Model rankings

Best Open-Source LLMs

The best open-source and open-weight LLMs in 2026 — DeepSeek V4, Kimi K3, MiniMax M3, GLM-5.2, Qwen 3.6, Llama 4, gpt-oss and Gemma 4 ranked by benchmark, licence, context window and cost, with self-hosting and hosted-inference picks for every use case.

Updated August 2026

Quick answer: The best open-source LLM you can run today is DeepSeek V4 — an MIT-licensed model with a 1M-token context that scores 80.6% on SWE-bench Verified, within a fraction of a point of last year’s proprietary leaders, and is free to self-host or use hosted at $0.14/$0.28 per million tokens. The most capable open model is now Moonshot’s Kimi K3 — open-weight since 26 July and the top open model on the independent intelligence index — but at 2.8T parameters (~594GB) it’s a serious-hardware self-host, which keeps DeepSeek V4 the best all-round pick for most teams. MiniMax M3 is the closest practical alternative for high-throughput, long-context work (80.5% Verified, 1M context). GLM-5.2 is the strongest pick for long-horizon agents; Qwen 3.6 and Llama 4 have the widest fine-tuning and tooling support; and Gemma 4 is the best model that runs on a single laptop, and Meta’s new Apache 2.0 Muse Glimmer is the pick for local agents on a 24GB GPU. The one caveat: “open source” is mostly a misnomer here — nearly all of these ship as open weights under custom licences, not fully open-source projects, so read the licence before you build a business on one.

This is an opinionated ranking of models you can download, inspect, fine-tune and run on your own hardware. It is not a scrape of any single leaderboard. We rank by consensus — across benchmarks, licences, real-world testing and our own model pages — and we cite the evidence behind every placement. Where a figure isn’t verifiable from a primary source, we mark it “data not available” rather than invent one. For the full field including proprietary frontier models, see our best AI models ranking; for the coding-specific view, see best AI for coding.

Open source vs open weight: the distinction that matters

Most models marketed as “open source” are not open source in the strict sense. Understanding the difference decides whether you can legally ship a product on one.

Open weight means the trained model weights are published for download, but the training data, training code and full pipeline may not be. Llama, Qwen, DeepSeek, GLM, Kimi and Gemma all sit here. You can run and fine-tune the model, but you cannot fully reproduce it.

Fully open source means weights, training code, training data and an OSI-approved licence are all public. Very few frontier-adjacent models qualify; Allen AI’s OLMo line is the usual reference point for genuinely open training.

The licences vary more than the marketing suggests. MIT (used by DeepSeek V4 and GLM-5.2) and Apache 2.0 (used by gpt-oss, the Mistral 3 family and, since April 2026, Gemma 4) are the cleanest for commercial use. Meta’s Llama Community Licence permits commercial use only up to 700 million monthly active users, above which an explicit Meta agreement is required. The practical rule: if you are building a commercial product, prefer an MIT- or Apache-licensed model, and read any custom licence in full before committing.

For most teams the honest reason to pick an open model isn’t ideology — it’s control. Open weights buy you self-hosting, data residency, offline operation, unlimited fine-tuning, no per-token bill and no vendor able to deprecate the model out from under you. Those are the axes this ranking weighs.

The ranking: best open-source LLMs (August 2026)

These are the strongest models you can download and run, ranked by overall capability and consensus standing, weighted toward repository-scale coding and agentic work (where the most comparable cross-model data exists), then reasoning, context and licence quality. Most of them sit in the “Open” tier of our master best AI models board; here we rank the open-weight field on its own terms — including Kimi K3, which ranks higher (flagship tier) on the master board on raw capability but is open-weight, so it belongs here too.

ᵛ Vendor-reported. Muse Glimmer’s figures come from Meta’s own model card with no independent evaluation yet, which is why a 30B model that beats Qwen 3.6 27B on most agentic benchmarks sits at #9 rather than higher — we treat vendor numbers as a ceiling.

Reading the table. SWE-bench Verified is the older, Python-only coding benchmark still quoted at most launches (llm-stats); we show it where a verified figure exists and mark “data not available” where the model’s own page is still in build rather than invent a number. Context is the maximum published window. Hosted prices are the cheapest first-party or major-host rate we can verify in USD per million tokens; every model here can also be self-hosted, where the only cost is your own compute. Licences are the weight licence, not the code licence — see the licence table below.

