Best Open-Source LLM 2026: We Benchmarked 8, GLM-5.2 Takes the Top Spot

Last updated: July 19, 2026. Every benchmark score, VRAM figure, and license in this guide was re-verified for this update, and the lineup changed hard: 5 of the 8 models here are new since March, only Llama 4 Scout, Phi-4, and Mistral Large 3 carried over unchanged. If a new release changes the order, this page updates within the month.

The best open-source LLM in July 2026 is GLM-5.2, and this isn't a leaderboard call from the sidelines, we ran it as our production coding model for three weeks before writing this update. Z.ai's 744-billion-parameter MoE scores 91.2% on GPQA Diamond and 62.1% on SWE-bench Pro, at a fraction of frontier API pricing. Kimi K3, Moonshot's 2.8-trillion-parameter release from July 16, already edges past Claude Opus 4.8 on Artificial Analysis's independent ranking, though its weights don't ship until July 27. DeepSeek V4 still wins on price, Qwen3.6 still wins on license freedom, and Gemma 4 12B now beats last year's 27B flagship at less than half the memory footprint.

Leading Open-Source LLMs: Benchmark Snapshot

The table below covers five widely-deployed open-source models scored on current public benchmarks. GPQA Diamond measures graduate-level science reasoning, the benchmark most 2026 leaderboards lean on now that MMLU is largely saturated at the frontier.

Google reports MMLU Pro instead of GPQA Diamond for Gemma 4, 77.2%, beating last year's Gemma 3 27B at 67.6% despite less than half the parameters. We re-verify this table every month; this pass was done July 19, 2026.

Open-source models don't just compete with closed frontier systems anymore, on agentic coding and reasoning, some are winning outright. Kimi K3's independent Artificial Analysis ranking now sits ahead of Claude Opus 4.8, and that's before its weights are even public.

This ranking cuts through the hype. Every model here is evaluated on benchmarks that matter, hardware you'll actually need, and the specific use case where each one shines brightest.

Quick Summary: Which Open-Source LLM Should You Pick?

Need the strongest all-around model? GLM-5.2, we ran it three weeks in production. Chasing closed-frontier reasoning and don't mind waiting a week? Kimi K3, once its weights land July 27. Want the cheapest tokens at scale? DeepSeek V4 Flash. Running on a single GPU? Gemma 4 12B or Qwen3.6. Need massive context? Llama 4 Scout's 10M tokens is still unmatched among models you can download today.

The VRAM numbers assume Q4KM quantization (parameters x 0.57 GB/param), cross-checked against our LLM VRAM requirements master table. Full precision (FP16) needs roughly 3.5x more. New to running models locally? Start with Gemma 4 12B or Qwen3.6, they're the most hardware-friendly options on this list.

Open-Source LLM Leaderboard: July 2026 Rankings

As of July 19, 2026, GLM-5.2 tops our open-source LLM leaderboard for agentic coding and reasoning, the same model we ran in production for three weeks. Kimi K3 is a close second on raw capability, but its weights aren't public until July 27, so treat it as pending rather than deployable today. DeepSeek V4 and Qwen3.6 round out the frontier-adjacent tier. The full ranking below covers all eight models scored in this guide, from multi-node clusters down to laptop-friendly picks.

These rankings reflect our monthly re-check of benchmarks, hardware needs, and license terms. This cycle folded in three releases that didn't exist when we last published: GLM-5.2, Kimi K3, and DeepSeek V4.

Now let's break down what makes each model worth your attention.

1. GLM-5.2, Best Overall Open-Source LLM

Zhipu/Z.ai's GLM-5.2 rolled out to Coding Plan subscribers June 13, 2026, with open weights following three days later. It's a Mixture-of-Experts model, 744 billion total parameters with roughly 40 billion active per token, and it carries a 1-million-token context window plus a 131K max output.

