Open weights / model directory
149 open-weight models from 10 provider groups, each with parameters, context window, list prices, license, capabilities and official sources. Every model has its own page.
Data checked September 6, 2026 · Derived from the preserved 419-record reference · Download the collection (JSON) ↗
A model is included when the reference records a license that is not proprietary: permissive licenses such as Apache 2.0 and MIT, community licenses such as Meta’s Llama agreements, and non-commercial licenses such as CC-BY-NC. Open weights are not always open source in the strict sense, so read the license before commercial use. Models with no recorded license are listed separately below for verification rather than presented as open.
Open weights or an API? Read the guide ↗ · Tools for running models locally ↗Search by name, size, license or capability. Filters work without an account and can be shared through the page URL.
149 open-weight models · Data checked September 6, 2026 · Not hands-on evaluations
Self-hosting a frontier-ish reasoning model on a single H100 with full chain-of-thought visibility and no per-token bill. OpenAI publishes no hosted…
On-device or low-latency local inference with tool calling; runs in ~16GB. No OpenAI-hosted price is published on the pricing page.
Largest dense open-weight Gemma for self-hosting on server-grade GPUs with a 256K window; hosted access via the Gemini API is free-tier only (the…
Best open-weight throughput-per-dollar in the family — MoE activating only 4B params per token, so it serves near-4B-dense speed at 26B-class…
Encoder-free multimodal mid-size open model with a 256K window — the sweet spot for single-GPU self-hosting; download-only (not exposed on the…
On-device/edge deployment with a 128K window; download-only, with QAT quantized builds for phone and laptop inference engines.
Smallest Gemma 4 — ultra-mobile and embedded inference at 128K context; download-only, no hosted endpoint.
The one Meta open-weight model with a genuinely permissive licence (Apache 2.0, no Community-License user-count or naming clauses) — a 30B dense…
The long-context option in Meta's open-weight line — a 10M-token window and single-H100 deployability at int4 make it the pick for whole-repository…
Higher-quality sibling to Scout with the same 17B activated cost per token but 400B total weights — better reasoning and image understanding, at the…
The best quality-per-GPU text-only Llama: 405B-class instruction following at 70B serving cost, which is why it remains the most widely hosted Llama…
A cross-attention vision adapter bolted onto Llama 3.1 70B — capable at charts, documents and captioning, but image+text is English-only and it is…
Small enough to fine-tune for a single narrow vision task (receipt or form extraction, screenshot classification) on one commodity GPU; not a…
On-device and edge summarisation, rewriting and query expansion where latency and privacy beat raw capability; note Meta's own quantized builds drop…
The smallest Llama, for phone- and microcontroller-class deployment or as a speculative-decoding draft model; expect to fine-tune it for one task…
Now mainly a teacher model: its licence explicitly permits using outputs to train other models, which is the remaining reason to run something this…
Superseded at the same size and cost by Llama 3.3 70B — keep it only for pinned reproducibility of existing 3.1 evaluations or fine-tunes. Weights…
Still the default single-GPU fine-tuning baseline across the open-source ecosystem because tooling support is universal; choose Llama 3.2 3B instead…
The default DeepSeek pick: near-Pro reasoning at a third of the price with the same 1M context, ideal for high-volume agentic coding and…
DeepSeek's frontier tier — worth the 3x premium over Flash only for hard agentic/production coding, deep reasoning and tool-heavy workflows where…
The only DeepSeek model that takes images — use it for screenshot-driven and chart/document-reading agents at identical Flash pricing; it is…
The published base weights behind Qwen3.8-Max and the largest open-weight model in existence — text-only, thinking-mode-only, 262K native context…
The standout self-host pick — a truly Apache-2.0, 27B dense multimodal model with 262K context and adjustable reasoning effort (xhigh/medium/low)…
Only 3B active parameters, so it runs fast on modest hardware while punching well above its weight on coding benchmarks — the sweet spot for local…
Dense 27B for workloads where MoE routing hurts quality consistency; now priced above the newer Qwen3.8-27B ($0.60/$3.60 vs $0.50/$3) — prefer…
The largest fully Apache-2.0 Qwen model — the one to self-host when licence purity matters more than raw scale and the custom-licensed Qwen3.8-2.4T…
Mid-size Apache-2.0 MoE that fits a single 8-GPU node in FP8 — a reasonable step down from the 397B when memory is the constraint.
