A practical, no-nonsense guide for founders, engineers, and AI teams deciding which open source or open-weight models are actually worth testing — by workload, benchmark profile, license fit, and hardware reality.
If you only need the short version, it is this: most teams evaluating open models in 2026 should begin with a strong 7B-32B text model, an explicit evaluation harness, and a clear hardware budget before they touch giant MoE systems, open video stacks, or "frontier" model marketing.
Qwen3.6, Llama 4, DeepSeek-V3.2, gpt-oss, and Gemma 4 26B MoE all use mixture-of-experts to deliver frontier quality at a fraction of the active-parameter cost. But dense models remain: Mistral Medium 3.5 is a 128B fully dense open-weight flagship, and Gemma 4 31B is also dense. Plan for MoE-friendly serving from day one — then check whether the model you're deploying actually uses it.
DeepSeek-V3.2 sits at GPT-5-class reasoning under MIT. Qwen3.6 and Mistral Large 3 compete with closed multimodal models. Mistral Medium 3.5 — a 128B dense open-weight model released May 22, 2026 — is now a credible flagship for coding and agentic work. OpenAI's gpt-oss runs on consumer hardware. For many real workloads, the open/closed gap has effectively closed.
License clarity matters at the checkpoint level, not the family level. Llama 4 is community-licensed with EU multimodal restrictions; Gemma 4 is Apache 2.0 but Gemma 3 used custom terms; FLUX.1-schnell is permissive but other FLUX variants are not. Voxtral TTS, released under CC-BY-NC 4.0 inherited from its reference voices, looks "open" but is restricted to non-commercial use — the same model-family-as-license shortcut that breaks here.
Native multimodal is now baseline. The 2026 step-change is omnimodal: Qwen3.5-Omni and MiniCPM-o 4.5 handle text, image, audio, and streaming speech in one model. Gemma 4 12B Unified (June 3, 2026) pushes further: an encoder-free design that unifies text, image, and audio into a single decoder-only stack. Grounded video understanding via Molmo 2 is also emerging as a distinct capability.
The right model is not the one with the biggest benchmark headline. It is the one that clears your real task evaluations, fits your hardware envelope, survives structured-output tests, and carries a license your company can live with.
This matrix is designed to help teams choose a sensible starting point instead of trying everything at once.
| Use case | Start here | Why | Hardware reality | Watch-out |
|---|---|---|---|---|
| All-in-one open-weight flagship (reasoning + coding + multimodal + agents) | Mistral Medium 3.5 | Dense 128B unifying instruction-following, reasoning, coding, and multimodality with a 256K context; strong on coding-heavy work like SWE-Bench Verified. | Self-hosting on as few as four GPUs per official Mistral announcement; datacenter-class for full throughput. | Open-weight under a modified MIT license, not OSI-style open source — confirm terms before standardizing. |
| General writing, chat, summaries, RAG | Qwen3.6-35B-A3B or Llama 4 Scout | Qwen3.6 is the current open-weight Qwen default: 35B total / 3B active, 262K native context, Apache 2.0. Llama 4 Scout offers strong multimodal performance on a single H100. | Qwen3.6-35B-A3B fits comfortably on 24–48GB VRAM with MoE serving. Llama 4 Scout fits a single H100 with INT4. | Confirm license fit before standardizing — Llama 4 has an EU multimodal restriction. |
| Coding assistant for real development work | Qwen3-Coder-Next | Purpose-built for coding agents: 80B total / 3B active, 256K context, Apache 2.0, optimized for long-horizon tool use. | Comfortable on 24GB VRAM thanks to MoE; Qwen recommends vLLM or SGLang for serving. | You still need tests, linting, sandboxing, and security review. |
| Tool-using agents and workflow automation | Qwen3.6-35B-A3B or Llama 4 Maverick | Native tool-calling support, strong structured output, broad ecosystem adoption. | 35B-A3B is the practical starting band; Maverick is datacenter-only at 400B total / 17B active. | Agent quality depends as much on orchestration and evals as on the base model. DeepSeek V3.2-Speciale drops tool calling — use standard V3.2 if you need agentic behavior. |
| Top-end open reasoning and large-scale inference | DeepSeek-V3.2 | Frontier-scale reasoning with integrated thinking and tool-use under MIT. Speciale variant is optimized for deep reasoning but drops tool calling. | Datacenter-class serving (685B parameters, DeepSeek Sparse Attention for long-context efficiency). | Too large for most local teams; API or cloud serving is the realistic path. |