We didn't just read GLM-5.2's spec sheet, we ran it. For three weeks we pointed Claude Code at Z.ai's Anthropic-compatible endpoint (api.z.ai/api/anthropic, model string GLM-5.2[1m]) as our primary coding model on real client repos. A normal week burned about 1,300 of our roughly 2,000 weekly Pro-tier prompts; two migration-heavy weeks hit the cap on day 5. It cleared a Next.js 16 API-route refactor across roughly a dozen files without a stumble. Full teardown, quota math and all, is in our GLM 5.2 Coding Plan review.

Benchmark highlights:

Those SWE-bench Pro and Terminal-Bench scores put GLM-5.2 within striking distance of Claude Opus 4.8 on real coding work, at a fraction of the API cost. Z.ai's own list pricing runs roughly $1.40 per million input tokens and $4.40 output, though third-party providers median closer to $0.55 in and $1.85 out.

Hardware requirements: At Q4KM, the full model needs roughly 424 GB of VRAM (744B x 0.57 GB/param), an 8x 80GB GPU node or a multi-node cluster, per our VRAM requirements master table. Not something you spin up on a gaming rig. Most teams run it through the GLM Coding Plan or a hosted API instead of self-hosting.

Who should use it: Teams doing agentic coding, long client-repo refactors, or anything where you'd otherwise reach for Claude Code or Codex. If your workload is mostly coding agents, this is the model we'd point you at first, because it's the one we actually depend on. For a side-by-side against the other two Chinese open-weight leaders, see our Qwen vs DeepSeek vs GLM comparison.

2. Kimi K3, Best for Long-Horizon Agents (Weights Land July 27)

Moonshot AI's Kimi K3 launched July 16, 2026, as the largest open-weight model released to date: 2.8 trillion total parameters, with only 16 of 896 experts active per token. It reads a million tokens of context, keeps reasoning switched on by default, and accepts text, image, and video input.

The launch numbers are real, and they're mixed. K3 debuted at no. 1 on Arena's Frontend Code leaderboard at 1,679 Elo, a 17-place jump from Kimi K2.6, ranking above Claude Fable 5. Independent Artificial Analysis places it 4th of 189 ranked models, trailing only Claude Fable 5 and GPT-5.6 Sol, and edging past Claude Opus 4.8. That's an open-weight model out-ranking a current-generation Claude Opus-tier model on an independent leaderboard.

Benchmark highlights:

One caveat worth repeating: Moonshot scored K3 inside its own KimiCode environment, versus Claude Code for Fable 5 and Codex for GPT-5.6 Sol. The scaffolding around a model moves its score, so read these numbers as directional. Our full Kimi K3 review breaks down every benchmark, including where K3 actually loses, DeepSWE and HLE-Full.

Hardware requirements: Here's the catch: you can't self-host K3 yet. Moonshot promises open weights July 27, 2026, ten days after launch. Once they land, expect roughly 1,596 GB of VRAM at Q4KM (2.8T x 0.57 GB/param), well beyond anything but a large multi-node cluster, even quantized. Until then, it's API-only at $0.30 cache-hit or $3.00 fresh input, and $15.00 output per million tokens, the priciest model any Chinese lab has shipped.

Who should use it: Nobody self-hosting today, the weights simply aren't out. Once they are, agentic research teams and long-horizon coding shops chasing frontier-adjacent performance without a Claude or OpenAI contract are the target buyer. If you need something you can download and run right now, skip to GLM-5.2 or DeepSeek V4 below.

3. DeepSeek V4, Best Price-Performance

DeepSeek V4 launched as a preview April 24, 2026, in two tiers: V4 Pro for reasoning and agentic work, and V4 Flash for fast, cheap, high-volume calls. Reported architecture is a 1.6-trillion-parameter MoE with 49B active for Pro, and 284B total with 13B active for Flash, both MIT-licensed with an official 1M-token context window.

The NIST CAISI evaluation of V4 Pro is the closest thing to an independent government-grade check this model has, worth reading if you're deploying it somewhere regulated. On benchmarks, V4 Pro posts 90.1% on GPQA Diamond, keeping pace with Kimi K2.6's 90.5%, and a reported ~80.6% on SWE-bench Verified, though that figure used an unverified test-time scaffold, so treat it cautiously until a third party reproduces it.

