Every open-weight family we track, with sizes, memory at Q4_K_M and a page per variant. Start from your VRAM, not from the leaderboard.
Written by Jakub Rusinowski · Last updated
The strongest models that fit each amount of VRAM or unified memory, with the build they were sized at. Each row opens the full list for that tier.
| Memory | Top picks | Everything that fits |
|---|---|---|
| 8 GB |
| 71 models → |
| 12 GB |
| 85 models → |
| 16 GB |
| 95 models → |
| 24 GB |
| 131 models → |
| 32 GB |
| 132 models → |
| 48 GB |
| 138 models → |
A research paradigm shift by Microsoft. BitNet b1.58 replaces 16-bit weights with ternary {-1, 0, 1} weights. This eliminates Matrix Multiplication (MatMul) in favor of simple addition, offering extreme speed and efficiency on CPUs.
DeepSeek R1 rivals proprietary top-tier models in reasoning and coding tasks using a Mixture-of-Experts architecture. Known for its exceptional logic and math capabilities. Previous generation — superseded by DeepSeek V4, which folds R1's reasoning strengths into a newer MoE architecture. Still widely deployed and supported.
Meta's latest flagship model. Llama 3.3 70B delivers state-of-the-art performance, matching the massive 405B model in many benchmarks while remaining efficient enough for high-end local setups.
The default first install and still the most-pulled model on Ollama by a wide margin. Llama 3.1 is legacy-but-safe rather than current-best: newer small-dense models beat it on coding and reasoning at the same size, but nothing matches it for tooling support, community familiarity, and being the model every guide assumes. Keep it as the smoke test, not the recommendation.
NVIDIA's specialized fine-tune of Llama 3.1 70B. It tops many leaderboards for helpfulness and instruction following, significantly outperforming the base Llama model in alignment.
Optimized specifically for RAG (Retrieval Augmented Generation) and tool use. Command R models excel at citing sources and reasoning over provided documents.
Microsoft's latest model focusing on high-quality synthetic data training. Phi-4 14B outperforms many models twice its size in reasoning benchmarks.
Microsoft's highly efficient small language models (SLMs). Perfect for mobile devices, edge computing, and low-VRAM environments without sacrificing basic reasoning.
The coding and math powerhouse of the open-weight world. Qwen 2.5 models consistently outperform Llama 3.1 at equivalent sizes in STEM tasks.
Built from the same research and technology as Google Gemini. Gemma 2 uses knowledge distillation for extremely high efficiency and creative performance.
The European champion of open AI. Known for efficiency and the 'Nemo' collaborative model with NVIDIA. Mistral models are production-ready workhorses.
High-performance models from 01.AI. Known for strong bilingual (English/Chinese) capabilities and coding performance.
IBM's enterprise-grade model family. Granite 3.0 is Apache 2.0 licensed and rigorously trained for coding, transparency, and safety in business environments.
Meta's groundbreaking Mixture-of-Experts (MoE) series. Llama 4 uses a sparse MoE architecture where only a fraction of parameters activate per token, delivering frontier-class intelligence with far lower hardware requirements than the total parameter count suggests.
Gemma 3 delivers exceptional quality-per-VRAM ratio, with the 4B model outperforming many 7B competitors. Features multimodal capabilities and a 128k context window across all sizes. Previous generation — superseded by Gemma 4 (E2B/E4B/26B-A4B/31B and the 12B Unified follow-up), which improves efficiency and multimodal support. Still widely used and supported.
Google's fourth-generation open model family. The first wave (E2B, E4B, 26B-A4B MoE, 31B dense) launched March 31, 2026 under Apache 2.0 with a fully multimodal architecture — text, image, and audio across all sizes, plus video on the 31B. A separate 12B 'Unified' encoder-free model followed June 3, 2026, adding native audio+video to a laptop-friendly 16GB footprint. The E2B and E4B Efficient variants run on 4–6 GB VRAM GPUs, while the 26B-A4B MoE and 31B dense models deliver near-frontier quality on a single RTX 4090.
Meta's multimodal and edge-optimized series. Llama 3.2 introduces vision capabilities (11B and 90B variants) and ultra-compact 1B/3B versions designed to run on smartphones and edge devices. All variants feature a 128k context window.
DeepSeek's groundbreaking non-reasoning frontier model. Trained for just $5.5M — 1/10th the cost of comparable proprietary models. Uses a 685B MoE architecture (37B active parameters per token), topping the open-source leaderboard for coding, math, and general instruction following at launch. Previous generation — superseded by DeepSeek V4, which expands context to 1M tokens and improves MoE efficiency. Still widely used and supported.
Alibaba's model series with a hybrid thinking mode. Qwen 3 models can toggle between fast response mode and deep Chain-of-Thought reasoning on demand — type /think for hard problems, /no_think for fast answers. Available in dense and MoE variants. Previous generation — superseded by Qwen 3.5 and Qwen 3.6, which extend context length and improve reasoning efficiency. Still widely used and supported.
Alibaba's powerful vision-language models that understand images, documents, charts, and videos alongside text. Qwen 2.5 VL 72B rivals GPT-4V on multiple visual benchmarks, while the 7B version delivers strong OCR and document parsing in under 6GB VRAM.
