39.8B total / 2.5B active · local coding · checked 25 October 2026Recursive Flint
An Apache 2.0 coding-agent MoE with a 262,144-token shared prompt/output limit. The official MLX INT4 package is about 22.6 GB. The official Q4_1 GGUF is 26 GB; its Metal recipe requires an Apple Silicon Mac with at least 48 GB unified memory. All experts stay stored, even though only about 2.5B parameters activate per token.
Use a runtime release with Flint architecture and tool-parser support. Recursive's coding and speed comparisons are vendor evaluations; measure accepted edits, recovery and context use in your own harness before changing your default model.
122B class · quantized workstation candidate · checked 25 October 2026Ornith-1.5-122B
A coding-agent candidate with a 262,144-token model window. Its roughly 250 GB BF16 repository exceeds a 128 GB personal machine. Ollama's specific 122B package offers a 73 GB NVFP4 route and a 77 GB Q4_K_M route; reserve additional memory for the operating system, runtime, context and concurrency. Download size does not establish full-context fit.
The Ollama package includes an MIT licence naming deepreinforce.ai; the upstream HF repository declares MIT but its standalone LICENSE path was missing at this check. Keep the licence with the exact redistributed package. Benchmark comparisons are vendor-reported, and the Qwen-compatible architecture alone does not establish the exact base checkpoint or training provenance.
24B total / 2B active · on-device agentsLiquid AI LFM2.5-24B-A2B
Released 13 October, this hybrid MoE targets private tool-using agents with a 131,072-token context. It always produces reasoning before its answer, so budget those tokens and time. Official GGUF weights range from 13.6 GB for QAD Q4_0 to 25.5 GB for Q8_0; reserve additional memory for context and the runtime. Unlike d1's fixed decisions, this is a generative agent model.
Vendor evidence: Liquid reports 48.09 on LiveClaw and 51.0 on SWE-bench Verified in OpenCode. These do not establish your harness's reliability. LFM Open License v1.0 permits free for-profit commercial use below $10 million annual revenue, counting the licence's controlled entities; obtain separate permission at or above that threshold.
Six compact models · released 27 October 2026Liquid AI LFM3
Choose the model for the input: LFM3-1.7B for text, LFM3-VL-2B for vision, LFM3-Audio-2B for speech, LFM3-Omni-2B for combined modalities, and d1-2B or d1-omni-2B for fixed decisions. These six 2B-class models share a backbone; the d1 variants serve a different role from a generative coding assistant.
Liquid specifies a 256K input/output window for the text model and about 1 GB of weights at 4 bits. That is not total working memory or a full-context fit for every modality. The text model card lists SGLang BF16, FP8 and NVFP4 routes and GGUF for llama.cpp clients. Pin an engine with explicit LFM3 support; a package or generic launcher example alone does not verify your installed build. Published quality and speed comparisons are vendor evidence.
Conditional commercial use: LFM Open License v1.0 permits free for-profit use below $10 million annual revenue, subject to its terms and Legal Entity definition, including controlled affiliates. Arrange separate permission at or above that threshold.
12B total / 2.5B active · Apache 2.0JetBrains Mellum2.1 Thinking
A self-hosted coding-agent MoE with 131,072-token context and explicit reasoning. The official GGUF repository is available: Q4_K_M is 8.1 GB, Q6_K 10.9 GB and Q8_0 12.9 GB, before context and runtime memory. Use a compatible llama.cpp, Ollama or LM Studio build and test the tool-call parser.
JetBrains-reported evidence: SWE-bench Verified is 47.0%, compared with 50.0% for Qwen3.5-9B in its shared Pi harness. Smaller active parameter count does not establish better completed-task quality. The weights and official GGUF builds use Apache 2.0.
600B total / 27B active · server scaleStepFun Step 5 Preview
Downloadable multimodal agent weights with up to 1M context. The published FP8 route starts at four B300 or eight H200 GPUs. This is a server deployment, with vendor-reported performance and custom commercial terms, rather than a laptop model.
816B total / 42B active · server scaleMeta Muse Spark Open
Released 15 October with multimodal input and 1M context. The NVFP4 checkpoint is about 537 GB of weights; Meta's model card budgets at least 700 GB VRAM. Its separate Muse Model License requires attribution and has a large-MaaS revenue gate. Meta's capability claims are vendor evidence.
0.5B backbone · MIT · checked 25 October 2026Ming-3.0-tts-0.5B
A smaller speech-synthesis alternative with voice design, expressive control and voice cloning. The 0.5B name describes its Qwen2.5 backbone; the checkpoint also includes a DiT head, audio tokenizer and speaker encoder. Budget the complete pipeline. Inference requires the separate official Ming-3.0-tts code repository and its dependencies.
The card lists Chinese, English, German, Spanish, French, Korean and Japanese, without Estonian. Its reference tests use data-centre NVIDIA GPUs; they do not establish a consumer-device RAM minimum. Evaluate speech quality and latency locally, and clone real voices only with consent. The project uses MIT; published quality comparisons remain vendor evidence.