One number worth sitting with: DeepSeek V4, MiniMax M3 and Alibaba’s closed Qwen3.7 Max all land within 0.2 points of each other (80.6%, 80.5%, 80.4%) on SWE-bench Verified, and all three are within a fraction of a point of Google’s proprietary Gemini 3.1 Pro. The open frontier has effectively caught last year’s closed leaders, at roughly a tenth of the price.

Segmented verdicts

The single-number ranking hides the fact that “best” depends on the job. Here are the decisive picks.

Best all-round open model: DeepSeek V4

DeepSeek V4 is the model to reach for first. Its MIT licence makes an 80.6%-SWE-bench-Verified model self-hostable outright — no usage caps, no competitive-use clause, no MAU ceiling — and its 1M-token context handles whole repositories and long documents in a single call. Hosted, its DeepSeek V4 Flash tier runs at $0.14/$0.28 per million tokens, roughly a 90x cost gap below a flagship like Claude Opus 5 on output. For most teams it is the best combination of capability, licence and cost available in open weights today.

Highest raw capability: Kimi K3

Kimi K3 is, on the evidence, the most capable open-weight model released — its open weights shipped on 26 July, and it leads every other open model on the independent Artificial Analysis Intelligence Index by a wide margin, passing proprietary flagships like Opus 4.8 on some knowledge-work evals. The catch is deployment: at 2.8 trillion parameters (~594GB) it is the largest open model ever, so self-hosting is a serious-hardware undertaking, and Moonshot has published no standardised SWE-bench figure. If you want the strongest open model and can run it (or call it hosted at $3/$15 per million tokens), it is the capability pick; if you want the strongest open model you can practically self-host, DeepSeek V4 remains the better answer.

Best for high throughput and long context: MiniMax M3

MiniMax M3 sits a hair behind DeepSeek V4 on SWE-bench Verified (80.5%) with the same 1M-token context, and its mixture-of-experts design is tuned for cheap, high-volume serving. Hosted at $0.30/$1.20 per million tokens, it is the pick when you are pushing large batches of long inputs through and want open weights you can also pull in-house later.

Best for long-horizon agents: GLM-5.2

GLM-5.2 is Zhipu’s MIT-licensed model built for multi-step, tool-using agent runs, with a 200K-token context. If you are building an autonomous agent that plans, calls tools and self-corrects over long sessions, it is the strongest open option to trial first — and the permissive licence means you can embed it in a commercial product without negotiation.

Best for fine-tuning and ecosystem support: Qwen 3.6 and Llama 4

For customisation, the deciding factor is how much tooling, documentation and community fine-tuning support already exists. Qwen 3.6 (Apache 2.0) and Llama 4 (Llama Community Licence) win here: both have the broadest coverage in training frameworks, quantisation formats and adapter libraries, so the path from base model to a task-specific fine-tune is the best-trodden. Note that Alibaba’s top-tier Qwen3.7 Max went closed and API-only, so the open Qwen you can self-host is the 3.6 line.

Best for the largest context: Llama 4

Llama 4 Scout reaches a 10M-token context window — the single largest of any model on this list, open or closed. For workloads that genuinely need to hold an entire codebase, book series or archive in context at once, it is the open pick, subject to the Llama Community Licence’s 700M-MAU commercial ceiling.

Best on a laptop or at the edge: Gemma 4

Gemma 4 is the standout for running locally on modest hardware. Its smaller variants are widely rated the most practical model to run on a single laptop or consumer GPU, making it the default for offline, private, on-device use. Gemma 4 also dropped the custom Gemma Terms of Use that constrained earlier generations: since the 2 April 2026 release the weights ship under Apache 2.0, so commercial use, fine-tuning and redistribution carry no royalties, no MAU ceiling and no bespoke agreement with Google. For the hardware side — which variant fits which card — see our best local LLMs guide.

Best open reasoning model: DeepSeek R1 / R2

For explicit chain-of-thought reasoning in open weights, DeepSeek R1 / R2 is the reference — an MIT-licensed reasoning line you can self-host for maths, science and multi-step logic where you want the model to show its working.

Best for a clean commercial licence: gpt-oss and the Mistral 3 family

If licence risk is your first concern, gpt-oss 120b (Apache 2.0) and Mistral Large / Medium 3.5 (Apache 2.0) are the safest picks — permissive, OSI-approved, no MAU ceiling. Mistral additionally suits teams that need EU data residency and a European vendor relationship.