Mistral's latest small model, now with vision capabilities and Apache 2.0 license. Mistral Small 3.1 24B features a 128k context window (4× larger than its predecessor), multimodal image understanding, and relicensed to Apache 2.0 for full commercial use — while outperforming Gemma 3 27B on most benchmarks.
Mistral's dedicated code generation model, trained on a massive programming corpus covering 80+ programming languages. Excels at fill-in-the-middle (FIM) completion — the key technique powering IDE autocomplete. Integrates natively with VS Code via continue.dev and Cursor.
Microsoft's ultra-efficient small model for mobile and edge deployment. Phi-4 Mini achieves remarkable reasoning capabilities in just 3.8B parameters using high-quality synthetic data — the same approach behind Phi-4 14B. Designed for on-device AI on phones and laptops without discrete GPUs.
The premier open-source code generation model. StarCoder 2 is trained on The Stack v2 dataset covering 600+ programming languages. Purpose-built for code completion, generation, and understanding — not a general-purpose model fine-tuned for code.
The world's most transparent LLM. OLMo 2 releases everything: weights, training code, training data, evaluation suite, and intermediate checkpoints. Perfect for researchers and compliance-sensitive deployments needing fully auditable AI. Competitive with Llama 3.1 on most benchmarks.
The smallest production-quality LLMs. Hugging Face's SmolLM2 models are designed to run on microcontrollers, phones, and browsers via WebAssembly. Despite tiny size, they show surprising intelligence thanks to careful data curation with the Smoltalk dataset.
The Technology Innovation Institute's latest generation of models, trained on 14 trillion tokens. Falcon 3 achieves state-of-the-art results for its parameter count and comes with a fully permissive Apache 2.0 license — making it one of the best choices for commercial deployment.
A powerful open-source model from China's Shanghai AI Lab. InternLM 3 excels at bilingual Chinese-English tasks, coding, and long document analysis. Particularly strong in math and STEM reasoning, making it the top open choice for Chinese language applications.
The world's best open multilingual model. Aya Expanse supports 23 languages and is specifically optimized for non-English speakers. A breakthrough for global AI accessibility — outperforms Llama 3.1 in Arabic, Chinese, Hindi, French, Spanish, Russian, and 17 more languages.
Alibaba's dedicated coding line. The 480B-A35B flagship holds state-of-the-art among open models on SWE-bench Verified and agentic coding at release (Qwen positions it comparable to Claude Sonnet 4), while the 80B-A3B variant delivers most of that quality with only 3B active parameters — cheap enough to self-host. Best-in-class open weights for repo-level coding, debugging, and code review.
Zhipu AI — rebranded internationally as Z.ai — released this February 2026 MoE model. GLM-4.7 uses a 32B active MoE architecture and scores on par with Claude Opus 4.5 on hallucination leaderboards. The reasoning-focused GLM-Z1 variant posts near-GPT-5.2 scores on the Artificial Intelligence Index. Strong multilingual support with particular depth in Chinese.
Cohere's February 17, 2026 release — the most capable multilingual open-weight model at the 3B parameter class. Tiny Aya outperforms Qwen3-4B, Gemma 3 4B, and Ministral 3B on multilingual benchmarks across 23 languages. Perfect for edge devices, phones, and resource-constrained deployments where global language support is critical.
LG AI Research's flagship open-source model family with top-tier real-world usability scores across 20 benchmarks. Excels at instruction-following and long-context tasks in both English and Korean. MIT licensed for commercial use.
Google's mobile-first multimodal model family. Uses the novel MatFormer nested architecture — a single model file contains multiple sub-models (E2B/E4B) that can run at different sizes. Processes text, images, audio, and video. Runs on phones without internet.
Meta's multimodal extension of Llama 3.2, adding a vision encoder for image understanding. The 11B version runs on a single 8–12 GB VRAM GPU. Supports image analysis, OCR, chart reading, and visual Q&A alongside text generation.
Mistral AI's lightweight model family for resource-constrained environments. Ministral 8B punches well above its weight on reasoning tasks and fits on any GPU with 6 GB VRAM. Ministral 3B is designed for phones and edge devices.
Alibaba's next-generation flagship family, released in waves from Feb 16 to Mar 2, 2026 (397B-A17B first, then the 122B/35B-A3B/27B tier on Feb 24, then the 0.8B–9B 'Small' series on Mar 2). Qwen 3.5 introduces a hybrid Gated DeltaNet + MoE architecture delivering frontier performance at a fraction of the compute, with native multimodal vision, ~256K context (extensible to 1M on the larger tiers), 201 languages, and Apache 2.0 license. Note: the 397B-A17B flagship is only available as a cloud-hosted Ollama tag (`:cloud`), not a true local download — everything 0.8B–122B-A10B has standard local Ollama tags.
Moonshot AI's 1-trillion parameter MoE flagship with 384 experts and a unique Agent Swarm architecture. Powers Cursor's Composer 2 IDE feature. Leads multiple coding and reasoning benchmarks including 93.33% on AIME 2026 and best Pass@1 on SWE-rebench. With 32B active parameters per token, requires datacenter infrastructure for self-hosting.
