Which models for which purpose?
Open models sorted by the business work they do. Every one can be hosted in your own cloud or on machines you control, and US and EU models are listed first.
Every model, by the work it does
Each card says what the model is good for and what it runs on. Open any card for the details.
Picks by purpose
Drafting, Triage & Review
Answer, plan, and draft across everyday work, from a laptop up to a large server.
See all 18Drafting & Review models-
Meta · 2024-12 US
General-purpose assistant across many languages
-
Google · 2026-03 US
Fast everyday writing and Q&A on a single workstation or laptop
-
OpenAI · 2025-08 US
Strong general reasoning on one workstation-class GPU
-
Qwen (Alibaba) · 2026-08
Long documents and multi-step work on a single large machine
Code, SQL & Internal Tools
Edit real repositories, call tools, and run multi-step agent loops.
See all 8Code & SQL models-
poolside · 2026-07 US
Coding across large codebases on one workstation-class GPU
-
Mistral AI · 2025-12 EU
Fixing code and running small coding tasks on one GPU
-
Qwen (Alibaba) · 2025-07
Fast coding help and tool calls on one GPU
-
Z.ai · 2026-08
Coding and multi-step agent work on a pair of servers
Document & Data Extraction
Read screenshots, scans, PDFs, and web pages into text or structure.
See all 5Documents models-
Ai2 (Allen Institute for AI) · 2025-10 US
Turning scanned PDFs into clean, searchable text
-
Jina AI · 2025-01 EU
Turning web pages into clean text for search and records
Non-commercial -
Qwen (Alibaba) · 2025-10
Reading scans, charts, screenshots, and video
-
OpenBMB · 2026-10
Reading images and documents on smaller machines
Knowledge Search & Q&A
Turn documents into vectors and rank what retrieval returns, for search and RAG.
See all 7Knowledge Search models-
Google · 2026-09 US
Small, fast search over text and images
-
Nomic AI · 2025-02 US
Lightweight multilingual document search
-
Qwen (Alibaba) · 2025-06
High-quality search over documents and code
-
BAAI · 2024-03
Putting the best search results at the top
Calls & Meetings
Transcribe meetings and calls, and speak answers back.
See all 6Calls & Meetings models-
OpenAI · 2024-10 US
Transcribing and translating meetings and calls
-
NVIDIA · 2025-04 US
Fast English transcripts with timestamps
-
Canopy Labs · 2025-03 US
Expressive English voice, streamed as it speaks
-
Qwen (Alibaba) · 2026-01
Transcribing calls in many languages
Marketing Images
Draft product shots, concept art, and marketing visuals on machines you control.
See all 3Images models-
Krea · 2026-06 US
Fast image drafts for marketing and design
-
Black Forest Labs · 2024-08 EU
High-quality concept art and drafts (non-commercial use only)
Non-commercial -
Qwen (Alibaba) · 2025-12
Images that include readable text, like signs and labels
Guardrails, Redaction & Routing
Screen prompts and outputs, catch injections, and route requests before a large model sees them.
See all 3Guardrails models-
Meta · 2025-04 US
Screening what goes into and comes out of a model
-
Meta · 2024-07 US
Catching jailbreak and injection attempts cheaply
-
Protect AI · 2024-04 US
Catching manipulated prompts before they reach a model
All 50 models
= verified in our lab
All 50 models shown
Drafting, Triage & Review
Answer, plan, and draft across everyday work, from a laptop up to a large server.
Beam 1.0 501B-A23B
Reflection AI · 2026-10 US
Very long documents, up to a million tokens, on a server cluster
- Params
- 501.3B
- Active
- 23.4B
- Context
- 1,000,000 tokens
License
Apache-2.0Formats
- BF16
- FP8
- NVFP4
Fits on
- Needs about 320 GB at 4-bit, more than any single box here. Plan for a multi-GPU server or cluster.