Cheapest hosted Qwen3.5 open model and an easy local run at 3B active params; superseded on quality by Qwen3.6-35B-A3B for a small price premium.
Dense 27B at the lowest price in the 27B line ($0.30/$2.40) — good value if you don't need the multimodal input that Qwen3.6-27B and Qwen3.8-27B add.
Best coding value in the catalogue — 5x cheaper than qwen3-coder-plus at the base tier and open-weight. Tiers: 0-32K $0.30/$1.50, 32-128K…
The heavyweight open coding model for hard multi-file agentic tasks; expensive and tier-sensitive (0-32K $1.50/$7.50, 32-128K $2.70/$13.50, 128-200K…
Laptop-to-workstation coding model with 3B active params; the practical choice for local IDE autocomplete and small refactors. Tiers: 0-32K…
Very cheap open reasoning at $0.15 input — hybrid-attention architecture with 3B active params makes long-context inference unusually fast per dollar.
Non-thinking sibling for latency-sensitive work where you don't want reasoning tokens billed — same price as the thinking variant, so choose on…
The 2025 open reasoning workhorse, still one of the cheapest large thinking models at $0.23 input; Qwen3.5-397B beats it but needs far more memory…
Excellent bulk-throughput value at $0.23/$0.92 — 2.5x cheaper output than its thinking twin, ideal for summarisation and extraction at scale.
Small MoE reasoner that self-hosts on a single 48GB card in FP8; output pricing is high relative to size, so prefer local deployment over the API…
Fast, cheap non-thinking small MoE — a solid drop-in for classification, routing and tool-dispatch layers.
Original hybrid-thinking Qwen3 flagship; output is $2.80 non-thinking / $8.40 thinking. The -2507 refresh is better and 3x cheaper — migrate.
The most widely deployed open Qwen3 dense model and a well-supported fine-tuning base; output $0.64 non-thinking. Newer 27B models beat it, but…
Original hybrid-mode 30B MoE; output $0.80 non-thinking / $2.40 thinking. Superseded by the -2507 split variants at the same price.
Oddly priced above Qwen3-32B on the API ($0.35 vs $0.16) — only worth it as a self-hosted fine-tune target on a single 24-32GB GPU. Output $1.40…
Edge/consumer-GPU tier — runs quantised on 8-12GB VRAM. Output $0.70 non-thinking / $2.10 thinking. Smaller 4B/1.7B/0.6B siblings exist on Hugging…
Largest open vision-language reasoner — use for chart/diagram reasoning and GUI-agent work where a small VL model fails; output is 2.5x the instruct…
Top open VL model for straightforward description/extraction at $0.40/$1.60 — no reasoning-token overhead.
Cheapest open VL reasoner on the API ($0.16/$0.64) and dense enough to self-host on two 40GB cards — strong default for visual QA pipelines.
Same price as the thinking variant with lower latency — pick this for OCR-adjacent and captioning work that needs no deliberation.
Fast MoE visual reasoner (3B active) — good local throughput, but on the API the 32B-thinking model is cheaper on output ($0.64 vs $2.40).
Best latency-per-dollar open VL model for high-volume image pipelines you host yourself.
Smallest open VL reasoner — the one to fine-tune for a narrow visual domain on a single consumer GPU.
Edge-deployable multimodal model for on-device captioning and screenshot understanding.
