Open Source & Open-Weight AI Models: A Practical Guide

Fully open

The OSI Open Source AI Definition calls for the data information, code, and parameters needed to study and modify a system. It does not simply require an Apache label or necessarily publication of every raw training record. OLMo is useful here because Ai2 publishes substantial training artifacts alongside its weights.

Open weight

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 / OpenMDW) and those under community, modified, or revenue-triggered licenses.

Source-available or restricted

Custom terms vary considerably. Llama 4 has a community license and separate acceptable-use policy; Muse Spark Open has its own Muse Model License. Qwen3.8-Flash-Next requires a separate commercial license for certain AI businesses regardless of revenue; GLM-5.3 has a much higher, narrowly defined security-review threshold. Non-commercial weight licenses such as Voxtral TTS and Qwen-Image 2.1 are a separate constraint.

Announced, not released

A hosted model announcement does not establish that a usable checkpoint exists. Check for actual weight files, a model card, and a license before adding a model to a self-hosting plan. Muse Spark Open cleared that check on October 15, and Mistral Large 4 on October 29, when FP8 and NVFP4 checkpoints appeared on Hugging Face with a license file. Alibaba’s September 22 roadmap describes Qwen 4 as in training, with no weights.

Kolibri-1

Aleph Alpha’s model card describes a 78B MoE with 3.46B active parameters per token. Its checkpoint license is Apache 2.0. The card lists a 1,048,576-token maximum but recommends no more than 262,144 for serving efficiency and complex tasks. The roughly 78 GB FP8 weight footprint excludes runtime allocations; active parameters do not make this a 3.46B-sized download. Evaluate German and English tasks separately on your intended serving setup.

Qwen3.8-27B

Qwen’s card describes a 27B language model with image and video input and 262,144 native context. The Unsloth UD-Q4_K_M language-model download is 16.46 GB, plus a separate vision projector. That makes modest-context testing on a 24 GB card plausible, but context and runtime settings still determine fit. Qwen exposes low, medium, and xhigh reasoning effort; xhigh is the default. Start with a budget that matches the task. Our launch analysis provides the earlier release context.

Muse Glimmer-30B

Meta publishes a 16.76 GB Q4 language-model file, a 1.40 GB vision file, and an optional 1.63 GB DFlash drafter. Count the components you enable rather than treating “17 GB” as total runtime memory. The model accepts text and images, with a 131,072-token context specified in the card. Apache 2.0 is accompanied by a usage policy. Our Glimmer article covers the launch.

Nemotron 3.5 Bolt and Super VL

NVIDIA compressed Bolt from the 120B / 12B-active Nemotron 3.5 Super VL, both released October 15 under OpenMDW-1.1. Bolt has 74B total and 9B active parameters; its native 2-bit build targets a footprint under 24 GB at batch size 1 and 128K context. NVIDIA’s published GGUF recipes ask for more: 28 GB at 65,536 context with a q8_0 cache (for a 64 GB RTX Spark or a 32 GB+ GPU), or 32 GB at 262,144 (a 128 GB RTX Spark or a 48 GB+ GPU). The card’s comparisons run Bolt at 2 bits against BF16 peers, and NVIDIA’s speed figures come from a single 128 GB Surface Laptop Ultra running llama.cpp. A smaller sibling, Bolt E4B, arrived on Hugging Face on October 23: an 8B (4B effective) model post-trained from Gemma 4 E4B, under OpenMDW-1.1 with the Gemma base under Apache 2.0. NVIDIA’s card lists text input and a 64K context for its 6.18 GB Q4_K_M file, while Ollama’s v0.40.10 package lists image input and 128K, so check which build you are running. Our Bolt and RTX Spark article covers the hardware side.

Recursive Flint

Recursive released Flint on October 21: 39.8B total parameters with about 2.5B active (2.2B without the output head) and a 262,144-token context. The weights are Apache 2.0; the card says hosted API terms are separate. Recursive reports 54.9% on Terminal-Bench 2.1 and 74.2% on SWE-bench Verified. Its 89.7% on “SWE-bench Pro v2” comes with no definition of that benchmark, so it cannot be compared with other SWE-bench Pro results. Recursive documents Ollama’s flint:40b-a2b-qat-int4 tag, a 23 GB download; the BF16 weights are about 80 GB. Memory guidance depends on the runtime: Recursive’s rec engine requires an M4 Pro, M5 Pro or M5 Max with at least 48 GB, while LM Studio lists 24 GB. Recursive’s throughput figures come from its own engine. Our Flint article covers the setup and the open questions.