LFM2.5-24B-A2B
Liquid AI · 2026-10 US
Fast, private agent steps and tool calls on a laptop
- Params
- 24B
- Active
- 2.45B
- Context
- 131,072 tokens
License
LFM Open License v1.0 · Commercial use limitedFormats
- BF16
- FP8
- GGUF
- MLX
- ONNX
Fits on
- MacBook Pro 24GB
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
- ★
Gemma 4 26B A4B (it)
Google · 2026-03 US
Fast everyday writing and Q&A on a single workstation or laptop
- Params
- 25.81B
- Active
- 3.8B
- Context
- 262,144 tokens
License
Apache-2.0Formats
- BF16
- QAT Q4_0
- FP8
- NVFP4
- GGUF
- MLX
Fits on
- MacBook Pro 24GB
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB(lab run on this box)
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
Gemma 4 31B (it)
Google · 2026-03 US
Everyday writing, analysis, and reading images on one workstation
- Params
- 31.27B
- Context
- 262,144 tokens
License
Apache-2.0Formats
- BF16
- QAT Q4_0
- FP8
- NVFP4
- GGUF
- MLX
Fits on
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
Nemotron 3 Super 120B-A12B
NVIDIA · 2026-03 US
Multi-step workflows and long documents on one large server
- Params
- 123.61B
- Active
- 12B
- Context
- 1,048,576 tokens
License
NVIDIA Nemotron Open Model License · Custom termsFormats
- BF16
- FP8
- NVFP4
- GGUF
- MLX
Fits on
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
Nemotron 3 Nano 30B-A3B
NVIDIA · 2025-12 US
Fast, low-cost answers and tool calls on modest hardware
- Params
- 31.58B
- Active
- 3.5B
- Context
- 262,144 tokens
License
NVIDIA Nemotron Open Model License · Custom termsFormats
- BF16
- FP8
- NVFP4
- GGUF
- MLX
Fits on
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
gpt-oss-120b
OpenAI · 2025-08 US
Strong general reasoning on one workstation-class GPU
- Params
- 116.83B
- Active
- 5.1B
- Context
- 131,072 tokens
License
Apache-2.0Formats
- MXFP4 (native)
- GGUF
- MLX
Fits on
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
gpt-oss-20b
OpenAI · 2025-08 US
Lightweight assistant for laptops and small servers
- Params
- 20.91B
- Active
- 3.6B
- Context
- 131,072 tokens
License
Apache-2.0Formats
- MXFP4 (native)
- GGUF
- MLX
Fits on
- MacBook Pro 24GB
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
- ★
Llama 3.3 70B Instruct
Meta · 2024-12 US
General-purpose assistant across many languages
- Params
- 70.55B
- Context
- 131,072 tokens
License
Llama 3.3 Community License · Custom termsFormats
- BF16
- FP8
- NVFP4
- GGUF
- MLX
- GPTQ (W4A16)
- EXL2
Fits on
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB(lab run on this box)
Saul-7B Instruct v1
Equall · 2024-02 US
Reading and drafting legal text on a laptop or small server
- Params
- 7.24B
- Context
- 32,768 tokens
License
MITFormats
- FP32
- GGUF
Fits on
- MacBook Pro 24GB
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
DeepSeek-V4.1-Flash
DeepSeek · 2026-09
Very long documents and reading images on a large server
- Params
- 552B
- Active
- 16B
- Context
- 1,048,576 tokens
License
MITFormats
- FP8 (native, mixed)
- NVFP4
- MLX
Fits on
- Needs about 350 GB at 4-bit, more than any single box here. Plan for a multi-GPU server or cluster.
- ★
MiMo-V2.6-Flash-RL
Xiaomi · 2026-09
Long, multi-step tasks that call your tools, on a pair of servers
- Params
- 309B
- Active
- 15B
- Context
- 1,048,576 tokens
License
MITFormats
- FP8 (native, mixed)
- GGUF
- MLX
- EXL3
Fits on
- 2x DGX Spark 256GB, assumed
- ★
Qwen3.8-27B
Qwen (Alibaba) · 2026-08
All-round writing, analysis, and tool use on one workstation
- Params
- 27.78B
- Context
- 262,144 tokens
License
Apache-2.0Formats
- BF16
- FP8
- NVFP4
- GGUF
- MLX
- AWQ
Fits on
- MacBook Pro 24GB
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB(lab run on this box)
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB(lab run on this box)
- ★
Qwen3.8-Flash-Next
Qwen (Alibaba) · 2026-08
Long documents and multi-step work on a single large machine
- Params
- 180B
- Active
- 6B
- Context
- 262,144 tokens
License
Qwen Community License 1.0 · Custom termsFormats
- BF16
- FP8
- NVFP4
- GGUF
- MLX
- AWQ
Fits on
- DGX Spark 128GB(lab run on this box)
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
Kimi K3
Moonshot AI · 2026-06
Very large assistant for complex, long-running work on big server clusters
- Params
- 2.78T
- Active
- 104B
- Context
- 1,048,576 tokens
License
Kimi K3 license (custom) · Custom termsFormats
- MXFP4 (native, QAT)
- NVFP4
- FP8
- GGUF
Fits on
- Needs about 1,760 GB at 4-bit, more than any single box here. Plan for a multi-GPU server or cluster.