MiniMax's 230B MoE model scoring 80.2% on SWE-bench — technically outperforming Claude Opus 4.6 on this benchmark — at 22.7× lower cost on output tokens. Supports up to 1M token context and includes multimodal capabilities. Available under a modified MIT license requiring attribution. Primarily API-accessed but self-hostable on multi-GPU infrastructure.
Xiaomi's 1T+ MoE model originally launched anonymously as 'Hunter Alpha' in March 2026 before being revealed as MiMo-V2-Pro. Led by ex-DeepSeek engineer Luo Fuli. Was the largest model on OpenRouter at launch. Strong on PinchBench and ClawEval agentic benchmarks with ~42B active parameters per token. Available via API through Xiaomi and OpenRouter.
DeepSeek's updated 685B MoE model delivering approximately 90% of GPT-5 quality at 1/50th the price. Significant performance jump over V3 — AIME 2026 at 91.67% in thinking mode. MIT-licensed and widely adopted. Input tokens at ~$0.28/1M make it one of the most cost-effective frontier models via API. Self-hostable on multi-GPU clusters (~370 GB at Q4). DEPRECATED: DeepSeek V3.2 is being retired (reported 2026-07-24) in favour of DeepSeek V4 (see the "DeepSeek V4" family) — kept here for users still running it.
NVIDIA's hybrid Mamba-2 + MoE model, released March 2026 as a single checkpoint: Nemotron-Cascade-2-30B-A3B. 31.6B parameters resident, 3.5B active per token, and the second open-weight model to reach gold-medal level at IMO 2025, IOI and the ICPC World Finals. NVIDIA also open-sourced the post-training recipe. NVIDIA Open Model License, 1M-token context, first-party Ollama tag.
OpenAI's open-weight family released under Apache 2.0 — their first substantial open-source release since GPT-2. Both sizes ship in MXFP4 precision, which is why they fit in far less memory than their parameter counts suggest: the 20B loads in ~14 GB and runs on a 16GB consumer card, while the 120B needs ~65 GB (dual RTX 3090 or a single 80GB A100). Signals OpenAI's strategic shift toward supporting on-premise enterprise deployments.
Mistral AI's coding specialists, spanning both generations. The original Devstral Small 2505 posts 46.8% on SWE-Bench Verified; its December 2025 successor Devstral Small 2 lifts that to 68.0% at the same 24B, which is the local-agent sweet spot — about 15 GB at Q4_K_M on one consumer card. The 123B flagship reaches 71.6%. Built for agentic software engineering: multi-file editing, repo navigation, and test-driven development. Apache 2.0 throughout.
Deep Cogito's hybrid reasoning model family released March 2026. Unique 'think-or-not-think' architecture lets the model decide at inference time whether to use chain-of-thought reasoning based on query difficulty — saving compute on simple tasks while using deep reasoning for hard ones. Apache 2.0 licensed with sizes from 3B to 70B, all runnable on consumer hardware.
Moonshot AI's cutting-edge coding and agentic model series. Kimi K2.5, K2.6 and the coding-specialised K2.7 Code rank among the top models globally for coding tasks, multimodal understanding, and autonomous agent workflows. Built for developers who need a model that can reason, use tools, browse the web, write and debug code end-to-end.
Moonshot AI's July 2026 flagship and its largest model to date — a 2.8-trillion-parameter Mixture-of-Experts that landed in the July 16–27 window. K3 roughly doubles the expert count of Kimi K2.5 (1T) while keeping active parameters modest (~48B per token), pushing Moonshot's lead on agentic coding, tool use, and long-horizon reasoning further ahead of K2.6. Natively multimodal with a 256K context. At this scale it is a data-center / API model — self-hosting needs well over a terabyte of memory even at Q4, so most users reach it through the Kimi API or the smaller K2.x community builds.
Zhipu AI's (Z.ai) fifth-generation model series, released under MIT license for maximum flexibility. GLM-5 and GLM-5.1 excel at agentic workflows, tool calling, and long-context reasoning. The MIT license makes them one of the most commercially permissive frontier-class models available, with zero restrictions on commercial use or distribution. Both ship only as a 744B Mixture-of-Experts model (40B active, SWE-bench Verified leader) trained entirely on Huawei Ascend hardware: a data-center model, with no smaller consumer checkpoint. The 9B, 32B and 72B sizes listed here earlier were never released.
Alibaba's April 2026 follow-up to Qwen 3.5. As of June 15, 2026 only two tiers have been released — a 27B dense model (Apr 21–22) and a 35B-A3B MoE model (Apr 16) — both Apache 2.0, with native text/image/video input and ~256K context (extensible to ~1M via YaRN). Qwen 3.6-Plus/Plus-Preview/Max-Preview exist but are proprietary, API-only, and not listed here.
DeepSeek's April 24, 2026 preview release, MIT licensed with a 1M-token context window. V4-Pro (1.6T total / 49B active) is server-class only — not runnable on consumer hardware even heavily quantized. V4-Flash (284B total / 13B active) is workstation-tier but local support is still WIP: llama.cpp has no merged upstream support as of June 2026 (only experimental forks), and Ollama's listings appear to route to cloud-hosted inference rather than a true local download. DeepSeek R2 remains unreleased/rumored as of June 15, 2026 and is intentionally not listed.