35B total / 3B active · workstation TTSFish Audio Drama 3
A released speech model for multi-character scenes, long-form narration, voice cloning and spoken direction. Fish's reference FP8 path requires at least 48 GB NVIDIA GPU memory on Linux or WSL. The SGLang-Omni cookbook also documents an int4 checkpoint with 20.8 GB of engine weights, before preprocessing, vocoder and cache allocations. An 8 GB community claim is not the documented FP8 requirement.
Fish Audio Community License v1.0 permits limited free commercial use below $500,000 aggregate annual revenue, including affiliates. Hosted TTS or voice-cloning services need separate permission, subject to the licence's research-demo exceptions; client deployments also have revenue conditions. Clone real voices only with consent. Published quality and serving results remain creator measurements.
27B dense / 35B-A3B · computer useHolo4
Evaluate H company’s released weights when a workflow needs screenshots, GUI actions, code and MCP or API tools. The 27B and 35B-A3B checkpoints have different author-reported OSWorld 2.0 results; reproduce the exact harness and task version before choosing. Holo4-27B uses CC BY-NC 4.0 and requires separate permission for commercial use. Holo4-35B-A3B uses Apache 2.0, which permits commercial use subject to its terms. Quantized weights still need runtime, context and action-harness headroom.
9B dense / 35B-A3B · Apache 2.0NorthCUA Nano / Mini
Computer-use candidates built on Qwen3.5 for screenshots, desktop applications and browser tasks. Nano has 9B total parameters and about 18.84 GB of predominantly BF16 files; Mini has 35B total / 3B active and about 70.23 GB. Reserve additional memory for image processing, activations, context and the execution harness. Mini is a 128 GB-class evaluation candidate; 3B active does not mean 3B of resident weights.
Use the official NorthCUA Harness and its normalized 0-1000 GUI coordinates. Author-reported benchmark results describe research checkpoints and protocols, without a verified mapping to the exact released export. Reproduce task outcomes in an isolated environment before comparison with Holo4.
Separate server tier: Pro is 397B total / 17B active, about 793.63 GB in BF16 files. Hugging Face listed no Inference Provider on 6 November 2026; a managed API route remains unverified.
About 4.7B total · Apache 2.0 · research candidateMicrosoft Mage-VL-Cowork
A compact multimodal model combining GUI clicks with terminal and code actions. The 4B language backbone sits in a roughly 4.7B-parameter checkpoint with 9.48 GB of BF16 weights, before runtime, screenshots and cache memory. Microsoft reports 66.0% on OSWorld-Verified; this is a vendor result and is not directly comparable with Holo4's OSWorld 2.0 scores.
Research scope: The weights carry Apache 2.0, while Microsoft's project README says the models are for research and are not intended for product or service deployment. Keep the licence grant and intended-use guidance distinct. Evaluate only in an isolated, disposable container or VM, without personal credentials or host access, and require approval for consequential actions.
35B total / 3B active · Apache 2.0Occamy-1.0
A candidate for bounded co-work through tools, code and files, with released weights and the Dressage RL stack. The authors’ cost/performance result uses four benchmarks, normalized scores and a fixed pricing protocol. It does not establish your local throughput or completed-task cost. Native visual browser and desktop control are outside its current co-work training interface. Store the full 35B checkpoint, reserve context memory and test tool calls, recovery, retries and final artifacts in your own harness.
16 GB · 14B Q4 9.1 GBMicrosoft Phi-4
A compact MIT-licensed reasoning model. Microsoft reports MMLU 84.8, GPQA 56.1 and HumanEval 82.6 for the 14B base instruction model; reasoning, mini and multimodal Phi-4 variants also ship.
Sizes14Bminimultimodalreasoning
8-32 GB · Apache 2.0IBM Granite 4
Enterprise-oriented models with governance disclosures, hybrid Mamba/Transformer variants and compact footprints. Shipped language sizes include Micro, H-Micro, H-Tiny, H-Small and dense 8B; evaluate the exact card because scores vary by variant.
SizesMicroH-MicroH-TinyH-Small8B
16 / 32 GB · fully open researchAI2 OLMo 3
A rare option with training code, data, checkpoints and detailed recipes. The family ships 7B and 32B Base, Instruct and Think variants with 65,536-token context; use 7B Q4/Q8 at 16 GB and 32B Q4 at 24-32 GB.
Sizes7B32BBaseInstructThink
Estonian specialist · prototypeTartuNLP EstLLM 8B
A locally testable Estonian-focused research checkpoint and evaluation suite. The authors clearly label it an early prototype with 4K context and no multi-turn chat; use it as a benchmark and fine-tuning reference, not an automatic production default.