The small open any-to-any model people actually self-host for offline voice assistants; output pricing varies by modality and is not stated as a…
Mistral's current flagship for long-horizon agentic work and agentic coding — but note it costs 3x the input and 5x the output of Mistral Large 3,…
The value pick of the whole lineup: frontier-size open-weight MoE at $0.50/$1.50, cheaper than Medium 3.5 and most rivals' mid-tier models — default…
Best price/performance workhorse — a hybrid instruct+reasoning+coding model with only 6.5B active params, so it is fast and cheap while still…
Largest edge-class Ministral — flat $0.20 in and out makes it the cheapest option when your workload is output-heavy; separate Base, Instruct and…
The sweet spot for high-volume classification, extraction and routing where you still want vision and 256K context; also the recommended local model…
Cheapest model Mistral serves and the one to run on-device — pick it for latency-critical edge inference or trivially structured tasks, not for…
Use when you need a 1M-token window on Mistral's EU-hosted infrastructure — it is a third-party Z.ai model served without Mistral modifications,…
Free Labs endpoint for Lean 4 formal proof engineering, autoformalization and automated theorem proving — genuinely $0 while Mistral gathers…
The one Voxtral that does audio *understanding* rather than plain transcription — chat and Q&A over speech via /v1/chat/completions. Text tokens are…
Live/streaming transcription at $0.006 per audio minute — double the batch model's rate, so only use it when you actually need low-latency partial…
Text-to-speech with zero-shot voice cloning (no transcript needed for the voice prompt), 9 languages and ~90ms time-to-first-audio. Billed per…
Compact self-hosted safety classifier: you write policy questions in natural language and it returns yes/no for prompt moderation, response…
Microsoft's safety- and unblocking-post-trained variant of DeepSeek-R1 (671B MoE, 37B active, MIT). It has dropped out of the current Foundry…
The dense 14B workhorse of the family: strongest general Phi quality per dollar for math, code and reasoning when a 16K context is enough — if you…
Cheapest hosted Phi and the default for high-volume classification, extraction and routing across 23 languages with a real 128K window; no native…
The only Phi that takes audio as well as images — good for cheap on-device-class speech understanding, OCR and chart reading in one 5.6B model.…
14B SFT-distilled reasoner at Phi-4 prices — the value pick for math and STEM chains of thought when you cannot justify MAI-Thinking-1 or a frontier…
RL-tuned sibling of Phi-4-reasoning at identical token prices — higher accuracy but noticeably longer chains of thought, so it costs more per answer…
The cheapest reasoning model Microsoft sells: 3.8B with a 128K window in and out, aimed at edge/embedded math tutoring and step-by-step solvers…
Hybrid SambaY/Gated-Memory-Unit architecture giving up to ~10x higher decoding throughput than Phi-4-mini-reasoning on long generations — self-host…
Newest Phi release — a 15B multimodal reasoner with explicit <think> traces for chart/diagram/document math and GUI element localization…
SlimMoE compression of Phi-3.5-MoE down to 7.6B total / 2.4B active — near-Phi-3.5-MoE quality at a third of the memory, but a hard 4K context kills…
Smallest SlimMoE variant (1.1B active) for CPU and NPU inference where every GB counts; 4K context and an Oct-2023 cutoff make it a component model,…
Research GUI-grounding model (Phi-3-V based) that maps a natural-language instruction to on-screen coordinates — a building block for computer-use…
Still deployable and still metered on Azure, but Microsoft has dropped it from the current Microsoft-models doc table and Phi-4-mini-instruct is…
The largest Phi ever shipped (16x3.8B MoE, 6.6B active) — historically interesting and still self-hostable under MIT, but superseded on…
Multi-image and video-frame reasoning at 4.2B; use Phi-4-multimodal-instruct instead unless you specifically need this checkpoint's multi-frame…
Legacy 14B long-context Phi-3; still metered and catalogued but strictly worse and pricier than Phi-4 — kept only for pinned reproducibility.
Short-context variant of Phi-3-medium; no reason to choose it today over Phi-4 at a lower price with a larger window.
Legacy 7B long-context model; superseded on every axis by Phi-4-mini-instruct at half the price.