Licence comparison: what you can legally ship

For a commercial product, the licence matters as much as the benchmark. This table summarises the weight licences, not the code repositories.

Bottom line: MIT and Apache 2.0 models (DeepSeek, GLM, gpt-oss, Qwen, Mistral, MiniMax, Gemma 4 and now Meta’s Muse Glimmer) carry the least legal friction. Llama’s licence is fine for the overwhelming majority of products but has a scale trigger worth knowing about — and note that Meta itself has now shipped an Apache 2.0 model, leaving Llama’s MAU ceiling as the older of its two licensing postures. Any vendor-specific “open weights” licence — Kimi’s included — should be read in full before you depend on it commercially.

How to run open-source LLMs

You do not need your own GPUs to use these models. There are two paths: self-host the weights, or call an open model through a hosted-inference provider that has already done the serving for you. The table below is built to be affiliate-ready; we take no commission and links are plain.

Practical guidance. For a laptop or a quick local trial, start with Ollama or LM Studio and a small model like Gemma 4 or a quantised DeepSeek variant. For production self-hosting, vLLM on your own GPUs gives the best throughput and an OpenAI-compatible endpoint. If you want open-model economics without running hardware, a hosted provider (Together, Fireworks, OpenRouter or Groq) lets you call DeepSeek V4 or MiniMax M3 by API. One serving note that materially affects quality: prefer bf16 over fp8 activation quantisation where you can, because aggressive quantisation degrades output on these models. Our companion best local LLMs page covers the hardware-and-deployment side in full — VRAM requirements by model, quantisation formats and what actually runs on consumer machines from an 8GB laptop up to a 128GB unified-memory desktop.

What changed this month

The freshness signal — the open-weight movements that reshaped the board through mid-2026.

Meta came back to open source — under Apache 2.0 (10 August). Meta released Muse Glimmer, a 30B dense multimodal model distilled from the proprietary Muse Spark 1.2 and built for local agentic work, under a plain Apache 2.0 licence — more permissive than any Llama model ever carried, with no MAU ceiling. It runs on a 24GB consumer GPU at 4-bit (~17GB) and beats Qwen 3.6 27B and Gemma 4 31B on most agentic benchmarks on Meta’s numbers, while trailing Qwen on computer use and terminal work. Two things to hold: every figure is vendor-reported, and at 30B it is not competing with the top of this board. The bigger prize is the promise attached to it — Meta says open weights for Muse Spark 1.2 itself are coming. If that lands, it changes this ranking far more than Glimmer does.

The open frontier caught the closed leaders. DeepSeek V4 (MIT), MiniMax M3 and Alibaba’s Qwen3.7 Max now all sit within ~0.2 points of Gemini 3.1 Pro on SWE-bench Verified, at roughly a tenth of the cost. The gap between open and proprietary that used to be 40–60 points wide is now a handful.

Kimi K3’s open weights shipped (26 July). Moonshot released the weights for Kimi K3 — the largest open model ever at 2.8T parameters and, on the independent intelligence index, the new raw-capability leader among open models. Its size makes it a serious-hardware deploy, so DeepSeek V4 stays the most practical open default, but the open capability ceiling moved up overnight.

Qwen’s top tier went closed — then started to reopen. Alibaba took Qwen3.7 Max to API-only, so the strongest Qwen was no longer downloadable and the open Qwen you can self-host became the 3.6 line. That reversed with the 3 August launch of Qwen3.8-Max, which Alibaba says will release open weights — the full 2.4T model plus a smaller Qwen3.8-27B for consumer hardware. As of this update those weights aren’t out yet, so the strongest downloadable Qwen remains the 3.6 line — a reminder that “open” status can change release to release, and worth checking before you standardise on a model.

Licences kept converging on the permissive end. The clear trend across 2025–2026 has been makers moving to MIT and Apache 2.0: DeepSeek to MIT with V3/R1, Qwen to Apache 2.0 with the Qwen3 generation, and new entrants such as gpt-oss and the Mistral 3 family launching directly under Apache 2.0. Gemma 4 joined them in April 2026, moving from Google’s custom Gemma Terms of Use to Apache 2.0. For commercial builders, open-weight licence risk is lower than it was a year ago — Llama’s MAU ceiling and the vendor-specific Kimi licences are now the main exceptions to read carefully.

Self-hosting got easier. Tooling like Ollama, LM Studio and vLLM has matured to the point where running a capable open model locally or in production is a matter of hours, not weeks — lowering the practical barrier that used to keep teams on closed APIs.