DeepSeek-V4-Flash
DeepSeek · 2026-04
Long reports and large document sets on a multi-GPU server
- Params
- 284B
- Active
- 13B
- Context
- 1,048,576 tokens
License
MITFormats
- FP4+FP8 (native, mixed)
- NVFP4
- GGUF
- MLX
Fits on
- 2x DGX Spark 256GB, assumed
Qwen3.5-35B-A3B
Qwen (Alibaba) · 2026-02
Quick answers and light coding help on one workstation
- Params
- 35.95B
- Active
- 3B
- Context
- 262,144 tokens
License
Apache-2.0Formats
- BF16
- FP8
- GPTQ (Int4)
- GGUF
- MLX
Fits on
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
DeepSeek-V3.2
DeepSeek · 2025-12
Large general-purpose assistant for teams with serious server capacity
- Params
- 671B
- Active
- 37B
- Context
- 163,840 tokens
License
MITFormats
- FP8 (native)
- NVFP4
- GGUF
Fits on
- Needs about 420 GB at 4-bit, more than any single box here. Plan for a multi-GPU server or cluster.
Code, SQL & Internal Tools
Edit real repositories, call tools, and run multi-step agent loops.
Laguna S 2.1
poolside · 2026-07 US
Coding across large codebases on one workstation-class GPU
- Params
- 117.56B
- Active
- 8B
- Context
- 1,048,576 tokens
License
OpenMDW-1.1Formats
- BF16
- FP8
- NVFP4
- INT4
- GGUF
- MLX
Fits on
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
Devstral Small 2 24B Instruct (2512)
Mistral AI · 2025-12 EU
Fixing code and running small coding tasks on one GPU
- Params
- 24.01B
- Context
- 262,144 tokens
License
Apache-2.0Formats
- FP8 (native)
- GGUF
- MLX
Fits on
- MacBook Pro 24GB
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
MiMo-V2.6-Pro-RL
Xiaomi · 2026-09
The largest coding and agent jobs, on a server cluster
- Params
- 1.02T
- Active
- 42B
- Context
- 1,048,576 tokens
License
MITFormats
- FP8 (native, mixed)
- GGUF
- MLX
Fits on
- Needs about 650 GB at 4-bit, more than any single box here. Plan for a multi-GPU server or cluster.
GLM-5.3
Z.ai · 2026-08
Heavy coding and long agent tasks on a server cluster
- Params
- 753.33B
- Context
- 1,048,576 tokens
License
GLM-5.3 license (custom) · Custom termsFormats
- FP8 (native)
- BF16
- NVFP4
- GGUF
Fits on
- Needs about 480 GB at 4-bit, more than any single box here. Plan for a multi-GPU server or cluster.
GLM-5.3-Flash
Z.ai · 2026-08
Coding and multi-step agent work on a pair of servers
- Params
- 321.32B
- Active
- 18B
- Context
- 1,048,576 tokens
License
MITFormats
- FP8 (native)
- BF16
- NVFP4
- GGUF
Fits on
- 2x DGX Spark 256GB, assumed
North-Mini-Code-1.0
Cohere Labs · 2026-06
Everyday coding help on a single GPU
- Params
- 30.48B
- Active
- 3B
- Context
- 262,144 tokens
License
Apache-2.0Formats
- BF16
- FP8
- INT4 (W4A16)
- GGUF
- MLX
Fits on
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
Ornith-1.0-9B
DeepReinforce (Ornith) · 2026
Coding agents that work in a terminal, on a single GPU
- Params
- 9B
- Active
- 9B
- Context
- 262,144 tokens
License
MITFormats
- BF16
Fits on
- MacBook Pro 24GB
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
Qwen3-Coder-30B-A3B-Instruct
Qwen (Alibaba) · 2025-07
Fast coding help and tool calls on one GPU
- Params
- 30.53B
- Active
- 3.3B
- Context
- 262,144 tokens
License
Apache-2.0Formats
- BF16
- FP8
- GGUF
- MLX
- AWQ
Fits on
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
Document & Data Extraction
Read screenshots, scans, PDFs, and web pages into text or structure.