LFM2.5-24B-A2B

Liquid’s card describes 24.0B total and 2.45B active parameters, 64 experts per MoE block with top-4 routing, and a 131,072-token context. It always reasons before answering, and Liquid recommends a low temperature (0.1). Its own table shows a clear AutomationBench lead over Qwen3.6-35B-A3B (27.69 vs 8.72) and an effective tie on BFCLv4 (65.01 vs 65.00), while Qwen leads on GPQA Diamond, SWE-bench Verified and PinchBench. The LFM Open License v1.0 does not license commercial use at $10M annual revenue or more. Our LFM2.5 article covers the license in more detail.

Qwen3.8-Flash-Next

The card separates a 125B backbone with 6B active parameters from 51B of n-gram embeddings and 4B of multi-token-prediction weights. The full checkpoint is roughly 180B parameters. It supports image and video input with 262,144 native context. Its new Community License 1.0 requires a separate license for commercial use by specified AI businesses regardless of revenue, with an internal-use exception. That makes it a different procurement decision from Qwen3.8-27B.

GLM-5.3 and GLM-5.3-Flash

GLM-5.3’s weights have been on Hugging Face since August 25. The published FP8 repository is roughly 753B parameters. GLM-5.3-Flash instead uses a new 320B / 18B-active multimodal architecture and MIT terms. Both cards document low, high, and max reasoning effort; max is the default. Use the matching serving recipe and template. The Flash name describes a different efficiency tradeoff, not a desktop-sized model.

DeepSeek-V4.1-Flash

The MIT-licensed checkpoint accepts images and text and generates text with up to 1M context. DeepSeek separates its 552B backbone from 196B of Engram conditional memory, with 8B active parameters during prefill and 16B during decode. The backbone count is not the complete checkpoint or runtime memory requirement. Evaluate the matching serving implementation and all required components. DeepSeek withdrew the planned September 14 Pro migration and continues to offer deepseek-v4-pro separately. The deepseek-flash API now uses V4.1-Flash-1017, released October 17. For 1017, use the matching Python encoding reference or deepseek-recipe toolkit; the card supplies no Jinja chat template. Inference, tool execution and HTTP transport remain the caller’s responsibility. Our release analysis covers the API migration; changes to hosted aliases do not remove the older downloadable V4 checkpoints below.

DeepSeek-V4 (Flash-0731 and Pro-0813)

The MIT-licensed V4-Flash-0731 and V4-Pro-0813 releases remain larger-model candidates. DeepSeek’s original V4 card describes 284B / 13B-active and 1.6T / 49B-active architectures with 1M context; the suffixed cards do not restate those figures. Use the complete suffixed repository name to distinguish them from preview artifacts. Published agent scores are tied to DeepSeek Harness. Compare the runtime recipe and actual checkpoint size before deciding whether to self-host; hosted API prices and features are a separate decision.

Muse Spark Open

Meta released the 816B / 42B-active model on October 15 under the Muse Model License, which it describes as having limited commercial terms. The card lists text, image, video and audio input with text output and a 1M context. Products must display “Built with Muse Spark Open”, and a licensee offering models as a service must notify Meta and sign a separate agreement once its aggregate gross revenue, with affiliates, exceeds $20M in any consecutive 12-month period. Internal use and features of the licensee’s own products are exempt. The download is gated. Meta’s guidance is at least 700 GB of VRAM for NVFP4 and 2 TB for BF16; it also says the model is optimized for two M5 Ultra Mac Studios with MLX, which S5 Labs has not tested. Apple puts the NVFP4 parameters at about 493 GB and says the model does not fit on one 512 GB Mac Studio. Meta calls it “the most capable US open-weights model to date”; its own table shows Kimi K3, GLM-5.3 and Qwen3.8-2.4T ahead on Terminal Bench 2.1. NVIDIA has published a 4.38B DFlash2 drafting model for it under separate NVIDIA terms. Our Muse Spark Open article covers hosted prices and Apple’s two-Mac setup.

MiniMax M3.1 Flash

MiniMax’s README describes a native multimodal MoE with about 428B total and 23B active parameters and a context of up to 1,000,000 tokens. The weights are public in the MiniMaxAI Hugging Face organization, and MiniMax’s release notes call it open-weight, yet its October 19 launch blog says weights “will be released later”. MiniMax has not explained the difference. The MiniMax Community License grants non-commercial use. Commercial products must display “Built with MiniMax M3.1-Flash” and send MiniMax a one-time notice, or obtain prior written authorization instead if the products or services generate more than $20M in yearly revenue. Its prohibited uses include any military purpose. MiniMax states no disk size or memory requirement. Our MiniMax M3.1 Flash article covers the hosted pricing.