| Omnimodal — speech + vision + text in one model | Qwen3.5-Omni | Native omnimodal model supporting text, image, audio, and audiovisual content. Strong audio-visual reasoning and streaming speech generation. | Datacenter-class for full quality; smaller real-time alternative is MiniCPM-o 4.5 (9B). | Production omni stacks need careful real-time orchestration; streaming speech latency is the main constraint. |
| Vision-language understanding | Gemma 4 (E4B, 12B Unified, or 26B MoE) or Qwen3.6-35B-A3B | Gemma 4 is Google's current open multimodal family with up to 256K context under Apache 2.0. The 12B Unified release (June 2026) adds an encoder-free multimodal design. Qwen3.6 handles vision natively. | Gemma 4 E2B/E4B run on consumer GPUs; 12B Unified and 26B MoE on prosumer hardware. | For true video grounding, pointing, and counting, prefer Molmo 2 — it is purpose-built for grounded visual reasoning. |
| Edge or low-footprint multimodal | MiniCPM-V 4.6 (1.3B) or Gemma 4 E2B | Edge-friendly visual models for phones, IoT, and lightweight servers. Apache 2.0. | Runs on consumer hardware with 4–8GB VRAM, sometimes CPU-only. | Capability ceiling is real — pick for footprint, not absolute accuracy. |
| Image generation | SDXL or FLUX.1-schnell | SDXL remains the mature ecosystem baseline. FLUX.1-schnell is the fast, permissive 12B alternative under Apache 2.0. | SDXL is happiest around 12GB VRAM; FLUX.1-schnell needs more for full quality. | FLUX licensing is checkpoint-specific — schnell is Apache 2.0, but other FLUX variants have stricter terms. |
| Image editing, control, inpainting | Diffusers + ControlNet + SAM 3.1 + LaMa | Editing is a stack problem, not a single-model problem. SAM 3.1 is Meta's current segmentation line, superseding SAM 2. | Consumer GPUs handle most workflows. | Workflow quality depends on masks, conditioning, and operator skill. |
| Speech recognition | Whisper large-v3 or Whisper turbo | Large-v3 remains the accuracy benchmark; turbo is the faster official derivative for streaming and low-latency use. Qwen3-ASR (Apache 2.0, 52 languages/dialects, forced aligner) is a newer open-source alternative worth evaluating for multilingual workloads. | Consumer GPUs handle both; large-v3 needs ~10GB VRAM. | Accuracy varies significantly across the long tail of languages — verify for your domain. Qwen3-ASR doesn't replace Whisper outright; test on your actual language mix. |
| Speech synthesis (TTS) | Qwen3-TTS (0.6B or 1.7B) or Chatterbox | Qwen3-TTS is Apache 2.0, multilingual, supports voice cloning, and has strong community adoption. Chatterbox (Resemble AI) offers Turbo (350M) for low-latency English and Multilingual V3 for broader language coverage. For high multilingual quality, Voxtral TTS (4B) is a strong open-weight option but carries CC-BY-NC weights. | Qwen3-TTS and Chatterbox run comfortably on consumer GPUs; viable for real-time agent voices. | Voice cloning has ethical and legal implications — confirm consent and disclosure rules. Voxtral TTS is source-available/non-commercial due to CC-BY-NC weight terms; confirm before any commercial use. |
| Open video experiments | Wan2.2 or Open-Sora 2.0 | Wan2.2 supersedes Wan2.1 with broader data and a 5B 720p TI2V path that is practical on strong consumer GPUs. Open-Sora 2.0 (11B, Apache 2.0) is the openly trained R&D counterpart for research into video generation economics and training transparency. | 16–24GB VRAM for comfortable workflows; the Wan2.2 5B TI2V model lowers the bar for high-resolution work. | Video quality, latency, and consistency remain uneven. Treat as R&D, not production-default. Verify exact LICENSE before commercial use. |
Choose the smallest model that reliably completes your real task with the right output shape. Then move up only if the gains are measurable.
This page separates fully open releases from open-weight releases and license-restricted releases because the market still collapses those categories into the same marketing label.
Weights, code, and meaningful training information are available. These releases are the closest match to the OSI-style vision of Open Source AI. Ai2's OLMo 3 leads this category with full model-flow traceability.
You can download and run the weights, but the full training data and recipe are not completely reproducible. This is where most high-performing "open" models sit today — including families under permissive licenses (Apache 2.0 / MIT) and those under community or modified licenses.
Some code and weights are available, but non-commercial clauses, behavioral restrictions, or geographic limits apply. Llama 4 carries an EU multimodal restriction; AudioCraft weights are CC-BY-NC; Voxtral TTS inherits CC-BY-NC 4.0 from its reference voices; some FLUX variants are non-commercial.
