Models — ArcFlare

Kitten TTS 2: a 1.7B speech model that clones a voice from 5-30 seconds of audio, with 47 built-in voices, emotion tags and a 0.95 GB ternary download that runs on the CPU too.

size: 1.7B
vram: ~4 GB
license: Stellon Labs Community

17% of a 24 GB card

#tts #speech #voice-cloning

voxcpm2

OpenBMB

VoxCPM2: the highest-quality model here, with 48 kHz output in 30 languages. It wants a CUDA GPU.

size: 2B
vram: ~8 GB
license: Apache-2.0

33% of a 24 GB card

#tts #speech #high-quality

qwen3.5

Alibaba Qwen

Qwen3.5: Qwen's 2026 generation, 2B to 35B-A3B, with a 256K native context and strong coding for a general model.

size: 27B
vram: ~18 GB
license: Apache-2.0

75% of a 24 GB card

#chat #code #reasoning

qwen3-coder-next

Alibaba Qwen

Qwen3-Coder-Next: an 80B MoE with only 3B active, a 256K context, and training aimed at agent harnesses like Claude Code and Cline.

size: 80B-A3B
vram: ~52 GB
license: Apache-2.0

2.2× a 24 GB card

#code #agentic #moe

qwen3-tts

Alibaba Qwen

Qwen3-TTS: Qwen's open speech model. Preset voices, 10 languages, tone control from a text instruction, and a variant that clones a voice from 3 seconds of audio.

size: 1.7B
vram: ~6 GB
license: Apache-2.0

25% of a 24 GB card

#tts #speech #multilingual

glm4.7-flash

Z.ai

GLM-4.7-Flash: a 30B-A3B MoE built for agentic coding. Only ~3B parameters active per token, so it runs at small-model speed.

size: 30B-A3B
vram: ~20 GB
license: MIT

83% of a 24 GB card

#code #agentic #moe

devstral2-small

Mistral AI

Devstral Small 2 (24B, 2512): Mistral's second-generation agentic coder, with a 256K context, under Apache-2.0.

size: 24B
vram: ~16 GB
license: Apache-2.0

67% of a 24 GB card

#code #agentic #tools

olmo-3-7b-instruct

Ai2

Open language model from Ai2, pre-trained on Dolma 3 and post-trained on the Dolci datasets; all code and checkpoints are released.

size: 7B
vram: ~6 GB
license: Apache-2.0

25% of a 24 GB card

#chat #instruct

qwen3-vl-8b-instruct

Qwen

Qwen3-VL 8B Instruct vision-language model with visual agent, spatial perception and OCR in 32 languages; native 256K context, expandable to 1M.

size: 8B
vram: ~7 GB
license: Apache-2.0

29% of a 24 GB card

#vision

lfm2-8b-a1b

Liquid AI

Liquid AI on-device hybrid MoE with 8.3B total and 1.5B active parameters and a 32,768-token context.

size: 8B-A1B
vram: ~6 GB
license: LFM Open License v1.0

25% of a 24 GB card

#chat #instruct #moe

granite-4-0-h-tiny

IBM

IBM 7B-parameter (1B active) hybrid long-context instruct model with improved instruction following and tool calling.

size: 7B-A1B
vram: ~6 GB
license: Apache-2.0

25% of a 24 GB card

#chat #instruct #moe

granite-4.0-h-small

IBM

32B-parameter long-context instruct model with improved instruction following and tool calling, finetuned from Granite-4.0-H-Small-Base.

size: 32B
vram: ~21 GB
license: Apache-2.0

88% of a 24 GB card

#chat #instruct #large

glm4.6

Z.ai

GLM-4.6: Z.ai's flagship ~357B MoE for coding and agents, with a 200K context. Server- or big-workstation class.

size: 357B
vram: ~220 GB
license: MIT

9.2× a 24 GB card

#code #agentic #moe

magistral-small-2509

Mistral AI

24B reasoning model built on Mistral Small 3.2 with SFT from Magistral Medium traces and RL; 128K context window.

size: 24B
vram: ~16 GB
license: Apache-2.0

67% of a 24 GB card

#reasoning

embeddinggemma-300m

Google

300M-parameter open embedding model from Google, built from Gemma 3 and trained on 100+ languages; 768-dim output with Matryoshka truncation.

size: 300M
vram: ~2 GB
license: Gemma

8% of a 24 GB card

#embedding #small

minicpm-v-4.5

OpenBMB

8B multimodal model on Qwen3-8B and SigLIP2-400M for single-image, multi-image and high-FPS video understanding, OCR and document parsing.

size: 8B
vram: ~7 GB
license: Apache-2.0

29% of a 24 GB card

#vision

seed-oss-36b-instruct

ByteDance Seed

ByteDance Seed open-source LLM for long-context, reasoning and agent tasks, with flexible thinking-budget control and native 512K context.