Benchmark highlights:

DeepSeek V4 Flash is the one to reach for on cost: $0.14 per million input tokens on a cache miss, but just $0.0028 on a cache hit, versus $0.28 output. For a RAG pipeline that re-reads the same context repeatedly, that's the cheapest option on this entire list by a wide margin.

Hardware requirements: V4 Pro needs roughly 912 GB of VRAM at Q4KM (1.6T x 0.57 GB/param), a dozen 80GB GPUs at minimum. V4 Flash is far lighter at roughly 162 GB, within reach of a small multi-GPU box. Practically, most teams call the API rather than self-host either tier.

Who should use it: Cost-sensitive, API-first products, especially anything with heavy repeated context. One deadline to flag: the older deepseek-chat and deepseek-reasoner endpoints retire 2026-07-24 at 15:59 UTC, five days after this update goes live. Migrate to deepseek-v4-pro or deepseek-v4-flash now if you're still pointed at the old endpoints. For the full three-way against Qwen and GLM, see our Qwen vs DeepSeek vs GLM breakdown.

4. Qwen3.6, Best for Self-Hosting

Alibaba shipped Qwen3.6 in two waves: the 35B-A3B Mixture-of-Experts variant on April 16, 2026, followed by a fully-dense Qwen3.6-27B on April 22. Both ship under Apache 2.0, the most permissive license on this list, and both stay in the family Alibaba has kept lightest to self-host.

The headline stat: Qwen3.6-35B-A3B posts 73.4% on SWE-bench Verified while activating just 3 billion parameters per token. That's a coding-capable model that runs at the inference speed of something a fraction of its size. The coding-focused variant, Qwen3-Coder-Next (80B-A3B), separately posts 70.6-71.3% SWE-bench Verified across evaluation setups, self-reported and state-of-the-art among open models without test-time scaling.

Benchmark highlights:

Alibaba's own closed, API-only Qwen3.7 Max preview scores higher still (56.6 on Artificial Analysis's Intelligence Index), but it isn't open-weight, don't confuse it with the models you can actually download.

Hardware requirements: At Q4KM, the dense Qwen3.6-27B needs roughly 15 GB of VRAM (27B x 0.57 GB/param), comfortable on a single 16 or 24 GB consumer card. The 35B-A3B MoE variant needs closer to 20 GB, for the same reason every MoE model does, all experts have to stay resident even though only 3B activate per token. For the full formula and a comparison against every other model in this guide, see our LLM VRAM requirements master table.

Who should use it: Anyone self-hosting on a single GPU who still wants frontier-adjacent coding performance, plus anyone who specifically needs Apache 2.0 over MIT for redistribution or patent terms. It's also the strongest pick here for multilingual work, a category the coding-only comparisons tend to skip. Pair it with the best tools for running LLMs locally if you're setting up a self-hosted box for the first time.

5. Llama 4 Scout, Best for Long Context

Meta's Llama 4 Scout is still the long-context specialist fifteen months after launch, and still the only model on this list you can download today with a 10-million-token context window. Meta hasn't shipped a Llama 5 despite claims circulating online this spring, Llama 4 (Scout and Maverick) remains the current generation as of this update.

Scout uses 16 MoE experts with only 2 active per token, giving it 109B total parameters but just 17B active. Meta claims it fits on a single H100 with INT4 quantization, though the 10M context window obviously demands far more memory for the KV cache on top of the weights.

Benchmark highlights:

That perfect Needle-in-a-Haystack score at 10M tokens still hasn't been matched by anything newer on this list. Most models start losing information well before 128K; Scout's inter-document attention masking keeps boundaries intact even at massive scale.

Hardware requirements: At Q4KM, the model weights need roughly 62 GB of VRAM (109B x 0.57 GB/param), a single H100 or A100 (80 GB) handles it comfortably. That figure is higher than what we published here back in March, we'd underestimated it; the corrected number comes from our VRAM requirements master table, cross-checked against Meta's model card. The catch remains the KV cache: actually using the full 10M context requires enormous memory on top of the weights, so most self-hosted deployments realistically cap at 128K-256K tokens.