Step-5-Preview

StepFun’s agentic MoE has 600B total and 27B active parameters, a 1M context, and FP8, BF16 and NVFP4 repositories. The FP8 repository is about 650 GB; StepFun’s deployment table puts it on one four-GPU B300 node or one eight-GPU H200 node. StepFun calls the license Apache-2.0-style, but it adds conditions: “Powered by StepFun AI” above 100M active users or $20M revenue in a calendar month, and separate agreements for device makers and for model-as-a-service businesses above $20M combined revenue in any 12 months. Internal use and StepFun-certified distribution channels are exempt. The card says its BF16 Transformers workflow has not been validated end to end.

Beam 1.0

Reflection’s 501.3B / 23.4B-active text model is on Hugging Face under Apache 2.0, ungated, with BF16, FP8 and NVFP4 releases and a 1M context. Reflection announced the public weights release on October 16. Our October 5 article predates the weights. The card’s scores use temperature 0.7 and top_p 0.99 on the BF16 model. Its DeepSWE result (43.3) trails its SWE-bench Verified result (81.6), so test it on repository-scale tasks before relying on the headline number.

MiMo-V2.6, Hy4-preview and Ling-3.1-flash

Xiaomi’s MiMo-V2.6-Pro-RL (1.02T / 42B active, MIT, text, image, video and audio input) reports agent scores close to DeepSeek-V4.1-Flash’s, but its card names no harness. Its Flash-RL sibling is 309B / 15B active, and a 9B distill is built on Qwen3.5-9B. Tencent’s Hy4-preview is 770B / 49B active under Apache 2.0, with a BF16 repository of about 1.56 TB. inclusionAI’s Ling-3.1-flash (560B / 25B active, October 13) is MIT per its card and tag, though the repository has no LICENSE file. All three need datacenter hardware.

Qwen3.8-2.4T-A95B

The downloadable checkpoint is text-only and thinking-only, while the hosted Qwen3.8-Max product has additional features. Its card labels benchmark results with the hosted product’s name, so do not assume feature parity. The custom license has an aggregate-revenue condition for specified AI businesses, separate UI-attribution conditions, and an internal-use exception. It differs from both Apache-licensed Qwen3.8-27B and Flash-Next’s Community License 1.0.

Mistral Medium 3.5

Mistral describes a 128B dense model with 256K context and multimodal input. Its card reports SWE-bench Verified 77.6%; that publisher result needs its harness context. The modified MIT license includes a $20M global consolidated monthly-revenue condition. Mistral has fixed a Transformers config error that degraded long-context performance; GGUF files made from the older config inherit the problem. Check the actual license and serving requirements before treating this as an interchangeable alternative to permissively licensed models.

Qwen3-Coder-Next

Qwen’s dedicated coding MoE has 80B total and 3B active parameters, 262,144 native context, and non-thinking output. The official Q4_K_M shards total 48.41 GB. Sparse activation reduces computation per token, but the other experts still occupy storage and generally memory. Add cache and runtime headroom or plan explicit offload; the file size alone does not establish that a 48 GB GPU fits.

Qwen3.6-35B-A3B

This Apache-licensed 35B / 3B-active model remains useful for comparing a small MoE against a dense local model. Its native context is 262,144 tokens. Measure prefill and decode separately on your machine: active-parameter count alone does not establish a speed multiplier.

Inkling

Thinking Machines publishes Inkling with Apache 2.0 metadata and text, image, and audio input. Its 975B / 41B-active architecture is a large deployment commitment. Evaluate it for a specific workload rather than inferring reliability from openness or size. The smaller Inkling-Small release is another candidate when the full checkpoint is impractical.

Kimi K3

Moonshot’s card describes a 2.8T / 104B-active model with MXFP4 weights. The custom license’s $20M test uses aggregate revenue of the licensee and affiliates operating a model-as-a-service business, not only revenue earned from that model. Internal-use and official-provider exceptions are explicit, and UI attribution has separate thresholds. Read those definitions before choosing a hosting arrangement.

Llama 4 Scout

Scout is a 109B / 17B-active multimodal model with a published 10M-token context ceiling. Meta’s quantized single-H100 guidance does not mean the maximum context fits in that configuration. The community license and separate acceptable-use policy include conditions that differ from Muse Glimmer’s terms.

