, by model size. Small dense models run locally; large MoE models need a serving stack or a hosted API.

Best Open Source LLM to Run Locally

For a laptop or a single consumer GPU, the practical ceiling is a 27B to 31B dense model in 4-bit quantization, which needs roughly 16 to 20 GB of VRAM for weights plus room for the KV cache. Qwen3.6-27B and Gemma 4 31B are the two picks that fit that budget, and Gemma 4 12B is the realistic option below 12 GB. Ollama pulls quantized builds of each. The per-model VRAM table, quantization tradeoffs, and pull commands live on best Ollama models, so this page keeps to the ranking.

A router is the efficient default when you want the best open model per task without owning the GPUs. It classifies each prompt and sends easy turns to a cheap small model and hard turns to a large one, which is how a 27B-class model and a 200B+ model coexist behind a single API. Morph's router classifies prompt difficulty in ~430ms into four tiers for 40-70% API cost savings. See What Is an LLM Router.

import OpenAI from "openai"

// Morph's API is OpenAI-compatible; point the SDK at it.
const client = new OpenAI({
  apiKey: process.env.MORPH_API_KEY,
  baseURL: "https://api.morphllm.com/v1",
})

// Served open-weight models on the fleet:
//   morph-kimik3, morph-glm53-744b, morph-glm53flash,
//   morph-dsv41flash, morph-dsv4flash
const res = await client.chat.completions.create({
  model: "morph-glm53-744b",
  messages: [{ role: "user", content: "Refactor this module to use dependency injection." }],
})

console.log(res.choices[0].message.content)

Previous Generation: Qwen3-Coder, DeepSeek-R1, Kimi K2, GLM-4.6

Most open-weight capacity in production today is one generation behind the list above. Weights do not expire, clusters are provisioned for a specific model, and a migration costs an evaluation cycle. These are the models that still hold that installed base, with the size, license, and context window each of them actually ships.

Qwen3-Coder-480B-A35B is the one worth keeping deliberately. Alibaba shipped it as a 480B Mixture-of-Experts with 35B active parameters and a 256K native window that extrapolates to 1M, under Apache-2.0. Independent SWE-bench Verified runs score it at 66.5% and SWE-bench Pro at 38.7%, roughly 14 points of Verified behind DeepSeek-V4-Pro, on a license with no conditions attached and a serving footprint a third the size of the current frontier.

DeepSeek-R1 is the reasoning model the rest of the field was distilled from. 671B total, 37B active, 128K context, MIT, and the license text states outright that derivative works and distillation for training other LLMs are permitted. DeepSeek-V3.2 shares the architecture. Both are still the cheapest way to get a large MIT-licensed reasoner onto your own hardware.

Kimi K2 is 1T total with 32B active over 384 experts, 128K context, Modified MIT. Moonshot reported 65.8 on SWE-bench Verified and 66.1 on Tau2-Bench without extended thinking. GLM-4.6 is 355B-A32B under plain MIT and was the release where Z.ai took the context window from 128K to 200K. MiniMax-M2 preceded M3 with 230B total and 10B active under a Modified MIT license, the lowest active-parameter count of that generation.

When to migrate: agentic work over hour-scale sessions, where Terminal-Bench 2.1 and SWE-Marathon separate Kimi K3 from everything above by more than 20 points. When to stay: single-issue coding, fill-in-the-middle completion, and reasoning at fixed cost per token. Morph runs both generations on the same fleet, so a router can send a routine turn to a previous-generation model and a hard turn to the frontier without changing the endpoint.

Frequently Asked Questions

What is the best open source LLM right now?

As of October 2026, Kimi K3: 88.3 on Terminal-Bench 2.1, 81.2 on FrontierSWE, 2.8T total parameters with 104B active, 1M context, weights published July 27, 2026. DeepSeek-V4.1-Flash (September 10, 2026, MIT) reports 90.6 Terminal-Bench 2.1 and 74.2 DeepSWE v1.1 from 552B total with 8B to 16B active, GLM-5.3-Flash (320B / 18B active) is Z.ai's strongest plain-MIT release, and Qwen3.8-27B reaches 61.7% SWE-bench Pro from a dense 27B under Apache-2.0.