olmOCR-2-7B-1025
Ai2 (Allen Institute for AI) · 2025-10 US
Turning scanned PDFs into clean, searchable text
- Params
- 8.29B
- Context
- 128,000 tokens
License
Apache-2.0Formats
- BF16
- FP8
- GGUF
- MLX
Fits on
- MacBook Pro 24GB
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
ReaderLM-v2
Jina AI · 2025-01 EU
Turning web pages into clean text for search and records
- Params
- 1.54B
- Context
- 512,000 tokens
License
CC-BY-NC-4.0 · Non-commercialFormats
- BF16
- GGUF
- MLX
Fits on
- MacBook Pro 24GB
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
MiniCPM-V-4.7-35B-A3B
OpenBMB · 2026-10
Reading images and documents on smaller machines
- Params
- 35.21B
- Active
- 3B
- Context
- 262,144 tokens
License
Apache-2.0Formats
- BF16
- FP8
- AWQ
- GPTQ
- INT4
- GGUF
- MLX
Fits on
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
Qwen3.6-35B-A3B
Qwen (Alibaba) · 2026-04
Reading screenshots and documents, then acting on them
- Params
- 35.95B
- Active
- 3B
- Context
- 262,144 tokens
License
Apache-2.0Formats
- BF16
- FP8
- NVFP4
- GGUF
- MLX
- AWQ
Fits on
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
Qwen3-VL-32B-Instruct
Qwen (Alibaba) · 2025-10
Reading scans, charts, screenshots, and video
- Params
- 33.36B
- Context
- 262,144 tokens
License
Apache-2.0Formats
- BF16
- FP8
- NVFP4
- GGUF
- MLX
Fits on
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
Knowledge Search & Q&A
Turn documents into vectors and rank what retrieval returns, for search and RAG.
EmbeddingGemma 2
Google · 2026-09 US
Small, fast search over text and images
- Params
- 744M
- Context
- 8,192 input tokens
License
Apache-2.0Formats
- BF16
- GGUF
- ONNX
- MLX
Fits on
- MacBook Pro 24GB
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
nomic-embed-text-v2-moe
Nomic AI · 2025-02 US
Lightweight multilingual document search
- Params
- 475M
- Active
- 305M
- Context
- 512 input tokens
License
Apache-2.0Formats
- FP32
- GGUF
Fits on
- MacBook Pro 24GB
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
jina-embeddings-v3
Jina AI · 2024-09 EU
Multilingual document search (non-commercial use only)
- Params
- 572M
- Context
- 8,192 input tokens
License
CC-BY-NC-4.0 · Non-commercialFormats
- BF16
- ONNX
- GGUF
Fits on
- MacBook Pro 24GB
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
Qwen3-Embedding-8B
Qwen (Alibaba) · 2025-06
High-quality search over documents and code
- Params
- 7.57B
- Context
- 32,768 input tokens
License
Apache-2.0Formats
- BF16
- GGUF
- MLX
Fits on
- MacBook Pro 24GB
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
Qwen3-Reranker-8B
Qwen (Alibaba) · 2025-06
Putting the best documents and code results first
- Params
- 8.19B
- Context
- 32,768 input tokens
License
Apache-2.0Formats
- BF16
- GGUF
- MLX
Fits on
- MacBook Pro 24GB
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
bge-reranker-v2-m3
BAAI · 2024-03
Putting the best search results at the top
- Params
- 568M
- Context
- 8,192 input tokens
License
Apache-2.0Formats
- FP32
- ONNX
- GGUF
Fits on
- MacBook Pro 24GB
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
BGE-M3
BAAI · 2024-01
Finding the right document across many languages
- Params
- 568M
- Context
- 8,192 input tokens
License
MITFormats
- FP32
- ONNX
- GGUF
Fits on
- MacBook Pro 24GB
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
Calls & Meetings
Transcribe meetings and calls, and speak answers back.
parakeet-tdt-0.6b-v2
NVIDIA · 2025-04 US
Fast English transcripts with timestamps
- Params
- 600M
- Context
- 1,440 s of audio per pass
License
CC-BY-4.0Formats
- NeMo checkpoint
- ONNX
- MLX
Fits on
- MacBook Pro 24GB
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
Orpheus 3B 0.1 (ft)
Canopy Labs · 2025-03 US
Expressive English voice, streamed as it speaks
- Params
- 3.78B
- Context
- Not stated
License
Apache-2.0Formats
- FP32
- GGUF
Fits on
- MacBook Pro 24GB
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
whisper-large-v3-turbo
OpenAI · 2024-10 US
Transcribing and translating meetings and calls
- Params
- 809M
- Context
- 30 s audio window
License
MITFormats
- FP16
- ONNX
- MLX
Fits on
- MacBook Pro 24GB
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
Qwen3-ASR-1.7B
Qwen (Alibaba) · 2026-01
Transcribing calls in many languages
- Params
- 2.35B
- Context
- 65,536 tokens
License
Apache-2.0Formats
- BF16
- GGUF
- MLX
Fits on
- MacBook Pro 24GB
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
Qwen3-TTS-12Hz-1.7B-CustomVoice
Qwen (Alibaba) · 2026-01
Custom brand voices in many languages
- Params
- 1.92B
- Context
- Not stated
License
Apache-2.0Formats
- BF16
- MLX
Fits on
- MacBook Pro 24GB
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
Kokoro-82M
hexgrad · 2024-12
Natural-sounding read-aloud on almost any machine
- Params
- 82M
- Context
- 512 phoneme tokens
License
Apache-2.0Formats
- FP32
- ONNX
- MLX
Fits on
- MacBook Pro 24GB
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
Marketing Images
Draft product shots, concept art, and marketing visuals on machines you control.