Gemma 4 family (incl. 12B Unified)

Google’s current card lists E2B, E4B, 12B, 26B A4B, and 31B variants under Apache 2.0. Text and image input span the family; audio input is supported on E2B, E4B, and 12B. Context limits and architecture differ by size. The smaller models use 128K context and the medium models support 256K. Pick a checkpoint for its modalities and memory needs rather than treating the whole family as one model.

gpt-oss (120b / 20b) and gpt-oss-safeguard

OpenAI’s cards describe text-only models with tool and structured-output support. The published deployment targets are a single 80 GB GPU for 120b and a 16 GB device for 20b. Those are vendor targets, not guarantees for every context or backend. The safeguard variants classify text against a supplied policy. Read the usage policy alongside Apache 2.0 and test the required output format.

OLMo 3.1 and OLMo Hybrid

Ai2 publishes model, training, and data artifacts for OLMo. OLMo 3.1 32B-Think provides an auditable reasoning checkpoint; OLMo Hybrid explores a mix of Gated DeltaNet and attention layers. These releases are useful when inspecting training matters as much as answer quality. Reproducing training still requires the relevant data, software, and compute.

Molmo 2

Ai2’s Molmo 2 family supports image and video understanding, including pointing and tracking. The model card distinguishes its Apache 2.0 artifacts from conditions on some third-party training datasets. Review those terms for the activity you intend, especially training or redistributing data. Grounding quality should be checked on your images and videos.

MiniCPM-o 4.5

OpenBMB combines text, vision, and speech capabilities in a 9B model. Its published memory table lists 19 GB in BF16 and 11 GB in INT4. Those figures are a starting point for evaluating the supported runtime, streaming mode, and latency on your hardware.

Whisper large-v3 + turbo

OpenAI’s repository lists about 10 GB VRAM for large-v3 and 6 GB for turbo. Turbo is an optimized 809M-parameter derivative; it is not trained for translation. For speech translated into English, use a multilingual non-turbo model. Measure transcription accuracy on your language mix, noise conditions, and vocabulary. Real-time chunking is an application concern; the model does not guarantee streaming behavior.

Qwen3-TTS (permissive speech)

Qwen provides Apache-licensed 0.6B and 1.7B variants with multilingual speech and voice-cloning capabilities. Check which checkpoint supports the required voice workflow. Obtain permission for voices used in cloning and evaluate pronunciation and latency on representative text.

Voxtral TTS (high-quality, non-commercial)

The Voxtral-4B-TTS-2603 card identifies CC-BY-NC 4.0 terms. Treat that as a constraint on the intended use rather than assuming every internal business application is non-commercial. A permissive alternative may be simpler when commercial deployment is required.

Wan2.2 (open video)

The Wan2.2 repository documents a 5B text/image-to-video model generating 720p at 24 fps, with a 24 GB RTX 4090 example using offload. That is a specific recipe, not a general memory guarantee. Larger A14B variants require more resources. Check motion consistency, generation time, and the exact model license before incorporating a video workflow into production.

Shieldstral

Mistral's 3B multimodal guard model (August 4, 2026): it scores text, images, and text-plus-image inputs against a moderation policy you write in natural language at inference time, and returns a calibrated probability in one forward pass. Fits a single 16 GB NVIDIA GPU in BF16, so it can sit next to the application rather than in a separate cluster. Mistral's own table has it within a point of Qwen3Guard-8B on text safety and well ahead of the larger OmniGuard-7B on multimodal moderation (VLGuard 97.7 vs 88.5). It is a component that sits in front of a model, not a base model in its own right.

License: Apache 2.0

gpt-oss-safeguard (120b / 20b)

OpenAI's October 2025 fine-tunes of gpt-oss that read a developer-supplied policy at inference and classify text against it with a reasoned explanation, at low, medium or high reasoning effort. Same footprints as the base models: the 120b fits one 80 GB GPU, the 20b runs in 16 GB. Built with ROOST, Discord and SafetyKit; OpenAI labels it a research preview intended for safety use cases. Heavier than Shieldstral, and the better fit when the policy is long and the verdict needs a rationale.

License: Apache 2.0 + OpenAI usage policy

Qwen3Guard-Stream

Qwen’s guard models with a token-level classification head that labels output as Safe, Controversial or Unsafe as it is generated, rather than after the response is complete. The 4B card covers 119 languages and dialects. The card gives no release date; the underlying Qwen3Guard technical report dates from October 2025. Use it when you need to stop a streamed response partway; it is not a replacement for a policy-reasoning classifier.