What is the best open source LLM for coding?

Kimi K3 leads independently measured open-weight coding at 88.3 Terminal-Bench 2.1 and 81.2 FrontierSWE. DeepSeek-V4.1-Flash reports 90.6 Terminal-Bench 2.1 and 74.2 DeepSWE v1.1 in its own harness under MIT. GLM-5.3 lifts the 753B GLM-5.2 base to 66.9 DeepSWE v1.1, and GLM-5.2 is still the strongest MIT-licensed coder on SWE-bench Pro at 62.1%. For a model that fits on one GPU, Qwen3.8-27B scores 61.7% SWE-bench Pro and 73.0 Terminal-Bench 2.1 under Apache-2.0.

What is the best small or local open source LLM?

Qwen3.8-27B, a dense 27B under Apache-2.0 with 262K native context, 61.7% SWE-bench Pro and 73.0 Terminal-Bench 2.1, published August 5, 2026. Gemma 4 is the alternative and now ships under Apache-2.0 as well, in five sizes from E2B to 31B with 128K to 256K context, which covers phones and laptops as well as workstations.

Are open source LLMs as good as GPT-5 or Claude?

On coding benchmarks the gap has closed. NIST's CAISI evaluation of GLM-5.2 found it was probably the most capable open-weight model at its release and similar to GPT-5.2 overall. Kimi K3 scores 93.5 on GPQA Diamond and 88.3 on Terminal-Bench 2.1. Closed frontier models still lead on the hardest reasoning sets, but the margin is small and open weights win on cost and control.

Which open source LLM license is best for commercial use?

MIT and Apache-2.0. GLM-5.2 and both DeepSeek-V4 variants are MIT. Qwen3.5, Qwen3.6, Qwen3.8-27B, Gemma 4, and gpt-oss-120b are Apache-2.0; the larger Qwen3.8 releases use custom Qwen licenses. Read the terms first on Nemotron 3 Ultra (OpenMDW-1.1), Kimi K3 (custom license with extra obligations for large companies and AI service providers), MiniMax M3 (non-commercial by default), Llama 4, and Cohere Command A (CC-BY-NC forbids commercial use).

How do I run the best open source LLM?

Models at or below 31B (Qwen3.8-27B, Gemma 4) run on a single high-memory GPU via vLLM, SGLang, Ollama, or llama.cpp. The frontier MoE models (Kimi K3 2.8T, DeepSeek-V4-Pro 1.6T, GLM-5.2 753B, MiniMax M3 428B) need a multi-GPU cluster. To skip the hardware, use a hosted inference API: Morph serves kimik3, glm52-744b, minimax3-428b, dsv4flash, gemma4-31b, and qwen38-27b through one OpenAI-compatible endpoint.

Is Qwen3-Coder, DeepSeek-R1, or Kimi K2 still worth running?

Yes, and most self-hosted open-weight capacity still runs on them. Qwen3-Coder-480B-A35B is 480B total with 35B active, 256K native context, Apache-2.0, scored at 66.5% SWE-bench Verified and 38.7% SWE-bench Pro in independent runs. DeepSeek-R1 is 671B / 37B active at 128K under MIT, which explicitly permits distillation for training other models. Kimi K2 is 1T / 32B active at 128K under a Modified MIT license, reported at 65.8 SWE-bench Verified without extended thinking. GLM-4.6 is 355B-A32B at 200K under plain MIT. All four sit a generation behind Kimi K3, GLM-5.2, and DeepSeek-V4 on agentic benchmarks, and all four cost less to serve.

Why are most open source LLMs actually open weight?

The OSI Open Source AI Definition requires the weights, the full training and inference code, and detailed training-data information. Almost no benchmark-topping model releases all three; they release weights only, making them open-weight. The practical difference is licensing freedom: open-weight models under Apache-2.0 or MIT can still be self-hosted, inspected, fine-tuned, and used commercially.

Related Resources

• Best Open Source Coding Model 2026
• Best AI Model for Coding
• What Is an LLM Router?
• LLM Gateway
• LLM Inference Optimization
• LLM Cost Optimization
• LLM Context Window