best 15 open-weight llms

zhipu's launch material claims 'pure open' with 'no regional limits', and the licence file is standard unmodified mit text with a 2026 copyright line. no revenue cap, no field-of-use clause, no geographic carve-out, no attribution mandate beyond keeping the notice. of the fourteen models here, this is one of the few where the marketing and the legal text say the same thing.

it also currently sits at the top of the artificial analysis intelligence index among open-weight models, having displaced kimi k2.6. it is a 753b sparse mixture-of-experts with roughly 40b active, a full one-million-token context, and it is built for long-horizon agentic and coding work rather than chat.

running it yourself is a datacentre exercise: community estimates put fp8 at around 750gb of gpu memory — an 8x h200 node — and full fp16 near 1.6tb, with int4 bringing it to roughly 400gb. most teams will rent it, where the price is identical at fireworks, baseten, novita and together.

the licence is plain mit, fetched from the repository rather than inferred: no revenue cap, no field-of-use restriction, no attribution requirement beyond the copyright notice. for a model reported at 80.6% on swe-bench verified and ahead of gpt-5.5 on codeforces and livecodebench, that is a remarkable thing to be able to write.

the architecture is a 1.6-trillion-parameter mixture-of-experts with 49b active and a one-million-token context. where it lags is breadth rather than depth — tool-use ecosystem, multimodality and very long agentic loops still favour the closed frontier models.

the caveat worth pausing on: this has been labelled a preview since april, and the stable build deepseek targeted for mid-july has not appeared in its changelog. that is a real consideration for anything going into production on downloaded weights, and it is why this sits second rather than first.

mistral has shipped flagships under its own research-and-commercial split licence before, so apache 2.0 here was not a foregone conclusion. the model card links the standard apache text directly and adds no custom terms — the announcement's 'all models released under apache 2.0' holds up against the repository.

at 675b total and 41b active with a 256k context, it is reported as the top open-source coding model on lmarena and second overall among open models at roughly 1418 elo, behind only gemini among all models at publication. the ministral 3 siblings at 14b, 8b and 3b carry the same licence, which makes this the most complete permissive family here.

two things temper it. december 2025 is a long time ago in this field, and mistral's ceo confirmed a larger sparse successor in early access in july that has not shipped. and 675b total means full-precision self-hosting is an organisational decision, not an individual one.

the modification to mit is a single scale-triggered clause: if your product passes 100 million monthly active users or $20 million in monthly revenue, you must display 'kimi k2.7 code' prominently in its interface. below that it behaves as plain mit — no revenue share, no field-of-use ban, no geographic limit. it is one of the milder terms in this category, and it is buried in the licence file rather than in any announcement.

the model is a 1t-parameter mixture-of-experts with 32b active across 384 experts and a 256k context, tuned for long-horizon real-world software engineering. moonshot reports it using roughly 30% fewer thinking tokens than k2.6 for comparable completion, which matters when you're paying per token.

it trails the closed frontier on raw scores — 62.0 against gpt-5.5's 69.0 on kimi's own code benchmark, 35.1 against claude opus's 42.8 on ml-research tasks — and sits at position 31 on the artificial analysis index. native int4 quantisation reportedly brings deployment down to around two a100 80gb cards, which is the most approachable footprint of any trillion-parameter model here.

released on 15 july, this is the first from-scratch open-weight model from mira murati's thinking machines lab: 975b total with 41b active, trained on 45 trillion tokens spanning text, image, audio and video, with up to a million tokens of context from the downloaded weights.

it is explicitly designed as a customisation base rather than a finished product, paired with the vendor's tinker fine-tuning platform, and its stated design goals include resistance to censorship. the model card's own admission that it is 'not the strongest overall model available today, open or closed' — swe-bench verified at 77.6% against roughly 95% for leading closed models — is the kind of candour this category could use more of.

one verification caveat we'd rather state than hide: apache 2.0 is confirmed from thinking machines' own model card, not from a raw licence file pulled out of the repository, which fell outside our budget. treat it as provisionally confirmed and check the repo before committing.