Short-context legacy 7B; migrate to Phi-4-mini-instruct. (The Foundry catalog record misreports its window as 131072; the model card and name are…
The model that launched the SLM category; still one of the most downloaded Phi checkpoints for offline/edge use, but on Azure Phi-4-mini-instruct is…
The canonical tiny Phi for phones and NPUs (also shipped as GGUF and ONNX DirectML/CUDA/CPU builds); pick it only for offline deployment where the…
Original Phi vision model, managed-compute/self-host only (no serverless per-token meter); superseded by Phi-3.5-vision-instruct and then…
Base (non-instruct) 2.7B research model, still the most-downloaded Microsoft checkpoint on Hugging Face and a common fine-tuning starting point —…
Cohere's flagship: the only model that folds vision, reasoning, agentic tool use and 48-language coverage into one set of weights, and it runs on 1x…
The long-context (256K) dense workhorse for RAG, tool use and 23-language agents, and the only Command A variant with a real 500 req/min production…
Pick when you need an explicit thinking budget on nuanced multi-step or agentic problems in 23 languages and can host it (4x H100 for production);…
A dedicated 23-language machine-translation model for regulated shops that must translate sensitive documents inside their own perimeter; the 8K…
Document/chart/OCR understanding with up to 20 images per request — but it does NOT support tool use, so for agentic multimodal work go to Command…
Still Live with a 500 req/min production limit, but at $2.50/$10.00 it is priced under the 'existing customers' FAQ and comprehensively beaten by…
The cheapest Cohere model with a published price, a real 500 req/min production limit and full tool use — the practical default for high-volume RAG…
Cohere's cheapest served model by a wide margin ($0.0375/1M in) with a 128K window — right for latency-sensitive chatbots, classification and…
Deprecated 2025-09-15 and 3.3x the price of command-r-08-2024 for worse results; the alias `command-r` also points here. No reason to choose it —…
Deprecated 2025-09-15 and the most expensive model Cohere still serves; the alias `command-r-plus` resolves here. Move to command-r-plus-08-2024 for…
Cohere's first agentic coding model: Apache-2.0, free API key plus free weights, and a 3B active footprint that runs locally — aimed at repo-level…
Open (Apache-2.0) 2B Conformer ASR covering 14 languages with a real-time factor up to 3x faster than similar-size models; no timestamps, no…
Arabic-specialised fine-tune of Cohere Transcribe — use it over the base model for any Arabic audio; same 25MB file cap, and it is not yet on…
Research-grade 23-language model with a 128K window at a flat $0.50/$1.50 — the cheapest way to test Cohere-family multilingual quality, but…
Open multilingual vision-language research model across 23 languages; the pricing FAQ covers only Aya Expanse, so its API price is unstated — for…
The balanced 3.35B/70-language Tiny Aya variant — the one to start with when you want broad low-resource language coverage on small hardware (GGUF…
Region-specialised Tiny Aya tuned for West Asian and African languages; choose it over Tiny Aya Global only when your traffic is concentrated in…
Region-specialised Tiny Aya tuned for South Asian languages; worth the swap from Tiny Aya Global for Indic-heavy workloads. No public API price.
Region-specialised Tiny Aya tuned for European and Asia-Pacific languages; the pick for EU/APAC-focused small-model deployments. No public API price.
Frontier long-context agent work where you can hit cache: the 10x cache-hit discount ($0.30 vs $3.00) matters far more than the headline rate, and…
The value pick for autonomous coding agents — roughly a third of K3's input cost and a quarter of its output cost, with 256k context and weights you…
Same weights as kimi-k2.7-code at exactly 2x the price — only worth it when interactive latency, not cost, is the constraint.
Cheapest multimodal option in the Kimi line — pick it over K2.7 Code when you need image/video input and over K3 when 256k context is enough.
Best frontier-class price/context ratio here — 1M context at $1.40 in / $4.40 out, but note reasoning cannot be disabled, so budget for thinking…
The cheapest capable multimodal model in this entire report and plain MIT weights — but the posted rate is a 50%-off promo running to 9 Sep 2026, so…
Same base model and same price as GLM-5.3 but under plain MIT and with thinking optionally off — the one to self-host, or to use when you need…
Cheaper than GLM-5.2/5.3 if 200k context is enough; otherwise the newer siblings are worth the extra 40 cents per million input.
Solid mid-tier coding/agent workhorse at under half GLM-5.2's price; step down to it when you don't need million-token context.
Legacy small model kept alive for existing integrations — new builds should start on GLM-4.7-FlashX or GLM-5.3-Flash instead.
Outstanding cost-per-context: 1M window at $0.30/$1.20 with open weights — just watch the tier break, since anything over 512k input bills at double.
Same price as M3 with a fifth of the context and no vision — only choose it if you have already validated against these exact weights.
A 2x latency surcharge on identical weights — justify it with a measured p95 requirement, not a hunch.
Absurdly cheap 224k-context reasoning (¥0.5/¥2 per 1M = $0.07/$0.30) — the budget choice for bulk long-document work if you can live with a…
Open weights only — Tencent sells no first-party per-token endpoint for it, so budget for GPUs (770B params) or a third-party host; the Apache-2.0…
Apache-2.0, 21B active — the most self-hostable strong model in this report, and the practical Tencent choice when Hy4-preview's 770B is too big for…
Small Apache-2.0 reasoning model that runs on a single modern GPU; not a line item in Qianfan's price table, so treat it as self-host-only.