Who should use it: Teams working with massive documents, legal analysis, code repository Q&A, or research paper synthesis, where nothing else on this list gets close to the same context ceiling.

6. Gemma 4 12B, Best for Single-GPU Deployment

Google DeepMind's Gemma 4 12B shipped June 3, 2026, and immediately made last year's Gemma 3 27B look oversized. At 11.95 billion parameters, dense, Apache 2.0, it scores 77.2% on MMLU Pro against Gemma 3 27B's 67.6%, beating a flagship more than twice its size.

It's an encoder-free, unified multimodal model: text, image, and native audio go in as raw input, no separate audio encoder bolted on. That design choice is exactly why it stays small enough to run on a mid-range consumer card. It carries a 256K-token context window and supports 140 languages.

Benchmark highlights:

Hardware requirements: At Q4KM, Gemma 4 12B needs about 6.8 GB of VRAM, fitting comfortably on an 8 GB card like the RTX 4060, per the official model card. One command gets you running: ollama run gemma4:12b (7.6 GB download). For the full spec sheet, sampling settings, and setup walkthrough, read our Gemma 4 12B deep dive.

Who should use it: Solo developers, small teams, and anyone who wants a capable multimodal model without renting cloud GPUs. It's also the model we'd recommend for prototyping, test your pipeline here, then scale to GLM-5.2 or DeepSeek V4 in production if you need more headroom.

7. Phi-4 Reasoning, Best for Edge and Resource-Constrained Deployment

Microsoft's Phi-4 Reasoning still proves that parameter count isn't everything. At just 14 billion parameters, it outperforms DeepSeek R1's old 70B distill on several reasoning benchmarks. Phi-4-reasoning-vision-15B, released March 2026, adds multimodal capabilities, scoring 17% higher than Gemma 3 12B on math-visual reasoning tasks.

The Phi-4 family's approach hasn't changed: curated, high-quality training data over raw scale. It still works. Phi-4 scores over 80% on MATH and MGSM, outperforming Gemini Pro and GPT-4o-mini class models on mathematical reasoning, even as the frontier above it has moved on.

Benchmark highlights:

Hardware requirements: Around 8 GB of VRAM at Q4KM (14B x 0.57 GB/param), a laptop GPU or an older desktop card handles it. The model runs comfortably on Apple Silicon MacBooks, making it genuinely portable, and it's one of the few models on this list that runs on-device with reasonable latency.

Who should use it: Mobile and edge developers, teams building on-device AI features, and anyone who needs strong reasoning on minimal hardware. Phi-4 is also a good pick for evaluating fine-tuned models, its small size makes training iterations fast and cheap. The MIT license means no commercial restrictions.

8. Mistral Large 3, Best for Multilingual and Enterprise Use

Mistral Large 3 remains Mistral's flagship going into the second half of 2026. At 675 billion total parameters (41B active), it's a full MoE model under Apache 2.0, a shift from Mistral Large 2's restrictive research license that's held up well over the past several months.

The model was trained from scratch on 3,000 NVIDIA H200 GPUs. On the LMSYS Chatbot Arena it still ranks no. 2 among open models. What keeps it relevant for European teams is multilingual strength, roughly 85.5% accuracy on a balanced multilingual MMLU test, an axis most of the newer coding-focused releases above don't even report.

Benchmark highlights:

One trait that still sets Mistral models apart: they're trained to say "I don't know" rather than hallucinate. That makes Mistral Large 3 a fit for RAG pipelines and enterprise applications where factual accuracy matters more than leaderboard bragging rights.

Hardware requirements: With 675B total parameters, Mistral Large 3 needs roughly 385 GB of VRAM at Q4KM (675B x 0.57 GB/param), corrected from the lower figure we published in March; five 80GB GPUs is the realistic minimum. Mistral also offers a lighter Small 4 variant (119B total, ~6B active) for teams that don't need the full flagship.

Who should use it: European companies that need multilingual capabilities and data sovereignty, and enterprise teams building RAG systems where hallucination reduction is critical. The Apache 2.0 license removes commercial friction entirely.