size: 36B
vram: ~23 GB
license: Apache-2.0

96% of a 24 GB card

#reasoning #large

nemotron-nano-9b-v2

NVIDIA

9B hybrid Mamba-2/Transformer model for reasoning and non-reasoning tasks; reasoning controllable via system prompt; up to 128K context.

size: 9B
vram: ~8 GB
license: NVIDIA Open Model License

33% of a 24 GB card

#reasoning

gemma3-270m

Google

Gemma 3 270M: a tiny model built to be fine-tuned for one narrow job. Runs on anything, including a CPU.

size: 270M
vram: ~0.5 GB
license: Gemma

2% of a 24 GB card

#tiny #fine-tuning #cpu

gpt-oss

OpenAI

gpt-oss: OpenAI's open-weight reasoning models. 20B runs in ~16 GB, 120B on one 80 GB card. Apache-2.0, adjustable reasoning effort.

size: 20B
vram: ~14 GB
license: Apache-2.0

58% of a 24 GB card

#reasoning #tools #moe

qwen3-4b-instruct-2507

Alibaba Qwen

Updated Qwen3-4B non-thinking model with better instruction following, reasoning and coding; 256K native context, no think blocks.

size: 4B
vram: ~4 GB
license: Apache-2.0

17% of a 24 GB card

#chat #instruct #small

gpt-oss-120b

OpenAI

OpenAI open-weight MoE reasoning model, 117B parameters with 5.1B active, configurable reasoning effort, MXFP4-trained MoE weights.

size: 117B-A5B
vram: ~64 GB
license: Apache-2.0

2.7× a 24 GB card

#reasoning #moe #large

qwen3-coder

Alibaba Qwen

Qwen3-Coder 30B-A3B: an agentic coding MoE with 256K context that runs at small-model speed on one GPU.

size: 30B-A3B
vram: ~19 GB
license: Apache-2.0

79% of a 24 GB card

#code #agentic #moe

qwen3-30b-a3b-thinking-2507

Alibaba Qwen

Qwen3-30B-A3B thinking-only update: 30.5B MoE with 3.3B active, deeper reasoning and 256K native context.

size: 30B-A3B
vram: ~20 GB
license: Apache-2.0

83% of a 24 GB card

#reasoning #moe #large

glm4.5-air

Z.ai

GLM-4.5-Air: a 106B MoE with 12B active, hybrid reasoning and strong agent and coding scores, under MIT. Workstation-class.

size: 106B-A12B
vram: ~64 GB
license: MIT

2.7× a 24 GB card

#moe #agentic #reasoning

qwen3-30b-a3b-instruct-2507

Alibaba Qwen

Qwen3-30B-A3B non-thinking update: 30.5B-parameter MoE with 3.3B active, 128 experts and 256K native context.

size: 30B-A3B
vram: ~20 GB
license: Apache-2.0

83% of a 24 GB card

#chat #instruct #moe

kimi-k2

Moonshot AI

Kimi K2: a 1-trillion-parameter MoE (32B active) tuned for agentic work. The top of the range: big servers, or very aggressive quants.

size: 1T-A32B
vram: ~250 GB
license: Modified MIT

10.4× a 24 GB card

#moe #agentic #code

nextcoder

Microsoft

NextCoder 32B: Microsoft's code-editing model, trained to apply changes to existing code rather than write it from scratch.

size: 32B
vram: ~22 GB
license: MIT

92% of a 24 GB card

#code #editing #refactoring

smollm3

Hugging Face

SmolLM3 3B: a fully open small model with switchable reasoning, multilingual output and long context. Its training recipe is published too.

size: 3B
vram: ~2.5 GB
license: Apache-2.0

10% of a 24 GB card

#small #reasoning #multilingual

ernie-4-5-21b-a3b

Baidu

Text MoE post-trained model with 21B total and 3B activated parameters per token and a 131072-token context length.

size: 21B-A3B
vram: ~15 GB
license: Apache-2.0

63% of a 24 GB card

#chat #instruct #moe

gemma3n-e4b

Google

Google's Gemma 3n E4B instruction-tuned model for low-resource devices, effective 4B parameters, 32K context; this GGUF is text-only.

size: E4B
vram: ~6 GB
license: Gemma

25% of a 24 GB card

#chat #instruct #small

mistral-small-3.2-24b

Mistral AI

Mistral Small 3.1 update with better instruction following, fewer repetitive outputs and more robust function calling; text and image input.

size: 24B
vram: ~17 GB
license: Apache-2.0

71% of a 24 GB card

#chat #instruct #vision

kimi-dev

Moonshot AI

Kimi-Dev 72B: Moonshot's open model for fixing real GitHub issues: it finds the bug, writes the patch, and checks it.

size: 72B
vram: ~42 GB
license: MIT

1.8× a 24 GB card

#code #swe #agentic

hunyuan3d-2.1

Tencent Hunyuan

Hunyuan3D-2.1: the highest-fidelity shapes of the family, from a 3.3B model, with fully open training code.