A shortlist of twelve models worth testing first, spanning text, code, multimodal, omni, speech, and video. License status is the first filter — read it before the parameter count.
| Model | License | Parameters | Key capabilities | Inference target | Maturity |
|---|---|---|---|---|---|
| Qwen3.6-35B-A3B | Apache 2.0 | 35B total / 3B active | General text, coding, agentic workflows, 262K context | Multi-GPU or strong prosumer for full context | Production-leaning |
| Qwen3-Coder-Next | Apache 2.0 | 80B total / 3B active | Coding agents, tool use, long-horizon reasoning, 256K context | Prosumer multi-GPU or quantized local | Production-leaning |
| Llama 4 Scout | Llama 4 Community License (open-weight) | 109B total / 17B active | Multimodal text+image, 10M context | INT4 on a single H100; else datacenter | Production-leaning |
| Gemma 4 31B / 12B Unified | Apache 2.0 | 30.7B dense or 11.95B unified | Text, coding, image; audio on 12B Unified | 12B on strong workstation; 31B multi-GPU | Production-leaning |
| DeepSeek-V3.2 | MIT | 685B total | Frontier reasoning and agentic behavior; V3.2-Speciale drops tool calling | Datacenter | Advanced / datacenter |
| gpt-oss-120b | Apache 2.0 + usage policy | 117B total / 5.1B active | Text reasoning, tool use, structured output | Single 80GB GPU | Production-leaning |
| Mistral Medium 3.5 | Modified MIT (open-weight) | 128B dense | Reasoning, coding, multimodal, agents | ~4-GPU self-hosting | Production-leaning |
| Molmo2-8B | Apache 2.0 (data-use caveats) | 8B | Image, video, grounding, pointing, tracking | Workstation / server | Research + production pilots |
| MiniCPM-o 4.5 | Apache 2.0 | 9B | Real-time omnimodal — text, vision, speech | Strong prosumer GPU; quantized local | Production-leaning for edge/omni |
| Whisper large-v3-turbo | MIT code; open weights | 0.8B | ASR, low-latency transcription | CPU or modest GPU (~6GB VRAM) | Mature / production standard |
| Qwen3-TTS 1.7B / 0.6B | Apache 2.0 | 1.7B or 0.6B | Multilingual TTS, voice cloning, streaming | Consumer / prosumer GPU | Production-leaning |
| Wan2.2 TI2V-5B | Verify repo LICENSE | 5B (within larger Wan2.2 family) | Text/image-to-video generation | High-end consumer GPU to workstation | Advanced prototyping / R&D |
"Open source" here means OSI-style permissive terms (Apache 2.0 / MIT). "Open-weight" means the weights are downloadable but the license is a community, usage-policy, or modified license — workable for most teams, but not the same legal object. Confirm the exact model card before you standardize.
Open AI is no longer one category. The ecosystem now includes general-purpose LLMs, code-specialized models, multimodal models, image generators, image-editing stacks, and increasingly capable video systems.
| Category | Best for | Start here | Move up to | Sweet spot | Watch-outs |
|---|---|---|---|---|---|
| Writing / general LLM | Chat, drafting, summarization, RAG, internal copilots | Qwen3.6-35B-A3B or Llama 4 Scout | Qwen 3.5 397B-A17B, Llama 4 Maverick, Mistral Medium 3.5, Mistral Large 3, DeepSeek-V3.2 | Qwen3.6-35B-A3B or gpt-oss-20b covers most team needs without datacenter overhead. | License terms (Llama 4 EU multimodal limits) and quantization quality matter more than leaderboard hype. |
| Coding | PR assistance, code generation, refactors, test writing, coding agents | Qwen3-Coder-Next | Qwen3-Coder-480B-A35B, Mistral Medium 3.5 (128B dense, strong SWE-Bench Verified results), Mistral Large 3, or DeepSeek-V3.2 | Qwen3-Coder-Next (80B total / 3B active, 256K context) is the most practical for real developer use. | Do not deploy without tests, sandboxing, and dependency/security review. |
| Agents | Tool use, workflow automation, multi-step task execution | Qwen3.6-35B-A3B or Llama 4 Scout | Llama 4 Maverick, Mistral Large 3, Mistral Medium 3.5 (capable all-in-one for reasoning + tool use + multimodality on four GPUs), or DeepSeek-V3.2 | Smaller models plus strong orchestration often beat oversized models with weak tooling. | JSON breakage, tool misuse, and cascading failures are the real bottlenecks. Avoid DeepSeek V3.2-Speciale here — it drops tool calling. |
| Multimodal / VLM | Document understanding, image Q&A, visual agents, OCR-heavy workflows | Gemma 4 (E4B or 26B MoE) or Qwen3.6-35B-A3B | Qwen 3.5 397B-A17B or Llama 4 Maverick | Gemma 4 26B MoE is the most practical local starting point with up to 256K context. | Grounding mistakes and OCR hallucinations still require checks; for true visual grounding, prefer Molmo 2. |
| Omnimodal (text + vision + speech) | Voice-first assistants, real-time audio-visual reasoning, streaming speech generation | MiniCPM-o 4.5 (9B) for local real-time | Qwen3.5-Omni for full audio-visual reasoning at datacenter scale; Gemma 4 12B Unified for an encoder-free architecture that handles text, image, and audio in one decoder-only stack | MiniCPM-o 4.5 gives most teams a real-time omni model that fits on prosumer hardware. | Some serving stacks still need patched support; streaming speech latency is the binding constraint. |