Best Open-Source LLM for Coding (July 2026)

For coding you can run today, GLM-5.2 is the top open-source pick: 62.1% on SWE-bench Pro and 81.0% on Terminal-Bench 2.1, close enough to Claude Opus 4.8 that our three-week production test genuinely replaced a chunk of our Claude Max usage. Want the model with the single strongest coding benchmark, full stop? Kimi K3 leads SWE Marathon and topped Arena's Frontend Code leaderboard above Claude Fable 5, but you'll wait until July 27 for the weights. On a single consumer GPU, Qwen3.6-35B-A3B's 73.4% SWE-bench Verified at 3B active parameters is the best ratio of capability to hardware on this list.

Curious how the three Chinese labs actually stack up against each other on coding specifically? Our Qwen vs DeepSeek vs GLM comparison runs the full benchmark table with sourcing notes on every number.

How to Choose: Decision Framework

The "best" model depends entirely on your constraints. Use this matrix to narrow your decision.

Still unsure? Start here: do you need something running today? If yes, and you have multi-GPU or cloud budget, GLM-5.2 or DeepSeek V4. On a single consumer card, Gemma 4, Qwen3.6, or Phi-4. Need 10M-token context? Llama 4 Scout. Willing to wait eight days for the strongest open weights available? Kimi K3.

How We Ranked These Models

Rankings aren't based on a single benchmark. Each model was evaluated across five dimensions:

1. Benchmark performance, GPQA Diamond, SWE-bench, AIME, and domain-specific agentic tests
2. Hardware accessibility, Can real teams actually deploy it, and are the weights even public yet?
3. License freedom, Apache 2.0 and MIT score higher than restrictive licenses
4. Ecosystem maturity, Available on Hugging Face, Ollama, vLLM, and major inference engines
5. Real-world adoption, Active community, production deployments, ongoing updates, including our own three-week GLM-5.2 test

Models that score well on benchmarks but require a GPU cluster nobody has get penalized, and models whose weights aren't public yet (Kimi K3, for now) get ranked on capability but flagged clearly on availability. Models with restrictive licenses that block commercial use lose points too. The goal is practical value, not leaderboard bragging rights.

This pass was re-verified July 19, 2026. If you want to test these models yourself, our LLM evaluation guide walks through setting up systematic benchmarks, and the best evaluation tools roundup covers the frameworks you'll need.

FAQ

What are the newest open-source LLMs in 2026?

The current open-weight frontier is led by GLM-5.2 (Zhipu/Z.ai), Kimi K3 (Moonshot, weights due July 27), DeepSeek V4, and Qwen3.6 (Alibaba). Gemma 4 12B (Google DeepMind) is the newest single-GPU pick, and Llama 4 Scout, Phi-4, and Mistral Large 3 round out the list. We re-check this leaderboard monthly; this pass replaced 5 of the 8 models from our March ranking.

How often is this open-source LLM leaderboard updated?

Monthly. We re-verify benchmark scores, VRAM requirements, and licenses at the start of each cycle and add new models as they ship. The "Last updated" date under the title reflects the most recent review.

What is the best open-source LLM in 2026?

GLM-5.2 currently leads on the broadest combination of benchmarks and real-world usability, we ran it as our production coding model for three weeks. Kimi K3 scores higher on some independent rankings but isn't self-hostable until its weights ship July 27, 2026. For pure price, DeepSeek V4 Flash wins.

Can I run open-source LLMs on consumer hardware?

Yes. Gemma 4 12B runs on an 8 GB card like the RTX 4060, and Qwen3.6-27B fits a single 16-24 GB consumer GPU. Phi-4 14B runs comfortably on a laptop GPU or Apple Silicon. Quantized versions (Q4, Q8) of larger models also work on multi-GPU consumer setups using tools like Ollama and llama.cpp.

Are open-source LLMs as good as ChatGPT or Claude?

On agentic coding specifically, yes, closer than ever: Kimi K3 edges past Claude Opus 4.8 on Artificial Analysis's independent ranking, and GLM-5.2 sits within striking distance of Opus on SWE-bench Pro. The gap widens on the hardest frontier reasoning (Claude Fable 5, GPT-5.6 Sol still lead there) and in creative writing and nuanced instruction following.