Mistral AI's March 16, 2026 release unifying its former Magistral/Pixtral/Devstral lines into a single 119B-total / ~6.5B-active MoE model (128 experts, 4 active per token) with text + image input, reasoning, and agentic coding. Apache 2.0 licensed, 256K context. At Q4 it fits a single 24GB GPU (RTX 4090).
IBM's April 29, 2026 release. The 3B and 8B tiers are dense transformers; the 30B tier pivots back to a hybrid Mamba-Transformer (interleaved SSM + attention) design rather than 4.0's MoE. Apache 2.0 licensed, ISO 42001 certified, 128K context (extendable to 512K). All three sizes have official Ollama tags, spanning budget edge hardware to mid-range 24GB GPUs.
MiniMax M3 — a natively multimodal MoE at roughly 428B total and 23B active with a 1M-token context, under the minimax-community licence (June 2026). Multimodality is built into the single checkpoint: there is no separate vision release.
Mistral's open-weight flagship, released 2 December 2025: a 675B MoE with 41B active parameters, Apache-2.0, 256K context. Datacentre-scale to self-host; most readers will rent it.
DeepSeek V4.1 — the Flash member shipped 10 September 2026 under MIT: a 552B-backbone MoE with a causal encoder-decoder design (8B active per input token, 16B per output token), a 196B-parameter Engram memory table and native image input. A V4.1 Pro was announced without a date and had not launched as of October 2026.
Qwen 3.7 shipped as API-only Max and Plus tiers in May and June 2026. Alibaba skipped the generation for open weights and went straight to Qwen 3.8, so nothing in this family is locally runnable.
GLM-6 has not been released. Z.ai's current generation is GLM-5.3 (August 2026), alongside GLM-5.3-Flash. The listings here were withdrawn; they redirect to the nearest real model.
Llama 4.5, reported as a June 2026 mid-cycle refresh of Llama 4 Scout and Maverick. UNVERIFIED: secondary sources disagree on whether this release happened, and no primary source has been read. Meta's confirmed 2026 open release is Muse Glimmer 30B (10 August, Apache-2.0).
PREVIEW (July 2026). The first 27B-class model that runs on a phone. Bonsai 27B is PrismML's extreme low-bit rebuild of Qwen3.6-27B: the architecture is unchanged, but the weights are squeezed to roughly 1–1.7 bits each. The 1-bit build is just 3.9 GB — small enough for an iPhone — while the ternary build is 5.9 GB and aimed at laptops. Both are multimodal (text + images), carry a 262K-token context, and ship under Apache 2.0. PrismML reports the ternary build keeps more than 95% of the full-precision (FP16) baseline across a 15-benchmark suite and the 1-bit build more than 90%, running about 11 tok/s on an iPhone 17 Pro. Specs are from launch coverage — verify against the Hugging Face model cards.
PREVIEW (July 2026). Thinking Machines Lab's first open-weights model — and the first ~1-trillion-parameter open model to natively accept text, image and audio input with a 1M-token context window. Inkling is a sparse multimodal Mixture-of-Experts transformer: 975B total parameters but only ~41B active per token (256 experts, top-6 routed plus 2 always-on shared experts). It was trained on 45 trillion tokens of text, images, audio and video, and reasons natively across modalities with seven selectable effort levels (none → max). Ships in full BF16 (~2 TB VRAM, Hopper+) and a well-calibrated NVFP4 4-bit build (~600 GB, Blackwell), both with MTP speculative-decoding layers for faster inference. Released under Apache 2.0 and pitched as a customization base to fine-tune via the company's Tinker platform — "built for customization, not leaderboard dominance." Datacenter-class hardware required; most users will reach it through a hosted API. Specs from launch coverage and the Hugging Face model card — verify before relying on them.
PREVIEW (2026). NVIDIA's open frontier World Foundation Model (WFM) for Physical AI — the first fully open omnimodel that natively understands and generates text, images, video, ambient audio and robot/vehicle actions in a single system, with leading physics accuracy. Cosmos 3 is a two-tower Mixture-of-Transformers (MoT): a reasoner tower (an autoregressive vision-language "brain" that interprets images/video/text and reasons about motion, object interactions and physical cause-and-effect) plus a diffusion-based generator tower that predicts physics-aware future video and action sequences conditioned on that reasoning — so it "thinks before it acts." The two towers keep separate norm layers and MLPs but share multimodal attention. Trained on ~20 trillion tokens of multimodal data (nearly a billion images, 400M+ real and synthetic videos, plus audio and human/robot action data), it compresses physical-AI training and evaluation from months to days by generating synthetic worlds and predicting actions for robots and autonomous vehicles. Ships in three scales — Super (64B), Nano (16B) and Edge (4B) — under the permissive OpenMDW-1.1 license, on Hugging Face (nvidia/) and NVIDIA NGC. NVIDIA reports open-source state-of-the-art results on R-Bench, PAI-Bench, Physics-IQ and RoboLab; treat launch numbers as vendor-reported pending independent replication. This is a robotics / autonomous-vehicle / world-simulation model, not a chat LLM — verify specs on the model cards before relying on them.