| Video grounding / pointing | Video understanding, object pointing, tracking, multi-image reasoning | Molmo 2 (4B, 8B, O-7B variants) | Custom Molmo 2 fine-tunes on domain data | One of the few open families in 2026 with genuine video grounding and pointing under Apache 2.0; note the third-party dataset restrictions in the model card. | Not a general chat model — pair with a chat-capable LLM for conversational interfaces. |
| Edge multimodal | Phones, IoT, lightweight servers, on-device VLM workloads | MiniCPM-V 4.6 (1.3B) or Gemma 4 E2B | MiniCPM-o 4.5 or Gemma 4 E4B when more capability is required | MiniCPM-V 4.6 punches above its weight for visual tasks on 4–8GB devices. | Capability ceiling is real — do not expect 27B-class reasoning at 1.3B. |
| Image generation | Concept art, marketing assets, ideation, product visuals | SDXL | FLUX.1-schnell (Apache 2.0) or other FLUX variants when licensing permits | SDXL remains the safest default for ecosystem compatibility; FLUX.1-schnell when you need speed and a permissive license. | Typography and exact prompt fidelity still need workflow iteration. FLUX licensing is checkpoint-specific. |
| Image editing | Inpainting, control, masking, pose/depth guidance, product edits | ControlNet + SAM 3.1 + LaMa + Diffusers | Project-specific editing stacks with custom masks and pipelines | Editing quality comes from stack design, not one magic checkpoint. | Commercial rights differ across base checkpoints and extensions. |
| Speech recognition | Transcription, translation, voice interfaces, audio understanding | Whisper large-v3 | Whisper turbo for faster streaming; Qwen3-ASR (open-source, 52 languages/dialects) as a newer alternative; or Qwen3.5-Omni for unified audio + text reasoning | Whisper large-v3 covers most transcription and translation needs; turbo is the official low-latency derivative. | Accuracy varies significantly across the long tail of languages — verify for your domain. |
| Speech synthesis (TTS) | Voice agents, narration, dubbing, expressive synthesis | Qwen3-TTS (Apache 2.0, 0.6B or 1.7B, multilingual, voice cloning) or Chatterbox-Turbo (350M) / Chatterbox-Multilingual V3 (500M) | Voxtral TTS (4B, CC-BY-NC — non-commercial only) for top multilingual quality and voice adaptation | Qwen3-TTS is the permissive default; step up to Voxtral TTS only when multilingual quality is the priority and non-commercial terms are acceptable. | Voice cloning has ethical and legal implications — confirm consent rules before deploying. Voxtral TTS is CC-BY-NC 4.0: non-commercial use only. |
| Video generation | Short exploratory clips, motion concepts, early creative prototyping | Wan2.2 TI2V-5B or smaller Wan2.2 variants | Larger Wan2.2 family extensions or Open-Sora 2.0 (11B, Apache 2.0) | Today, open video is a prototyping tool more than a production default. | Temporal flicker, identity drift, and long render times remain common. Verify the repo LICENSE before commercial use. |
| Video editing | Interpolation, inpainting, retiming, experimental edit pipelines | RIFE, ProPainter, Wan2.2 VACE-style workflows | Custom pipelines for domain-specific video tasks | Use specialized tools rather than expecting one general model to handle everything. | Workflow complexity is high; results are sensitive to clip quality and masking. Some tools (e.g., ProPainter) ship under research-only S-Lab terms — verify license fit. |
The biggest change from 2024 to 2026 is not just raw model quality. It is the breadth of credible open options across text, coding, multimodal, and media generation.
Short, opinionated notes on the families that matter most right now — what each is genuinely good at, and where it stops being the right tool.
Released May 22, 2026, Mistral Medium 3.5 is the most significant post-snapshot addition to this page. It is a dense 128B model with a 256K context window that combines instruction-following, reasoning, coding, and multimodal input in a single set of weights — and Mistral's official materials claim practical self-hosting on as few as four GPUs. Use it when you need a unified open-weight option that does not force you to choose between chat quality and code quality. The license is a modified MIT with open weights, not OSI-style open source, so read the exact terms before standardizing on it.
The current default open-weight pick for teams that want a capable, well-documented model without immediately jumping to datacenter-scale MoE. Official materials describe 35B total parameters, only 3B active at inference, and a native 262,144-token context window. Its editorial value is balance: large enough to be useful on real work, much easier to reason about operationally than 400B+ families. Use it as a serious private-assistant or internal-agent baseline. Do not treat it as a tiny local model — it needs real serving infrastructure and benefits from vLLM or SGLang.
When the workload is repository-aware coding, debugging, or scaffolded tool use, Qwen3-Coder-Next is the right specialist, not a generic chat model. It is an 80B total / 3B active MoE model with a 256K context window and direct support for SGLang and vLLM. The active-parameter count matters here: 3B active means the inference footprint is manageable even though the total weight is 80B. Verify the exact license in the official model card before deploying commercially — retrieved excerpts do not surface it directly.