openai's open-weight release carries unmodified apache 2.0. the only restrictions are apache's own patent-litigation termination clause and ordinary trademark limits; a separate usage policy exists but is explicitly not part of the licence and imposes nothing contractual. after the licences elsewhere in this category, that plainness is worth something.

the practical case is the footprint. 117b total with only 5.1b active per token means it fits a single 80gb gpu at native mxfp4 — roughly 60gb of weights with room left for kv cache — and the small active path makes cpu and consumer-gpu inference viable at lower throughput. nothing else here with real agentic and tool-calling ability runs on one card.

the age shows. released august 2025 with no successor, it is text-only while the 2026 field has gone multimodal, and its 128k context is a quarter of what the models above it offer. it has been passed on the intelligence index by inkling, gemma 4 and deepseek v4.

worth stating clearly because comparison tables get it wrong: v4 flash is a separate repository and a separate release at 284b total with 13b active, not a quantisation or distillation of v4 pro's 1.6t. it carries the same plain mit licence.

the economics are the story. hosted at fireworks it runs $0.14 in and $0.28 out per million tokens against v4 pro's $1.74 and $3.48 — a twelfth the cost — and it scores 40.3% on the artificial analysis intelligence index, which for a model of this size and price is a strong showing.

self-hosting is meaningfully more achievable than the pro tier too, with 13b active rather than 49b. if your workload is high-volume and doesn't need the last increment of reasoning, this is the entry on this page most likely to save you real money.

gemma 1 through 3 shipped under google's own gemma terms of use: a prohibited-use policy, redistribution and notice obligations, downstream pass-through requirements and termination on breach. widely called 'open', never actually open source. gemma 4's hugging face card carries a plain apache-2.0 tag, and google's announcement says the same. that is a genuine and underreported change.

the family is designed for local deployment: e2b and e4b variants handle text, image and audio on edge hardware, the 12b targets laptops, and the 31b dense and 26b sparse variants sit comfortably on one 24-48gb consumer card when quantised. 256k context and 140+ languages.

two loose ends we could not close. the repositories remain gated behind an accept-terms click-through, and google's old gemma terms page is still live — we could not confirm whether that click-through still invokes it or whether the page is simply stale. and the training data stops at january 2025, which shows on anything recent.

the licence file is standard unmodified apache 2.0 with a 2026 alibaba cloud copyright and no appendix. at 35b total with just 3b active, alibaba describes it as performing on par with models ten times its active size, reporting 73.4 on swe-bench and 51.5 on terminal-bench 2.0 — vendor figures, not independently verified.

the confusion worth avoiding: qwen3.7-max, announced in may as the flagship, has no hugging face repository, no weights and no committed release date. it is api-only through alibaba cloud and resold through openrouter and together. roundups that place '3.7' alongside the open qwen line are describing two different products, and only one of them is open.

the honest limit is capacity. 35b total is a fraction of the frontier models above, so raw knowledge breadth and long-tail reasoning trail them. what you get instead is something that runs on a single high-end card.

the licence file linked from the hugging face card is the linux foundation's openmdw-1.1: no revenue cap, no field-of-use limit, no naming mandate, with an ip-litigation termination clause and responsibility on you to clear third-party rights. read alone, that is reasonably permissive.

but nvidia's own legal-agreements pages describe nemotron models as governed by a custom nvidia nemotron open model licence, effective december 2025, which adds a naming mandate on derivatives and a broad user-indemnification clause that openmdw contains nothing like. we could not determine whether this model was quietly relicensed or the card is mislabelled, and nvidia's materials are inconsistent on it.

the model itself is strong — 550b total with 55b active and a one-million-token context, built for long-context agentic reasoning and reliable tool-calling. it needs a multi-node cluster to serve, so the audience is exactly the kind of organisation that will notice an unresolved licence.