Poolside's July 2026 open-weight coding model, and one of the few frontier-grade agentic coders small enough to run on a single consumer GPU. Laguna XS 2.1 is a 33B-total / 3B-active Mixture-of-Experts built for long-horizon software work — the 2.1 point release lifted SWE-Bench Verified from 69.9% to 70.9% and SWE-bench Multilingual from 57.7% to 63.1%. Architecturally it mixes sliding-window and global attention 3:1 across 40 layers, with sigmoid per-head gating and an FP8 KV cache, which is how it holds a 256K context without a datacenter behind it.
NVIDIA's 120B open Mixture-of-Experts, activating just 12B parameters per token for compute efficiency on multi-agent workloads. Like the rest of the Nemotron line it generates a reasoning trace before its final answer, and NVIDIA publishes open weights alongside training data and recipes rather than weights alone. Positioned for collaborative agents and high-volume automation such as IT ticket triage. First-party Ollama tags exist at Q4_K_M, Q8_0 and BF16, so this is genuinely self-hostable — on a multi-GPU box rather than a laptop.
The top of NVIDIA's Nemotron 3 line: a 550B-parameter hybrid that interleaves Mamba-2 state-space layers with latent mixture-of-experts layers and a minority of attention layers, with 55B active per token. Pre-trained with an NVFP4 recipe and published in BF16 and NVFP4 under OpenMDW-1.1. NVIDIA's own minimum for the BF16 checkpoint is an 8-GPU H200 or B200 server.
Z.ai's June 2026 flagship and, by Artificial Analysis' ranking at the time, the strongest open-weight model available — fourth overall including closed models. A 744B-parameter Mixture-of-Experts with roughly 40B active per token and a genuinely usable 1M-token input context (output up to 131,072). Two selectable reasoning-effort levels. Released in stages between 13 and 17 June 2026: the Coding Plan first, then open weights on Hugging Face, then documentation. The MIT license is the notable part — full commercial use, modification and redistribution on a frontier-scale model.
MiniMax's March 2026 agentic MoE: 230B total parameters with 10B active per token across 256 experts — a 4.3% activation rate — over 62 layers, using multi-head causal self-attention with RoPE. Weights are published openly on Hugging Face and GitHub, so this is self-hostable rather than API-only, though the scale still implies a multi-GPU box. NVIDIA has upstreamed fused QK RMS Norm and FP8 MoE kernels from TensorRT-LLM for it. ⚠️ Released under a NON-COMMERCIAL license — commercial deployment requires a separate agreement with MiniMax.
The Granite 4.0 generation, listed here for the one tier that has no Granite 4.1 counterpart: the 32B-total / 9B-active hybrid mamba-2 build. Granite models are finetuned from their base models on permissively-licensed open instruction datasets plus internally-collected synthetic data, with improved instruction-following and tool-calling aimed squarely at enterprise RAG and agent workloads. Apache 2.0. For the smaller sizes, see Granite 4.1 — those tiers are newer and supersede their 4.0 equivalents.
Hugging Face's fully-open 3B: weights, training data, recipe and 100+ intermediate checkpoints all published. Dual-mode — a deep-thinking path for hard problems and a fast path for everything else — with 128K context and six natively supported languages. The reference model for anyone fine-tuning at the small end, which makes it the natural companion to the datasets hub.
MIT-licensed reasoning fine-tunes that argue the case for small models: VibeThinker-3B scores 94.3 on AIME26, matching what DeepSeek V3.2 reaches at 671B. Trained with the Spectrum-to-Signal pipeline — explore solution diversity during SFT, then reinforce the correct signal with RL. Vendor-reported and contested; treat the headline comparison as a claim about a narrow benchmark family, not general capability.
Mistral's reasoning line: Mistral Small with long chain-of-thought added via SFT on Magistral Medium traces plus RL. Apache 2.0, vision-capable, and sized so the whole thing fits a single RTX 4090 or a 32 GB MacBook once quantized — which is the point. Reasoning traces degrade past roughly 40K tokens even though the context window runs to 128K.
A reliability release rather than a capability one. Same 24B architecture as Small 3.1, with instruction-following, function-calling robustness and output stability fixed — Arena Hard more than doubles (19.56% to 43.10%) and the infinite-repetition rate halves. If you were running 3.1, this is a drop-in upgrade with no new hardware cost.
The first mainstream omni-modal open model that a single consumer GPU can actually hold: text, image, speech and video in, text out, at 3B active parameters from a 30B MoE. 260K context under the NVIDIA Open Model License, which permits commercial use. The efficiency ratio here is the story — video understanding at roughly the per-token cost of a 3B dense model.
Alibaba's August 2026 wave. The 27B is a natively multimodal dense model under Apache 2.0 with a 262K context that extends to 1M via YaRN — small enough for one consumer card, which is the combination most of this site's visitors are looking for. Qwen3.8-Max sits at the other extreme: 2.4T parameters with roughly 95B active, open-weighted under its own licence. Qwen3.8-Flash-Next is neither — an experimental preview of the Qwen4 architecture, 180B resident but only 6B active, under the Qwen Community License rather than Apache.