How we rank

Our methodology is deliberately plural, because no single source is trustworthy on its own.

We start with benchmarks, read critically: we quote SWE-bench Verified where a verified figure exists, label it as the older Python-only measure it is, and mark “data not available” rather than guess where a model’s page is still in build. We weight memorisation-resistant and unsaturated evals over ones the top models have already saturated.

We layer in licence quality, because an open model you cannot legally ship is not useful, and real-world testing — how a model behaves on long agentic runs, whether it flags its own uncertainty, and how it holds up outside a benchmark’s distribution. And we weight operational reality: context window, hosted price, and how much tooling exists to run and fine-tune the model.

Every figure on this page is either cited to a primary source or linked to our own model pages, where the detail lives. Where a number can’t be verified, we write “data not available”. This ranking is re-scored monthly; the date at the top is the last full pass. The honest limitation: several open-weight models here have capability that is well-established in community testing but not yet captured in a standardised, verified benchmark figure — we mark those cells rather than fill them with vendor claims.

Frequently asked questions

What is the best open-source LLM in 2026?

The best open-source LLM you can run today is DeepSeek V4 — MIT-licensed, 1M-token context, and 80.6% on SWE-bench Verified, within a fraction of a point of last year’s proprietary leaders. It is free to self-host and cheap to run hosted ($0.14/$0.28 per million tokens). MiniMax M3 is the closest alternative for high-throughput, long-context work, and GLM-5.2 is the strongest pick for long-horizon agents.

Is there a difference between open source and open weight?

Yes, and it matters commercially. Open weight means the model’s trained weights are downloadable, but its training data and pipeline may not be — this covers Llama, Qwen, DeepSeek, GLM, Kimi and Gemma. Fully open source means weights, training code, training data and an OSI-approved licence are all public, which very few frontier-adjacent models are (Allen AI’s OLMo is the usual example). Most models marketed as “open source” are technically open weight.

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

For most routine work, close. DeepSeek V4, MiniMax M3 and Alibaba’s Qwen line sit within ~0.2 points of Google’s proprietary Gemini 3.1 Pro on SWE-bench Verified, at roughly a tenth of the cost. Proprietary flagships like Claude Opus 5 still lead on the hardest tasks and on long-horizon agentic reliability, but paired with a strong harness an open model handles the bulk of everyday tasks.

Which open-source LLM has the best licence for commercial use?

MIT and Apache 2.0 models carry the least legal friction. DeepSeek V4 and GLM-5.2 are MIT; gpt-oss, Qwen 3.6, Gemma 4 and Mistral Large / Medium 3.5 are Apache 2.0 — all permit unrestricted commercial use. Llama 4’s Community Licence is the main exception, permitting commercial use only up to 700 million monthly active users.

Can I run an open-source LLM on my own computer?

Yes. For a laptop or single GPU, install Ollama or LM Studio and pull a smaller model such as Gemma 4 or a quantised DeepSeek variant. For production self-hosting on your own GPUs, vLLM gives high throughput and an OpenAI-compatible API. If you want open-model economics without owning hardware, hosted providers such as Together AI, Fireworks AI, OpenRouter and Groq serve open models by API.

What is the best open-source LLM for coding?

DeepSeek V4 leads the open field on SWE-bench Verified (80.6%) and is the default open coding model, with Kimi K2.6 strong for agentic coding workflows and Qwen 3.6 well supported for fine-tuning on your own codebase. For the complete picture including proprietary tools and agents, see our best AI for coding guide.

What is the best open-source LLM for agents?

GLM-5.2 is the strongest open pick for long-horizon, tool-using agents, with Kimi K2.6 a close alternative for agentic and coding tasks. Both are permissively licensed enough to embed in a commercial agent product; trial both on your specific tool-calling workload before committing.

Which open-source LLM has the largest context window?

Llama 4 Scout has the largest published context window of any model on this list at up to 10 million tokens. DeepSeek V4 and MiniMax M3 both offer 1 million tokens, which is enough for whole-repository or long-document work for most teams.

This ranking is re-scored monthly and updated as new open-weight models ship and benchmarks evolve. Benchmark scores vary by harness; SWE-bench Verified is Python-only and quoted at launch, and we mark cells “data not available” rather than publish unverified figures. Licence, pricing and availability current as of 4 August 2026 and subject to change — always confirm a model’s licence with its maker before commercial use.