Size8BBF16 checkpointcommunity quants
13 / 29 GB VRAM · local image specialistFLUX.2 Klein 4B + 9B
The 4B and 9B checkpoints combine text-to-image generation and multi-reference editing in one four-step model. Black Forest Labs says the Apache-2.0 4B model fits about 13 GB VRAM on an RTX 3090 or 4070-class GPU; the stronger 9B model fits about 29 GB and uses the FLUX Non-Commercial License. Both are actual WaveSpeed endpoints, but their released weights make local use credible without the API.
Creator-published fit4B · about 13 GB VRAM9B · about 29 GB VRAM4 inference stepsText + multi-reference editing
Under 16 GB VRAM · Apache 2.0Z-Image Turbo
Tongyi-MAI's 6B image model targets photorealism and bilingual text rendering in a much smaller package than the server leaders. The creator says it runs on consumer GPUs with under 16 GB VRAM and publishes the code and weights. That makes the Turbo checkpoint a genuine local alternative, not merely a WaveSpeed product label.
Creator-published fit6B parametersUnder 16 GB VRAMText-to-imageEnglish + Chinese text
14 GB VRAM with offload · videoHunyuanVideo 1.5
Tencent's 8.3B text-to-video and image-to-video release is the strongest current lightweight video candidate in this catalogue. Tencent reports state-of-the-art open-model quality and motion coherence, publishes training and inference code, and documents a 14 GB minimum with model offloading. The step-distilled 480p image-to-video checkpoint runs in 8 or 12 steps and was measured by Tencent at under 75 seconds on an RTX 4090.
Creator-published fit8.3B parameters14 GB VRAM minimumText + image to video480p distilled + 720p
24 GB VRAM · 720p videoWan 2.2 TI2V 5B
The 5B Wan checkpoint handles both text-to-video and image-to-video at 720p and 24 fps on one 24 GB RTX 4090 with offloading. Alibaba describes it as one of the fastest open 720p models and publishes the weights and inference code. The larger A14B route belongs on the server page; this 5B checkpoint is the consumer-workstation cut.
Creator-published fit5B parameters24 GB VRAM720p24 fps
8-24 GB VRAM with offload · music generationMiniMax Music 3
MiniMax Music 3 generates complete songs of up to five minutes from lyrics and a detailed music description, including vocals and 32 kHz stereo audio. The official Diffusers path fits the full-precision pipeline in under 24 GB VRAM; automatic CPU offload uses about 22 GB, and layer-by-layer language-model streaming can reduce the GPU requirement to 8 GB at the cost of speed. Inference currently requires CUDA and does not stream the generated audio.
Creator-published fitUnder 24 GB · full precisionAbout 22 GB · CPU offload8 GB · layer streamingUp to 5-minute songs
0.6B / 1.7B · speech specialistQwen3-TTS
Qwen releases 0.6B and 1.7B speech models for custom voices, voice design and cloning. On Qwen's Seed-TTS evaluation, the 12 Hz 1.7B Base checkpoint reports 0.77 Chinese and 1.24 English word error rates, ahead of the commercial and open baselines listed in that table. The official project does not state a minimum VRAM, so treat the small parameter count as a starting signal and measure the complete audio stack on the target GPU.
Released sizes0.6B1.7BCustom voiceVoice design + cloning
4-24 GB VRAM · music generationACE-Step 1.5
ACE-Step is the clearest local music-generation addition in the hosted catalogue. The standard model can run with under 4 GB VRAM, while its 4B XL decoder needs at least 12 GB with offload and recommends 20 GB. The creators report under ten seconds per full song on an RTX 3090 and publish benchmark, profiling and GPU-tier tools, including Mac, AMD, Intel and CUDA paths.
Creator-published fitStandard: under 4 GB VRAMXL: 12 GB with offloadXL: 20 GB recommendedUp to 10-minute songs
10-29 GB VRAM · 3D assetsHunyuan3D 2.1
Tencent releases the full 3.3B shape and 2B PBR texture pipeline, including training code. Its creator evaluation leads the listed open and closed baselines on shape following and textured-asset metrics. Hardware scales by task: 10 GB for shape, 21 GB for texture, or 29 GB for the complete pipeline. Note that its community licence excludes the EU, UK and South Korea.
Creator-published fitShape: 3.3B + 10 GB VRAMTexture: 2B + 21 GB VRAMComplete: 29 GB VRAMPBR materials
3.44 GB checkpoint · segmentationSAM 3.1
Meta's current Segment Anything release handles text-prompted image segmentation, video segmentation and multi-object tracking. The official checkpoint is about 3.44 GB. On Meta's public tables, SAM 3.1 improves six of seven video-object-segmentation benchmarks over SAM 3 and adds an approximately sevenfold 128-object speedup on one H100 test. The official project does not publish a consumer-GPU minimum, so measure your prompt count, image size and video workload before promising a device tier.
Released route3.44 GB checkpointImage masksVideo trackingText + visual prompts