Xiaomi's MIT-licensed MoE pair, and the clearest licence story at the trillion-parameter tier — MIT, not a bespoke open-weight agreement with revenue thresholds. Both carry 1M-token context and agentic multimodal capability. Consistently omitted from best-of roundups despite matching Kimi K2.6 on aggregate index scores.
Xiaomi's September 2026 flagship and the successor line to MiMo-V2.5 and MiMo-V2-Pro. One released checkpoint, MiMo-V2.6-Pro: a 1.02-trillion-parameter sparse MoE with 42B active per token that takes text, image, video and audio in, with a 1M-token context, under MIT. Every expert must be resident, so this is a datacenter model: Xiaomi's own serving recipes are SGLang and vLLM on multi-GPU servers.
The edge tier of the Mistral 3 family — 3B, 8B and 14B dense, every size shipped in base, instruct and reasoning variants with image understanding, all Apache 2.0. Built for laptops, phones, drones and anywhere a datacenter is not. The 3B runs in about 4 GB of video memory; the 14B reasoning variant scores 85% on AIME 2025.
The September 2026 agentic family from Nex-AGI, an innovation alliance initiated by the Shanghai Innovation Institute. Three sizes, all post-trains of other labs' base weights: mini and Pro continue the multimodal Qwen3.5 lineage of Nex-N2, while Max is the alliance's first trillion-scale post-train, built on a 1.6T text-only DeepSeek-V4-Pro base. Vision is treated as an action interface rather than an input mode — the models drive computers and browsers and check their own work from what appears on screen. Max takes BrowseComp at 92.6, ahead of every model in Nex's own comparison table. Only the 35B mini is realistically self-hostable.
Nex-AGI's Apache-2.0 agentic pair, post-trained on Qwen3.5 and built around what Nex calls Agentic Thinking: Adaptive Thinking lets the model choose when and how deeply to reason, Coherent Thinking keeps one reasoning style across chat, tool calls and terminal work. Pro reaches 75.3 on Terminal-Bench 2.1 and 90.7 on GPQA Diamond. The 35B mini is the size that matters for local use — 21.9 GB at Q4_K_M with 3B active parameters per token. This is the generation whose parameter counts Nex states outright.
Nex-AGI's first agentic series, and the one that set the pattern: take someone else's open base weights, post-train them for tool use and autonomy, publish the result. The flagship is a DeepSeek-V3.1 post-train scoring 80.2 on tau2-Bench and 70.6 on SWE-bench Verified. Superseded by Nex-N2, but kept because the question 'what does a 671B agentic post-train need?' has a factual answer that does not change when the model stops being recommended.
The first natively multimodal model in the GLM-5 series, and the one Z.ai actually released under MIT — 320B total parameters with 18B active, text plus image and video in, a 1M-token context, and hybrid attention (linear attention for local dependencies, sparse attention for distant retrieval). Z.ai claims 63.4 on DeepSWE against GLM-5.2's 46.2. Note this is NOT the GLM-5.3 flagship: that is a separate 753B model under a bespoke licence with a hyperscaler security-review clause.
Z.ai's 753B-parameter flagship, released in August 2026. It shares GLM-5.2's base model, so every gain is post-training: Z.ai's card leads with agentic coding and cyber-security benchmarks. Text only, 1M-token context, and a bespoke licence: MIT-style, except that a Model-as-a-Service operator with more than $10 billion in annual revenue must pass Z.ai's security review before commercial use. For the natively multimodal 320B sibling, see GLM-5.3-Flash.
Kolibri 1 is Aleph Alpha's mixture-of-experts reasoning model for German and English, released in October 2026 with Apache-2.0 weights. It was trained from scratch, supports adjustable reasoning effort and tool calling, and uses a mostly sliding-window attention design to keep long-context cache costs low. The memory trade-off is real: every expert must be resident, so it is a server and high-memory-workstation model, not a laptop model.
Trinity is Arcee AI's open-weight sparse mixture-of-experts family. Trinity Mini (26B total, about 3B active) is the locally practical member with an official GGUF; Trinity-Large-Thinking (about 398B total, about 13B active) is a reasoning and agent model for multi-GPU servers. The whole family moved from Apache-2.0 to OpenMDW-1.1 on 29 May 2026, so older Apache wording is stale.
Nemotron 3 Nano 4B is NVIDIA's compact text model, compressed from Nemotron-Nano-9B-v2 into a hybrid Mamba-2 and attention design with only four attention layers. It targets edge devices and RTX PCs, offers a switchable reasoning mode, supports up to 262K context, and ships with an official Q4_K_M GGUF and an Ollama tag.
Seed-OSS-36B is ByteDance Seed's dense 36B open-weight reasoning model with an Apache-2.0 licence, a controllable thinking budget, tool-call parser support and a native 512K context window. It dates from August 2025 and is kept in the library as a supported legacy option; the long context is a model capability, not a promise that a single consumer GPU can serve it.
OLMoE-1B-7B is Ai2's small, fully documented mixture-of-experts research model with 64 experts per layer and roughly 1.3B parameters active per token out of 6.9B. The January 2025 Instruct release has official GGUF files, which makes it a lightweight option for phones, tablets and Apple Silicon, but it is English-focused, short-context and older, so it is listed as a niche legacy entry.