DeepSeek-V3.2 is released under MIT at 685B parameters — one of the strongest frontier-scale reasoning and agentic families available outside closed API silos. One distinction the article must make explicit: V3.2 and V3.2-Speciale are not interchangeable. Speciale is optimized for deep reasoning and drops tool-calling support, so teams building agentic workflows should use standard V3.2. Neither variant is practical for most local teams; API or cloud serving is the realistic deployment path. The MIT license is genuinely permissive — rare at this scale; the limit here is training transparency, not usage rights, which is why it sits in the open-weight tier rather than fully open.
Scout is a 109B total / 17B active multimodal model with a 10M-token context window and the ability to run INT4-quantized on a single H100. That is useful operational guidance and one of the reasons it belongs on this list. What it is not is open source — Scout is released under the Llama 4 Community License, which carries EU multimodal restrictions and acceptable-use limits that Apache-2.0 does not. Confirm those terms before treating Scout as a standard open-weight default, particularly for EU deployments.
Gemma 4 is the clearest Apache 2.0 recommendation across the multimodal tier, spanning E2B through 31B dense with up to 256K context. The June 3, 2026 addition of 12B Unified changes the editorial story: it is an encoder-free design that handles text, image, and audio in a single decoder-only stack, which makes it editorially distinct from the other size variants. For most readers, Gemma 4 is not the highest raw performer, but it is the most straightforward family to recommend when license simplicity and deployment flexibility are the primary constraints.
The gpt-oss family fills a narrow but important role: strong text-only open-source reasoning with explicit tool-use and structured-output support. Both models are released under Apache 2.0 plus a separate usage policy — read both documents, not just the license header. The 120b is 117B total / 5.1B active and targets a single 80GB GPU per the official HF page; the 20b is 21B total / 3.6B active and fits in 16GB. Present these as leading reasoning-focused open-source models, not general multimodal replacements — they are text-only.
Molmo 2 is the rare fully open family that advances grounded video understanding rather than generic chat. Ai2 open-sourced the full codebase in March 2026, and the family includes Apache 2.0 8B and 4B Qwen-based variants plus a 7B OLMo-backed variant for teams that care about full-stack openness. Note that the Apache 2.0 terms include third-party dataset restrictions — worth reading before commercial use. Use it for perception-and-grounding workflows — image, multi-image, video, pointing, tracking. Pair it with a separate conversational model if you need a polished assistant interface; Molmo 2 is a specialist, and stretching it into a chat model produces predictably worse results.
At 9B parameters combining SigLip2, Whisper-medium, CosyVoice2, and a Qwen3 backbone, MiniCPM-o 4.5 occupies the narrow band where real-time omni interaction on serious local hardware is actually achievable. Full-duplex streaming, strong visual and speech capabilities, and official support for quantized local deployment via llama.cpp and Ollama make it the right pick for omni experiments on prosumer GPUs. The distinction between "possible to run" and "easy to productize" still applies — streaming speech latency and serving stack compatibility are the primary constraints before treating this as a production default.
Whisper remains the most stable and widely deployed ASR standard in practice, and both variants belong in the article. Large-v3 is the quality-first default at roughly 1.55B parameters and 10GB VRAM; turbo is the 809M optimized derivative with roughly 6GB VRAM and materially faster throughput for streaming and low-latency use. The known risks are worth stating plainly: hallucinated text, uneven performance across the long tail of languages, and weaker translation suitability for turbo. Whisper code is MIT-licensed; the model weights are distributed via OpenAI's official channels.
Qwen3-TTS is the open-source end of the TTS spectrum: Apache 2.0, multilingual, with 0.6B and 1.7B variants and voice cloning support. The GitHub repo has attracted meaningful community uptake. Use it when permissive licensing for downstream commercial work matters, or when you want a TTS model you can actually inspect, fine-tune, and deploy without non-commercial restrictions. Voice cloning carries consent and misuse risks regardless of the technical license — build those guardrails into the application layer, not as an afterthought.
Voxtral TTS is a 4B open-weight Mistral release with strong multilingual voice generation and voice adaptation quality. The critical caveat is that the published model inherits CC BY-NC 4.0 because of the reference voices it includes — making it source-available and non-commercial, not permissively open. The right use is evaluation, research, and internal tooling where non-commercial terms are acceptable. For production commercial deployments needing TTS, Qwen3-TTS is the cleaner license choice.
Wan2.2 replaces Wan2.1 as the current recommendation in the Wan family. The official repo positions it as a significant upgrade: broader data scale, MoE-informed innovations, and a 5B 720p text-to-image-to-video model intended to be practical on strong consumer hardware. That is meaningful progress relative to where open video was a year ago, but the category framing should not change — temporal flicker, identity drift, and long render times remain real problems. Treat Wan2.2 as a prototyping and creative R&D tool. Verify the LICENSE file for the specific checkpoint before commercial use.