Long-document grounded QA where you need a 256K window and citations-faithful answers on a budget; the hybrid SSM design keeps long-context cost far…
High-volume RAG and summarisation over long inputs; one of the cheapest 256K-context commercial endpoints available.
Self-hosted enterprise QA where you want 256K context and Apache-2.0 freedom; answers without the token overhead of a reasoning model. Not on AI21's…
On-device RAG on iOS/Android/desktop when you need an unusually large 256K window from a 3B model.
Local reasoning at 3B scale; evaluate against Qwen and LFM2.5 thinking models before committing, since AI21 publishes no hosted endpoint for it.
Cheapest way to bolt image/video understanding onto a high-volume pipeline, and the weights are downloadable if you'd rather run it yourself; check…
The strongest fully-reproducible reasoning model: choose it when you must audit or re-derive the training pipeline, not when you need best-in-class…
General chat variant of the fully-open 32B; Ai2 flags it as intended for research and educational use, so read the Responsible Use Guidelines before…
The base checkpoint to fine-tune when you need documented provenance for every training token (Dolma 3, ~9.3T tokens) plus intermediate checkpoints.
Small fully-open chat model for academic baselines and ablation studies where a licence-clean, data-transparent 7B is the requirement.
Cheap local long-chain-of-thought experiments; pair with the RL-Zero checkpoints if you are studying RL recipes rather than deploying.
Best open choice when you need pointing/grounding output — coordinates and timestamps rather than prose — on short video and multi-image inputs;…
The efficiency-tuned Molmo 2 for single-GPU video grounding when the 8B is too heavy.
The only end-to-end open VLM here whose language backbone is also fully open (Olmo, not Qwen) — pick it when backbone provenance is the point.
Frontier-scale open weights for on-prem agentic workloads — but it needs 8x GB200/B200 or 16x H100 minimum, so it is a datacentre commitment, not a…
The practical sweet spot of the Nemotron line: 12B active params keeps throughput high for high-volume ticket automation and RAG while retaining a…
Newest and most deployable Nemotron: 256K context on a single H100 with only 3B active params, and reasoning can be switched off per-request to cut…
Document intelligence — invoices, forms, charts — at up to 4 images of 2048x1536 per request; a strong self-hosted alternative to paid OCR/VLM APIs.
Apache-2.0 enterprise workhorse with 128K native context (512K extensible) and three effort levels, so you can dial reasoning cost per request;…
The best value in the Granite line for self-hosted agents: 128K context and reliable tool calling on a single mid-range GPU, 12 languages, no…
Smallest Granite 4.2 for CPU or edge deployment where you still want a 128K window; GGUF, nvfp4 and mxfp4 builds ship alongside.
Largest LFM: strong agentic tool use at 1.5B active params and 18.5K tok/s at high concurrency — but Liquid explicitly says it is weak on heavy…
The default on-device model here: 131K context, 16 languages, and competitive with models 4x its size on tool use — just note the bespoke lfm1.0…
Vision on a laptop or NPU: ~3.3GB memory and 228 tok/s on an M5 Max, at the cost of a comparatively short 32K window.
MIT-licensed multimodal reasoner sized for a single GPU; ServiceNow claims ~30% fewer reasoning tokens than Apriel 1.5, which matters more than raw…
Narrow but useful: a small model RL-trained specifically for multi-turn MCP tool calling. Also ships at 4B and 14B. Treat it as a research artifact,…
No models match those filters. Try a broader search or another size.
These 27 records share a family name with an open-weight model, but the reference does not record their license. Check the vendor’s model card before treating any of them as open.
Every record comes from the preserved reference dataset of 419 models, checked against official vendor documentation on September 6, 2026. Parameter counts, output modalities, notes and source links are taken from each model’s archived reference page. Size bands use total parameters; for mixture-of-experts models the active parameter count is shown alongside. Prices are list prices per million tokens where a vendor publishes them; a recorded $0.00 is preserved as recorded and does not promise free access. Nothing here is a hands-on evaluation or a benchmark result.