What does MoE (Mixture-of-Experts) mean for hardware?

MoE models have a large total parameter count but only activate a fraction per token. However, the entire model must be loaded into memory so the router can select experts. GLM-5.2, for example, is a 744B-parameter MoE with only ~40B active, but you still need VRAM for all 744B, you don't save memory, you save compute.

Which open-source LLM is best for coding?

GLM-5.2 leads among models you can run today, at 62.1% SWE-bench Pro and 81.0% Terminal-Bench 2.1. Kimi K3 posts stronger launch numbers on some agentic coding benchmarks but isn't downloadable until July 27. For coding on a single consumer GPU, Qwen3.6-35B-A3B's 73.4% SWE-bench Verified at 3B active parameters is the best ratio of capability to hardware.

Is Llama 4 Scout really 10 million tokens of context?

The architecture supports it, and Meta demonstrated perfect Needle-in-a-Haystack retrieval at 10M tokens. But actually using the full context requires enormous KV cache memory. Most self-hosted deployments realistically cap at 128K-256K tokens. Cloud providers like Together AI and Fireworks offer longer context windows.

What license should I look for in an open-source LLM?

Apache 2.0 and MIT are the most permissive, full commercial use, modification, and redistribution. Llama's custom license allows commercial use but has restrictions for apps with 700M+ users. Some newer Chinese releases ship under a custom "community license" with commercial-use conditions rather than a standard permissive license, always read the actual terms before production deployment.

How much VRAM do I need for a 70B model?

At Q4KM, a 70B dense model needs roughly 40 GB of VRAM (70B x 0.57 GB/param). A single A100 (80 GB) handles it easily, or two RTX 4090s (48 GB total). At full precision (FP16), you're looking at roughly 140 GB, effectively two A100s or four RTX 4090s.

Can I fine-tune open-source LLMs for my use case?

Absolutely. QLoRA makes fine-tuning a 70B model possible on a single 24 GB GPU. Smaller models like Phi-4 and Gemma 4 are even more accessible for fine-tuning experiments. Apache 2.0 and MIT licensed models have no restrictions on derivative works.

What about open-source multimodal LLMs?

Gemma 4 12B handles text, image, and native audio input with no separate encoder. Kimi K3 adds video on top of text and image, once its weights ship. Llama 4 Scout and Phi-4-reasoning-vision-15B both handle text and image, with Scout supporting up to 20 hours of video input within its massive context window.

Is Kimi K3 really open source if the weights aren't out yet?

Not yet, and that distinction matters. Moonshot calls K3 "open-weight," but as of this update you can only reach it through the API, there's nothing to download, self-host, or independently verify. Moonshot has committed to publishing the weights July 27, 2026. Until then, treat every K3 benchmark as vendor-reported and directional rather than something you can reproduce yourself.

Is there an open source LLM leaderboard?

Yes. Our July 2026 leaderboard above ranks all eight models in this guide, Artificial Analysis maintains an independent cross-lab ranking, and Hugging Face hosts the community-run Open LLM Leaderboard for broader coverage. We re-verify our rankings monthly against benchmarks like GPQA Diamond, SWE-bench, AIME, and Terminal-Bench.

Choosing a tool is the easy half. Getting it to run reliably inside a real product is where most teams stall, and that is exactly what our AI integration team builds for clients, from RAG pipelines to custom agents. Want a second opinion on your stack? Get a free consultation.

Sources

• GLM-5.2 - Hugging Face
• Z.AI Developer Docs
• Kimi K3 Launch Coverage - VentureBeat
• DeepSeek API Pricing - Official
• CAISI Evaluation of DeepSeek V4 Pro - NIST
• Qwen3.6-27B - Hugging Face
• Gemma 4 12B Official Launch - Google
• Meta Llama 4 Official Page
• Mistral Large 3 Announcement
• Phi-4 Reasoning Technical Report - Microsoft Research
• Artificial Analysis - Independent Model Rankings
• Open LLM Leaderboard - Hugging Face