Some of the most useful tools in open AI are not foundation models. They are conditioning architectures, segmentation models, and workflow utilities that sit on top of generators. They do not belong in a head-to-head leaderboard with Qwen, Gemma, or DeepSeek, but they are essential to most real production pipelines.
Control architecture for Stable Diffusion-class workflows. Adds pose, depth, edge, and segmentation conditioning to diffusion image generation. A workflow component, not a standalone foundation model.
License: Apache 2.0 code; OpenRAIL weight distribution
Meta's current promptable segmentation line, succeeding SAM 2 (which remains widely used in existing pipelines). Used as a pipeline primitive for masking, editing pipelines, and visual agents — not a top-level answer to "best vision model." SAM 2 is still the more common integration target given ecosystem maturity, but new pipelines should track SAM 3.1.
License: Apache 2.0-style — verify the current SAM 3.x model card
Local video workflow tool for next-frame prediction inference. Helps produce long, consistent video clips on constrained consumer hardware.
License: Stable version unspecified — verify repo LICENSE before commercial use
Resolution-robust inpainting model. Best for clean object removal and background reconstruction inside larger editing stacks.
License: Apache 2.0
Frame interpolation model for smooth slow motion and frame-rate upscaling. The official repo reports ~30+ FPS on 2× 720p on a 2080 Ti; the 4.22.lite variant is well-suited for diffusion-video post-processing where a lighter interpolation pass is enough. Pairs well with video generation pipelines.
License: MIT (code) — verify weight distribution
Mask-aware video inpainting and object removal. Specialized for video editing; not a general-purpose video tool. S-Lab research terms are a hard stop for most commercial production uses — confirm the license before building on it.
License: S-Lab License (research / non-commercial — verify before production)
Use these components alongside foundation models, not instead of them. ControlNet adds conditioning to image generation; SAM 2 adds grounding to vision pipelines; FramePack, RIFE, and ProPainter extend video workflows.
These numbers are useful as a map, not as a verdict. Benchmark settings vary. Prompt formatting moves scores. Preference benchmarks can overstate real operational reliability. Use this as the first filter, then test on your own workload.
| Model | General | Reasoning | Coding | Notes |
|---|---|---|---|---|
| Qwen3.6-35B-A3B | Current Qwen open-weight default | Improved multimodal and agentic behavior vs. Qwen 3.5 | Stronger real-world coding than the Qwen 3.5 baseline | 35B total / 3B active MoE. 262,144 native context. Apache 2.0 (verify model card before publishing). Released April 14, 2026. |
| Qwen 3.5 397B-A17B | MMLU-Pro 87.8, SuperGPQA 70.4 | AIME26 91.3, GPQA Diamond 88.4 | LiveCodeBench v6 83.6 | Frontier MoE with only 17B active params. Native multimodal, 201 languages. Numbers from Qwen 3.5 announcement materials. |
| DeepSeek-V3.2 | GPT-5-class open-weight reasoning | Frontier-scale math and reasoning; Speciale variant is reasoning-specialized but drops tool calling | SWE-bench competitive with closed frontier models | 685B parameters under MIT License. DeepSeek Sparse Attention for long-context efficiency. V3.2 and V3.2-Speciale are not interchangeable — use standard V3.2 for agentic workflows. |
| Mistral Medium 3.5 | New open-weight flagship; strong all-round | Combined instruction-following and reasoning in one set of weights | Strong on coding-heavy work including SWE-Bench Verified (~77.6%) | 128B dense. 256K context. Open weights under a modified MIT license — not OSI-open. Self-hosting on as few as four GPUs per Mistral. Released May 22, 2026. |
| Mistral Large 3 | Strong general-purpose multimodal model | Mid-to-high tier reasoning | Competitive on standard coding benchmarks | 675B total / 41B active MoE. 256K context. Multimodal. Open-weight — verify the Mistral model card for exact terms. Realistically datacenter-class (8×A100/H100-class deployment). |
| Llama 4 Maverick | MMLU 85.5, MMLU-Pro 80.5 | GPQA Diamond 69.8 | HumanEval 82.4 | 400B total / 17B active, 128 experts. Natively multimodal, 1M context. Open-weight under Llama 4 Community License — not OSI-style open source. |
| Llama 4 Scout | MMLU 79.6, MMLU-Pro 74.3 | GPQA Diamond 57.2 | HumanEval 74.1 | 109B total / 17B active, 16 experts. 10M context. Fits on a single H100 with INT4. Open-weight under Llama 4 Community License — not OSI-style open source. |
| gpt-oss-120b | MMLU-Pro 90.0 | AIME 2025 97.9 (with tools) | Near o4-mini on competition coding | 117B total / 5.1B active MoE. Apache 2.0 with usage policy. Fits on a single 80GB GPU. 128K context. |
| gpt-oss-20b | Matches o3-mini on common benchmarks | Strong for its size class | Competitive with o3-mini | 21B total / 3.6B active MoE. Apache 2.0 with usage policy. Runs on 16GB devices. 128K context. |
| Gemma 4 | Current Google flagship open family | Improved over Gemma 3 across reasoning benchmarks | Strong for size class | Apache 2.0. Family spans E2B (2.3B effective), E4B (4.5B), 12B Unified (11.95B), 26B A4B MoE (25.2B total / 3.8B active), and 31B dense (30.7B). Up to 256K context. Text and image across the family; audio on E2B, E4B, and 12B Unified. The 12B Unified (June 3, 2026) is encoder-free and unifies text, image, and audio in a single decoder-only stack. Family launch Mar. 31, 2026. |
| OLMo 3 32B-Think | Leading fully-open reasoning model | Strongest open-traceable reasoning in late-2025/early-2026 | Competitive among fully-open releases | Released late 2025 (paper posted Dec. 15, 2025). Full model-flow traceability and open training recipe. The benchmark for fully-open transparency — not the highest raw capability, but the most auditable stack available. |
Use one academic snapshot table, one real-work evaluation table, and one reliability table. If a model only looks good in one of those three, it is not production-ready for your team.