Krea 2 Turbo
Krea · 2026-06 US
Fast image drafts for marketing and design
- Params
- 12.82B
- Context
- Not stated
License
Krea 2 Community License · Commercial use limitedFormats
- BF16
- FP8
- GGUF
- MLX
Fits on
- MacBook Pro 24GB
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
FLUX.1 [dev]
Black Forest Labs · 2024-08 EU
High-quality concept art and drafts (non-commercial use only)
- Params
- 11.9B
- Context
- Not stated
License
FLUX.1 [dev] Non-Commercial License · Non-commercialFormats
- BF16
- FP8
- GGUF
- ONNX
- MLX
Fits on
- MacBook Pro 24GB
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
Qwen-Image-2512
Qwen (Alibaba) · 2025-12
Images that include readable text, like signs and labels
- Params
- 20.43B
- Context
- Not stated
License
Apache-2.0Formats
- BF16
- FP8
- GGUF
- MLX
Fits on
- MacBook Pro 24GB
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
Guardrails, Redaction & Routing
Screen prompts and outputs, catch injections, and route requests before a large model sees them.
Llama Guard 4 12B
Meta · 2025-04 US
Screening what goes into and comes out of a model
- Params
- 12B
- Context
- Not stated
License
Llama 4 Community License · Custom termsFormats
- BF16
- FP8
- GPTQ (W4A16)
- GGUF
Fits on
- MacBook Pro 24GB
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
Prompt-Guard-86M
Meta · 2024-07 US
Catching jailbreak and injection attempts cheaply
- Params
- 279M
- Context
- 512 input tokens
License
Llama 3.1 Community License · Custom termsFormats
- FP32
- ONNX
Fits on
- MacBook Pro 24GB
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
deberta-v3-base-prompt-injection-v2
Protect AI · 2024-04 US
Catching manipulated prompts before they reach a model
- Params
- 184M
- Context
- 512 input tokens
License
Apache-2.0Formats
- FP32
- ONNX
Fits on
- MacBook Pro 24GB
- RTX 4090 24GB
- RTX PRO 6000 96GB
- DGX Spark 128GB
- 2x DGX Spark 256GB, assumed
- Tenstorrent QuietBox 2 128GB
No models match.
How this list is built
We check each open-weight entry against its public Hugging Face repository: license, total weight size, and creation month from the repo, context length from the model config, active parameters from the model card.
A star means we served the model on our own machine. The card says which machine; the run details stay private. A model that is staged on a lab machine but has not been run does not get a star.
Footprints are weights only, from total parameters. Fit uses 4-bit weights plus 15% headroom. For a MacBook Pro with 24 GB, we count about 18 GB as available for weights. Release month is the public release where known, otherwise the repository creation month.
Fit formula
Weights only. Footprint = total parameters (billions) x bytes per weight: BF16 2.0, FP8 1.0, q4 0.55 GB per billion. MoE models use total parameters, not active. A precision fits when footprint x 1.15 (15% headroom) is at most the box's memory. KV cache, context length, activations, and runtime overhead are not included, so a long-context workload can fail to load even when this says it fits.
On DGX Spark, a fit above the lab's practical budget of about 110 GB is marked tight. Any fit at 90% or more of memory is also tight. A model that only runs at 2-bit is not counted as a fit, even when we ran it that way.
What the labels mean
- Verified
- We served it on one of our own machines and publish what we found.
- Locally hostable
- The weights are public, so it can run in your own cloud or on machines you control.
Licenses to check
3 entries carry a license that rules out commercial use, and every card flags it: ReaderLM-v2 (CC-BY-NC-4.0), jina-embeddings-v3 (CC-BY-NC-4.0), FLUX.1 [dev] (FLUX.1 [dev] Non-Commercial License). Two more, LFM2.5-24B-A2B and Krea 2 Turbo, limit free commercial use under their licenses. Eight others use the publisher's own license, which allows commercial use under conditions, and their cards say Custom terms. Check the license before you put any model into a product.
Talk to us
Test a model on your own work.
Tell us the job, the models you are considering, and where it has to run. We email you a confirm link first. Once you click it, a person reads your note.
Prefer email? Write to contact@telchines.io.