GLM-4.5-Air is Z.ai's 106B-total, 12B-active hybrid-reasoning mixture-of-experts model (thinking and non-thinking modes) with MIT-licensed weights. Released in mid-2025 and since followed by newer GLM generations, it stays popular as a big-memory local option; it is listed as legacy and cross-linked to the newer GLM entries.
Cohere's first model aimed at developers rather than enterprises, and the opening entry in their North family of code agents. A 30B sparse MoE activating ~3B per token, Apache 2.0, 256K context and 64K maximum generation — small enough to run locally while posting a Coding Index of 33.4, ahead of several open models many times its size.
A fully open model family from MBZUAI's Institute of Foundation Models — weights, training data, recipe and code all published under Apache 2.0, with intermediate checkpoints released too. Three sizes: a 3.7B and a 7B dense model for small machines, and a 36B mixture-of-experts model that activates about 4B parameters per token. All three have a 512K native context. On the independent Artificial Analysis Intelligence Index they score 16, 21 and 25, the strongest results in their size classes when they were checked.
IBM's first family of dense decoder-only reasoning models, released August 2026 in 3B, 8B and 30B. Granite 4.2 drops the hybrid Mamba-2 MoE design of Granite 4.0 and continues the dense line Granite 4.1 started. Each size is pretrained from scratch on roughly 15T tokens with a five-phase context extension to 512K, then RL-post-trained for tool calling, RAG, structured JSON and an explicit thinking mode. All three are Apache 2.0.
OpenBMB's pocket-sized multimodal line, backed by ModelBest and Tsinghua University, built around one goal: image and video understanding that actually runs on a phone. Version 4.5 is the 8B that beat much larger proprietary models on OpenCompass; version 4.6 is the 1.3B successor that cut visual-encoding FLOPs by more than half and ships on iOS, Android and HarmonyOS.
A 3B vision-language model built for one job — turning document pages into structured text — and a demonstration of "contexts optical compression": rendering text as an image and decoding it costs roughly 10x fewer tokens at near-lossless fidelity, or 20x at about 60% precision. A DeepEncoder compresses each page into vision tokens which a DeepSeek-3B MoE decoder reads, activating only ~570M parameters. A second generation, DeepSeek-OCR-2, followed in January 2026.
Liquid AI's on-device line, built around their hybrid LIV convolution architecture rather than a stack of attention blocks. The 8B-A1B is the flagship: a sparse MoE small enough for phones, laptops and robots, reasoning-only (it emits an explicit chain of thought before answering), and trained on 38T tokens with large-scale RL. Liquid claims tool-calling quality comparable to models up to 4x its active size.
Meta's return to open weights, and its first since Llama 4. A 29.6B dense model distilled from the closed Muse Spark, with its own perception encoder, built specifically for agentic work on one consumer card: multi-step reasoning, tool use and failure recovery in a single checkpoint that runs with no network access.
NVIDIA's agentic line, built on a hybrid stack that interleaves Mamba-2 state-space layers and MoE blocks with a minority of attention layers. Only the attention layers hold a KV cache, which is what lets it advertise a 1M-token context at this size.
Z.ai's vision-language line. The Flash member is the 9B efficiency variant, sharing GLM-4.6V's multimodal capabilities at a fraction of the compute and small enough for an 8 GB card.
Fully open, not just open-weight: Ai2 publishes the weights, the training data (Dolma 3), the code and the intermediate checkpoints under Apache-2.0. Base, Instruct and Think variants at each size, which makes this the family to reach for when the training process itself has to be auditable.
Poland's open LLM, trained on Polish large-scale computing infrastructure (PLGrid) on a corpus curated by the SpeakLeash team. v3 replaced the tokenizer with APT4, built specifically for Polish, which is why it does more with its parameters on Polish text than a general multilingual model of the same size.
Trained from scratch on the MareNostrum 5 supercomputer at the Barcelona Supercomputing Center, covering all 24 official EU languages. The European answer to models whose multilingual coverage is an afterthought: EU languages are the training target, not a side effect.
Switzerland's open model, from EPFL, ETH Zurich and the Swiss National Supercomputing Centre. Trained on roughly 15 trillion tokens spanning more than 1,000 languages, with weights, data and process published under Apache-2.0 — the European sovereignty argument made concrete.
PrismML's second ternary family: four vision-language models whose weights are only −1, 0 or +1, stored in packed formats about nine times smaller than FP16. The 2B, 4B and 9B are rebuilt from Qwen3.5 and the 27B from Qwen3.8-27B, so all four keep a 262,144-token context and a hybrid attention backbone (three linear-attention layers to every full-attention layer), which also keeps the KV cache small. They ship ONLY in these packings — there is no Q4_K_M or Q8_0 build — and the smallest files need PrismML's own llama.cpp fork or mlx-vlm to load. Apache 2.0. The 27B succeeds the July Bonsai 27B, which was built on Qwen3.6.
Meta's open-weight release of its Muse Spark line, published on 15 October 2026: an 816B-total, 42B-active mixture-of-experts model that takes text, images, video and audio in, with a 1M-token context. Every expert must be resident, so even Meta's 4-bit (NVFP4) checkpoint is a 533 GB download and Meta asks for at least 700 GB of GPU memory to run it. It is a multi-GPU server model, not a desktop one. If you want Meta's weights on one consumer card, Muse Glimmer 30B is the local member of the line; Ternary Bonsai 2 27B and Qwen3.8 27B are the other single-GPU alternatives in this library.