apache 2.0 on the weights, and alongside them the dolma 3 training corpus, the dolci post-training datasets, the training and evaluation code, and the intermediate checkpoints — all linked from the model card. if you need to audit what went into a model, or reproduce it, this is the only real option on this page.

the family is 7b and 32b dense, each in base, instruct, think and rlzero variants, which makes the reasoning-focused ones genuinely useful for research into how these models learn. the 7b runs on a 16-24gb consumer card and the 32b on a single 48-80gb card.

the trade is capability. a 32b dense model does not compete with 675b and 1.6t sparse rivals, the 32k context is the shortest here, and it did not feature in july's intelligence index coverage. that is the honest cost of full transparency at a non-profit's budget.

microsoft has kept every phi-4 variant under plain mit, from the 3.8b mini through the 14b reasoning models to this 15b reasoning-vision release from march. no field-of-use terms, no attribution beyond the copyright line. consistency across a seven-model family is rarer than it should be.

the appeal is deployment reach: the 15b fits a single 24gb consumer card in bf16 or int8, and the mini variants run on cpu or entry-level edge hardware through ollama or foundry local. the reasoning was distilled from o3-mini traces and now extends to vision-grounded tasks.

the 16,384-token context is the problem. against 256k and 1m elsewhere on this page, that rules out long documents, large codebases and most agentic loops. this is a model for bounded, well-specified tasks close to the device, and it is good at those.

the llama 4 community licence carries four things worth knowing. any product above 700 million monthly active users needs a separate licence from meta. you must display 'built with llama', and any derivative model's name must begin with 'llama'. meta's acceptable use policy is incorporated by reference. and your rights terminate automatically if you bring ip litigation against meta over the model or its outputs.

the benchmark history compounds it. meta submitted a differently-tuned, unreleased experimental checkpoint to lmarena in april 2025 that scored far above the weights it actually published — yann lecun acknowledged in january 2026 that 'results were fudged a little bit'. trackers now disagree wildly, with one aggregator placing it 205th of 215.

the architecture — 400b total, 17b active, 1m context, multimodal — was competitive when it landed in april 2025. it now clearly trails deepseek v4 and qwen 3.6 on coding, and everything above it here has a cleaner licence.

the sequence matters. minimax announced 'minimax open sources m2.7' on 12 april, and its previous two models, m2 and m2.5, shipped under standard mit — so the reasonable expectation was more of the same. shortly after release the licence file was revised: commercial use now requires minimax's prior written authorisation, requested by email.

'commercial use' is defined broadly enough to catch almost anything — any paid product or service, commercial api use, and commercially deployed fine-tunes or derivatives. even authorised users must display 'built with minimax m2.7'. a use-restriction appendix additionally bans military applications, illegal content, disinformation and harassment. the hugging face community flagged it immediately as not osi-compliant, and it isn't.

which is a shame, because the model is good: 230b total with 10b active, around 200k context, 56.22% on swe-pro matching gpt-5.3-codex, and 57.0% on terminal bench 2. it sits at position 46 on the intelligence index. if the terms were mit it would rank in the top half of this page.

moonshot launched k3 on 16 july through its app and api, describing it as the largest open-weight model ever released at roughly 2.8 trillion parameters with a one-million-token context and native multimodal input. the hugging face repository exists as an upcoming-release countdown; on the day we published this, it held no weights and no licence file.

moonshot's own recent history is the reason not to assume what the terms will be. kimi k2.7 code shipped as 'modified mit', and the modification — an obligation to display the model's name in your product's interface above 100 million monthly users or $20 million in monthly revenue — appears in the licence file and in almost none of the coverage. whatever k3 arrives with should be read in the repository rather than taken from launch day.

everything else about it is currently unverifiable. there is no public leaderboard entry because there are no weights to evaluate, at least one outlet has flagged an undisclosed hallucination risk pending independent testing, and the circulating hardware figures — around 594gb for the native mxfp4 download and roughly 1.4tb of fast memory at 4-bit — are press estimates with no release to check them against. we will re-rank this entry properly once the files land.