License: Apache 2.0

ControlNet

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

SAM 3.1

Meta’s promptable segmentation model for masks and visual pipelines. The current SAM repository uses a custom license with use restrictions; do not copy the Apache license assumption from SAM 2.

License: Custom SAM License

LaMa

Inpainting model for object removal and background reconstruction inside larger editing stacks.

License: Apache 2.0

RIFE

Frame interpolation for slow motion and frame-rate conversion. Check the model version and test motion artifacts on the intended footage.

License: MIT (code) — verify weight distribution

ProPainter

Mask-aware video inpainting and object removal. Specialized for video editing; not a general-purpose video tool. S-Lab research terms rule out most commercial production uses. Confirm the license before building on it.

License: S-Lab License (research / non-commercial — verify before production)

FramePack

A video-generation implementation designed to reduce inference memory use. Its code license does not relicense the Hunyuan-derived model weights used by a workflow.

License: Apache 2.0 code; inspect base-model weight terms separately

DeepSeek Harness (dsh)

The agent runtime DeepSeek open-sourced on August 13, 2026, and the software that produced the Terminal-Bench and DeepSWE numbers on its V4 model cards ("minimal mode": one system prompt, a bash tool and a file editor). Model-agnostic, plugin-based, and labelled a developer preview with breaking changes expected. Useful in two ways: as a harness to run agents locally, and as the thing to point at when someone asks how a self-reported agent score was produced.

License: MIT

d3 and pplx-decider (decision models)

Decision models return probabilities for multiple-choice, yes/no and scale questions instead of generating text, which suits routing, triage and scoring. vLLM Semantic Router’s d3 is 26.09B parameters, including a 0.46B vision encoder, fine-tuned from Qwen3.8-27B; it accepts text, JSON, images and video. Its card puts it first of 149 on the Jev Decision Index 0.3.2 text board at 66.3, ahead of Perplexity Decider v1.1 at 62.8. The card says vLLM support is on main but not yet in a release, and the CUDA nightly has not been tested on an NVIDIA GPU. Ollama v0.40.9 (October 22) packages d3 at 2B, 4B, 9B and 27B; the packages are quantized, and Ollama notes that the publisher’s scores describe the original checkpoints. Perplexity’s pplx-decider (27B, also from Qwen3.8-27B) reached Ollama in v0.40.7. Perplexity’s own v1.1 card uses a different Decision Index version (61.56), so do not compare its numbers directly with d3’s board.

License: Apache 2.0 (both, per their Hugging Face cards)

EmbeddingGemma 2

Google's October 6, 2026 embedding model. It maps text, code, images, video and audio into one 768-dimensional space. One ungated checkpoint (1.49 GB in bf16) loads as four models: 270M text-only, 440M with the image encoder, 570M with audio, or the full 740M. Matryoshka truncation to 512, 256 or 128 dimensions is supported, though Google says 128 suits text-only work best. The 8,192-token context is shared across modalities. LiteRT-LM bundles for on-device use are 165, 388 and 485 MB; run it in bf16 or fp32, because fp16 returns NaN or degraded embeddings. Scores are Google's own, and the MTEB submission had not been accepted at launch. Switching embedding models means re-embedding the corpus, and Google does not say the EmbeddingGemma 1 and 2 vector spaces are compatible.

License: Tagged Apache 2.0 in repo metadata and stated on the card (no LICENSE file; card still cites the Gemma Prohibited Use Policy)

The snapshot below identifies the publisher for each result. Even a familiar benchmark name can hide different task revisions, harnesses, timeouts, context lengths, and reasoning budgets. Cross-lab rows are not a controlled leaderboard.

Independent evaluators can make comparisons more consistent, but their index versions and model settings change. This guide links to external evaluations without carrying forward undated rankings or treating a composite score as a deployment recommendation.

This is a source review, not a hands-on benchmark. Hardware figures are published artifact sizes or vendor deployment guidance. We have not measured inference speed, accuracy, or reliability across this shortlist.

Single GPU, 12–16GB VRAM, 32–64GB RAM

What fits: Gemma 4 E2B/E4B and 12B Unified (16 GB laptop per Google), MiniCPM-V 4.6, gpt-oss-20b, Shieldstral (16 GB), EmbeddingGemma 2 (1.49 GB bf16 weights only; runtime memory not tested by S5 Labs), FLUX.2 [klein] 4B, 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, a guard model beside the app

Watch-outs: Named quantizations of Qwen3.8-27B and Muse Glimmer-30B are larger than this VRAM band before runtime allocations. Model downloads and system RAM do not automatically increase dedicated GPU memory.