The real tradeoff is not "open is better" or "closed is better." It is whether you want control, customization, and privacy enough to take on the systems burden yourself.
| Dimension | Open / open-weight | Closed ecosystem |
|---|---|---|
| Control | Self-host, fine-tune, inspect, and route however you want. | Fastest path to strong capability with less systems work. |
| Cost model | Infrastructure, ops, and engineering replace per-token API pricing. | Usage-based pricing is simple but can become expensive at scale. |
| Privacy and data boundary | Best option when prompts, outputs, and logs must stay inside your environment. | Provider policy and retention controls matter more. |
| Customization | Adapters, quantization, routing, and domain tuning are the major advantages. | Prompting is easy; deep model customization is limited. |
| Operational burden | You own serving, evals, security, and reliability. | You inherit better managed infrastructure and usually better SLAs. |
| Best fit | Teams with repeatable workloads, privacy needs, or platform ambitions. | Teams optimizing for speed, simplicity, and managed frontier access. |
The fastest way to waste time in open AI is to choose models before you define the serving envelope. Pick the hardware tier first, then shortlist models that fit. If you are buying a dedicated box to run these models locally, our best local LLM machines guide compares Mac Studio, DGX Spark, Framework, HP Z2 and more by memory, bandwidth, and software stack.
What fits: Gemma 4 E2B/E4B, MiniCPM-V 4.6, gpt-oss-20b, smaller Qwen3.6 variants, lightweight coding models, SDXL, Whisper turbo (~6GB VRAM), Qwen3-TTS 0.6B/1.7B
Best for: Local testing, lightweight RAG, first agents, edge multimodal, image generation, on-device ASR and TTS
Watch-outs: Do not expect comfortable 35B+ MoE serving or serious open video production.
What fits: Qwen3.6-35B-A3B, Qwen3-Coder-Next, Gemma 4 26B MoE / 31B, gpt-oss-120b, MiniCPM-o 4.5 (9B), Voxtral TTS (4B), Wan2.2 TI2V-5B, small video stacks
Best for: Serious private assistants, agentic coding, local omni experiments, MoE serving, open-weight video generation prototyping
Watch-outs: Open video is still slow and multi-step agent stacks need careful tuning.
What fits: Qwen 3.5 397B-A17B, Llama 4 Maverick (single H100 DGX node), Mistral Large 3 (8×A100/H100), Mistral Medium 3.5 (four-GPU minimum for self-hosting), DeepSeek-V3.2, Qwen3.5-Omni, larger Wan2.2 variants
Best for: Internal copilots, agent platforms, omnimodal services, governed deployment
Watch-outs: Reliability, governance, and evaluation matter more than raw model choice at this tier.
For most real teams, the 14B-32B band is the easiest place to get strong quality without crossing into difficult multi-GPU operations. Giant MoE systems make sense later, not first.
Hallucinations are only one part of the reliability story. Open models also fail through prompt sensitivity, poor tool arguments, visual grounding errors, license misunderstandings, and brittle long-context behavior.
The most common failures are fabricated facts, false confidence, stale knowledge, malformed JSON, and plausible-but-wrong code. Code models can also generate insecure or license-sensitive output.
Expect OCR misses, object misidentification, incorrect grounding, and overconfident descriptions of partially visible content.
The main problems are prompt drift, poor typography, inconsistent identity, and weak fine-grained control unless you add editing and conditioning tools.
The biggest issues remain temporal flicker, identity drift, motion incoherence, and long runtimes for short clips.