JetBrains' open coding models. Mellum2.1 Thinking is a 12B mixture-of-experts model with 2.5B active parameters, post-trained mostly with reinforcement learning inside real repositories (shell and file-editing tools, rewarded when tests pass), for agentic software engineering and as a sub-agent. Apache 2.0, 131,072-token context. At about 8 GB in Q4_K_M it is one of the few agentic coding models that fits a mid-range laptop.
Mistral's first merged flagship: one dense 128B model that folds the reasoning (Magistral) and coding (Devstral 2) lines into a single checkpoint, with image input and a 256K context. Dense means every parameter is read for every token, so it is slower per token than a mixture-of-experts model of the same size, but at about 78 GB in Q4_K_M it fits a 128 GB Mac or a pair of 48 GB cards. Its "Modified MIT" licence withholds every right from companies above $20M monthly revenue.
Bespoke Labs' 9B decision model, a LoRA on Qwen3.5-9B trained with contrastive data pairs. The default decision model in Ollama's library.
Together AI's experimental Jev-inspired decision models, supervised fine-tunes of Qwen3.5 that pick one option from a structured decision. Available in Ollama.
Gemma 4 fine-tunes for typed decisions, shipped as GGUF with a llama.cpp-based server that serves /v1/systemone and chat/vision from one loaded model.
A family of open decision models on Qwen3.5 bases (0.8B to 35B-A3B) with one interface, up to 255 options per question and a TypeSafe-compatible server.
A 4B decision model on Qwen3.5-4B with a TypeSafe-style server and GGUF builds for llama.cpp on NVIDIA, AMD, Intel, Apple GPUs or CPU.
InternLM's multimodal structured decision models on Qwen3.5 that answer every question in a schema from one forward pass.
A 27B multimodal decision model on Qwen3.8-27B, the highest-scoring downloadable model on the community Decision Index as of 28 Sep 2026.
Small encoder-based decision models (ModernBERT / mmBERT) for fast routing and guardrails in 100+ languages, installed with pip.
Cloudflare's decision models, Clef 27B and Clef-Flash 9B: typed questions in, a probability per option out, in one scoring pass with no text generation. Both are in Ollama's library and accept images.
Surogate's decision model on Gemma 4 26B-A4B (mixture-of-experts, text and images). The highest-scoring downloadable model on the community Decision Index 0.2.1 snapshot of 28 September 2026.
Open decision models distilled from calibrated teachers: a 400M encoder (nano) and a 4B LoRA on Qwen3-4B (small), installed with pip and served by opendecider serve, a /v1/systemone drop-in. Larger 30B and 80B MoE variants need NVIDIA GPUs and are not listed here.
Hanno Labs' small decision models: rank-16 LoRAs on Qwen3-0.6B and Qwen3-1.7B with 256 learned decision tokens, so one question can carry up to 255 candidates. Served by the Jev-compatible server at /v1/systemone.
Liquid AI's edge decision models: d1-3B (text and images, on LFM2.5-VL-3B) and the experimental d1-omni-600M (text with images or audio). A Python library loaded through Transformers, not a server; released under the LFM Open License v1.0.
HAL-X AI's 322M multilingual decision model: an mmBERT-base encoder with a decision head, post-trained in 18 languages. Besides choice, yes/no and score it returns numbers, verbatim excerpts and citations from a document.
Alibaba's February 2026 flagship generation. Qwen 3.5 immediately topped multiple open-source reasoning benchmarks, scoring 88.4 on GPQA Diamond and 76.4 on SWE-bench Verified. The MoE architecture delivers 8–19× higher decoding throughput vs earlier Qwen3 while supporting 200+ languages. Apache 2.0 licensed for full commercial use.
Google DeepMind's breakthrough open-weight model. Gemma 4 27B delivers GPT-4 level performance at 14 GB VRAM, making frontier-quality AI accessible on consumer GPUs. Features a hybrid architecture with interleaved local and global attention, multi-image understanding, and 128k context window. Achieves 85 tokens/second on RTX 4090.
Rank #1 open-weight model on every major benchmark. DeepSeek V4 Pro uses a massive 1.6 trillion parameter Mixture-of-Experts architecture with only 49 billion parameters active per token — giving frontier-tier intelligence at a fraction of the compute cost. Dominates coding, math, reasoning, and multilingual tasks.
Alibaba's most ambitious open-weight model release. Qwen 3.5 (397B) is a natively multimodal model supporting text, vision, and video understanding in a single architecture. Ranks among the top 3 open-weight models globally and competes directly with GPT-4o and Gemini 2.0 Pro on multimodal benchmarks.
DeepSeek-V3.2-Exp, the experimental sparse-attention snapshot released in September 2025 ahead of the V3.2 line.
Decision models (Jev-style System One: they answer typed questions with a probability per option, not with text) are in the library too. Start with the decision-models guide, or generate a request for one; this explainer explains the idea. They are left out of the rankings on purpose.
Models section
Tools