24–48GB VRAM, 64–128GB RAM, fast NVMe, optional multi-GPU

What fits: Qwen3.8-27B (16.46 GB UD-Q4_K_M language model plus 0.93 GB BF16 vision projector), Muse Glimmer-30B (16.76 GB Q4 language model plus 1.40 GB vision file), Gemma 4, and smaller speech or vision workloads. Nemotron 3.5 Bolt’s 2-bit GGUF build has NVIDIA recipes for 28 GB (65,536 context, q8_0 cache) and 32 GB (262,144 context); NVIDIA pairs them with 32 GB+ and 48 GB+ GPUs. Recursive Flint’s Ollama build is 23 GB; Recursive’s own rec engine requires 48 GB of unified memory. FLUX 3 Image Klein peaks at 25–31 GB on BFL’s B200 test without its 9B prompt upsampler, which adds about 28 GB; those are datacenter peaks, not minimums, and BFL says its INT8 build runs on RTX GPUs with as little as 8 GB. These are candidate configurations, not tested runtime guarantees. Qwen3-Coder-Next’s Q4_K_M download totals 48.41 GB and needs additional memory or offload.

Best for: Serious private assistants, agentic coding, local omni experiments, MoE serving, open-weight video generation prototyping

Watch-outs: Budget caches, runtime buffers, and the optional drafting model. Larger contexts and concurrent requests raise memory use. Keep sufficient free unified memory for the operating system.

Multi-GPU clusters, high-bandwidth networking, optimized serving

What fits: GLM-5.3, GLM-5.3-Flash, Qwen3.8-Flash-Next, DeepSeek V4, Kimi K3, Inkling, Muse Spark Open, Step-5-Preview, MiMo-V2.6, Beam 1.0, Hy4-preview, Ling-3.1-flash, and other large models. Meta asks for at least 700 GB of VRAM for Spark Open’s NVFP4 build and 2 TB for BF16. Flash-Next’s 125B backbone is only part of its roughly 180B-parameter checkpoint. Use the exact release’s serving recipe and account for all files.

Best for: Internal copilots, agent platforms, omnimodal services, governed deployment

Watch-outs: Some models can run with quantization and offload on smaller machines at a substantial speed cost. That is different from keeping the full model resident or serving multiple users.

Founder or operator testing AI internally

Test Qwen3.8-27B or Muse Glimmer-30B on the hardware you have, with a modest context budget and a small set of representative tasks. Use a smaller Gemma 4 or gpt-oss-20b configuration when memory is tight.

Developer building a local coding copilot

Compare Qwen3.8-27B with Qwen3-Coder-Next on your repositories. Coder-Next’s 48.41 GB Q4_K_M artifact needs runtime headroom beyond the download size; use a larger-memory machine or explicit offload.

Creative team evaluating image and video

Use SDXL or an appropriately licensed FLUX checkpoint for image experiments; FLUX 3 Image Klein’s free commercial use of outputs stops at US$5M gross annualized revenue, counted with affiliates. For video, compare the documented Wan2.2 or LTX-2.5 workflow with your footage and hardware. Confirm output quality, processing time, and current weight terms before production use.

Team building voice or omnimodal interfaces

Compare MiniCPM-o 4.5 and Qwen3-Omni-30B-A3B-Instruct for audio-visual interaction, Whisper or Qwen3-ASR for transcription, and Qwen3-TTS or Chatterbox for synthesis. Check whether each variant produces speech or only text; benchmark end-to-end latency and review voice permissions.

Team shipping an agent that reads untrusted input

Restrict tool permissions, sandbox execution, and evaluate prompt injection with representative untrusted inputs. A guard model is one layer. Meta’s AgentDojo result describes one benchmark setup and is not a measured baseline for every agent.

Enterprise team with privacy and governance requirements

Prioritize license clarity (per-checkpoint, not per-family), eval discipline, and serving fit over raw leaderboard rank. vLLM-class serving plus Qwen3.8-27B, Muse Glimmer-30B or DeepSeek-V4-Flash-0731 (MIT) is usually the right first step. If a checkpoint's license carries a revenue trigger (Kimi K3, Muse Spark Open, Step-5-Preview, Qwen3.8-2.4T-A95B, Mistral Medium 3.5, LFM2.5-24B-A2B), have counsel read it before the evaluation starts.