License fit is not cleanup work after the benchmark review. It is one of the first filters. Many teams waste time evaluating models they cannot legally or economically ship.
| License pattern | Best for | Examples | Watch-out |
|---|---|---|---|
| Apache 2.0 / MIT | Commercial deployment and broad integration | OLMo 3, Whisper code, Qwen3.6, Qwen3-Coder-Next, Qwen3-TTS (0.6B/1.7B), Qwen3-ASR, gpt-oss (with usage policy), Gemma 4, FLUX.1-schnell, MiniCPM-V 4.6, MiniCPM-o 4.5, Molmo 2 (training data restrictions apply — check model card), ControlNet code, SAM 2 (SAM 3.1 is the newer line — verify its card), DeepSeek-V3.2 (MIT). Mistral Medium 3.5 ships open weights under a modified MIT license — useful for most commercial work, but not OSI-open; treat it as open-weight, not fully permissive. | Verify each model card — checkpoint-level terms can differ even when the family is described as permissive. Qwen sub-model licenses should be confirmed directly in each official repo. |
| Llama 4 Community License | Commercial use with strong ecosystem momentum | Llama 4 Scout, Llama 4 Maverick | Permissive for many uses, but not OSI-style open source. The policy includes an EU restriction for multimodal use and broader acceptable-use limits. |
| Gemma terms / custom terms (legacy) | Practical use of older Gemma generations | Gemma 3 and earlier | Gemma 4 moved to Apache 2.0 in April 2026, but Gemma 3 and earlier remain under Google's custom terms. The family name alone is not a license signal. |
| OpenRAIL / Responsible AI licenses | Creative or research use where behavioral restrictions are acceptable | SDXL (CreativeML Open RAIL++-M), ControlNet weight distributions, BigCode OpenRAIL-M | Behavioral restrictions and downstream obligations can affect productization. |
| Community / revenue-threshold licenses | Early testing before full commercialization | Some Stability releases | Revenue thresholds and enterprise terms can change the total cost of ownership. |
| Non-commercial weight licenses | Research, experimentation, internal evaluation | Non-schnell FLUX variants, AudioCraft weights (CC-BY-NC 4.0), Voxtral TTS weights (CC-BY-NC 4.0, inherited from included reference voices), some video editing tools (e.g., ProPainter under S-Lab research terms) | This is a hard stop for many production uses. Verify the LICENSE file in every video/tooling project before publishing legal language. |
Treat every checkpoint as its own legal object. Do not assume the family name tells you the full commercial story.
Choosing a model without choosing a serving and evaluation stack is incomplete. The stack determines latency, batching, observability, and how painful future model swaps will be.
Best for: Fastest path to local testing
Strengths: Great for laptops, desktops, and quick internal prototypes.
Limits: Not the best fit for serious multi-user production serving.
Best for: High-throughput production inference for MoE models
Strengths: Paged attention, strong batching, and a mature serving ecosystem. SGLang is explicitly recommended by Qwen for Qwen3-Coder-Next.
Limits: More ops-heavy than local tools.
Best for: NVIDIA-centric optimized serving
Strengths: Best when you want GPU-specific performance tuning at scale.
Limits: More specialized setup and infra assumptions.
Best for: Custom workflows and research flexibility
Strengths: Best ecosystem for model experimentation, adapters, and editing pipelines.
Limits: Requires more assembly than end-user desktop tools.
Best for: Creative image and video workflows
Strengths: Visual pipeline building, strong community extensions, easy iteration. The standard path for Wan2.2 and open-video generation workflows.
Limits: Operational governance is weaker than code-first stacks.
Best for: Agents, tool use, and workflow orchestration
Strengths: Useful abstractions for state, retrieval, and multi-step execution.
Limits: They do not fix weak evals or poor model choices for you.
Use these as default launch points, not as permanent architecture decisions.
Start with Qwen3.6-35B-A3B or Llama 4 Scout, run it through a small RAG layer, and measure task completion before chasing bigger models.
Start with Qwen3-Coder-Next (80B total / 3B active, Apache 2.0), then step up only if your evals show clear gains on your real repos.
Use SDXL or FLUX.1-schnell for image work first. Treat open video (Wan2.2, Open-Sora 2.0) as an R&D lane, not your default production pipeline. ComfyUI is the standard entry point for Wan2.2 and most open video workflows.
Start with MiniCPM-o 4.5 for local real-time omni, Whisper large-v3 or turbo for ASR (or Qwen3-ASR if you need broader language support), and Chatterbox or Qwen3-TTS for TTS. Voxtral TTS offers high-quality multilingual synthesis but carries a non-commercial CC-BY-NC license — confirm that constraint fits your use case before committing. Step up to Qwen3.5-Omni when you need datacenter-class audio-visual reasoning.
Prioritize license clarity (per-checkpoint, not per-family), eval discipline, and serving fit over raw leaderboard rank. vLLM-class serving plus Qwen3.6-35B-A3B or Llama 4 Scout (mind the EU multimodal restriction) is usually the right first step.
Pick one workflow, one evaluation harness, one hardware target, and three candidate models. Anything broader becomes expensive research theater.
This page is built from model cards, technical reports, official repositories, standards bodies, and tooling documentation. The goal is practical decision support, not hype-driven ranking.