size: 3.3B
vram: ~10 GB
license: Tencent Hunyuan Community

42% of a 24 GB card

#3d #image-to-3d #high-fidelity

qwen3-embedding

Alibaba Qwen

Qwen3-Embedding: multilingual text and code embeddings at 0.6B, 4B and 8B, near the top of the MTEB leaderboard.

size: 0.6B
vram: ~1 GB
license: Apache-2.0

4% of a 24 GB card

#embedding #retrieval #multilingual

qwen3-embedding-4b

Qwen

Qwen3 text embedding model with 4B parameters, 32K context, 100+ languages and embedding dimensions up to 2560 (user-defined sizes).

size: 4B
vram: ~6 GB
license: Apache-2.0

25% of a 24 GB card

#embedding #small

deepseek-r1-0528

DeepSeek

DeepSeek-R1-0528: the updated 671B reasoning model, plus an 8B distill on Qwen3 that runs on a gaming GPU. MIT.

size: 8B
vram: ~6 GB
license: MIT

25% of a 24 GB card

#reasoning #math #moe

devstral

Mistral AI & All Hands

Devstral Small 24B: built for agentic coding. It navigates a repo, edits files and runs tests, the way the ArcFlare agent and OpenCode work.

size: 24B
vram: ~16 GB
license: Apache-2.0

67% of a 24 GB card

#code #agentic #tools

outetts

OuteAI

OuteTTS 1.0 0.6B: a small TTS model in 14 languages that clones a voice from a clip.

size: 0.6B
vram: ~2 GB
license: Apache-2.0

8% of a 24 GB card

#tts #speech #voice-cloning

phi-4-reasoning-plus

Microsoft

14B dense reasoning model fine-tuned from Phi-4 with chain-of-thought SFT plus reinforcement learning; 32K context.

size: 14B
vram: ~11 GB
license: MIT

46% of a 24 GB card

#reasoning

qwen3

Alibaba Qwen

Qwen3 runs from 0.6B to 235B, dense and MoE, with thinking and non-thinking modes in one model. Apache-2.0 across the board.

size: 8B
vram: ~6 GB
license: Apache-2.0

25% of a 24 GB card

#chat #reasoning #tools

mellum

JetBrains

Mellum 4B: JetBrains' completion model. A base model built for fill-in-the-middle autocomplete, not chat.

size: 4B
vram: ~5 GB
license: Apache-2.0

21% of a 24 GB card

#code #autocomplete #fim

seed-coder

ByteDance Seed

Seed-Coder 8B: a strong small coding model, with a separate reasoning variant. MIT licensed.

size: 8B
vram: ~7 GB
license: MIT

29% of a 24 GB card

#code #small #reasoning

qwen3-0-6b

Alibaba Qwen

Qwen3 0.6B causal language model with switchable thinking and non-thinking modes, 32K context and 100+ languages.

size: 0.6B
vram: ~2 GB
license: Apache-2.0

8% of a 24 GB card

#chat #instruct #small

qwen3-1-7b

Alibaba Qwen

Qwen3 1.7B causal language model with switchable thinking and non-thinking modes, 32K context and 100+ languages.

size: 1.7B
vram: ~3 GB
license: Apache-2.0

13% of a 24 GB card

#chat #instruct #small

qwen3-14b

Alibaba Qwen

Qwen3 14.8B dense model switching between thinking and non-thinking modes; 32K context, 131K with YaRN, 100+ languages.

size: 14B
vram: ~10 GB
license: Apache-2.0

42% of a 24 GB card

#chat #instruct

qwen3-32b

Alibaba Qwen

Qwen3 32.8B dense model switching between thinking and non-thinking modes; 32K context, 131K with YaRN, 100+ languages.

size: 32B
vram: ~21 GB
license: Apache-2.0

88% of a 24 GB card

#chat #instruct #large

mixtral

Mistral AI

Mixtral 8x7B — a sparse mixture-of-experts model offering strong quality at high inference efficiency.

size: 8x7B
vram: ~28 GB
license: Apache-2.0

1.2× a 24 GB card

#moe #reasoning #instruct

llama4-scout

Meta

Llama 4 Scout: a natively multimodal 109B MoE (17B active, 16 experts) with an extremely long context window.

size: 109B-A17B
vram: ~64 GB
license: Llama 4 Community

2.7× a 24 GB card

#vision #moe #long-context

moondream

vikhyat

A tiny, fast vision-language model that runs almost anywhere — captioning and visual Q&A on the edge.

size: 1.8B
vram: ~3 GB
license: Apache-2.0

13% of a 24 GB card

#vision #small #captioning

chatterbox

Resemble AI

Chatterbox: clones a voice from a few seconds of audio, with 23 languages in its multilingual model. MIT.

size: 0.5B
vram: ~4 GB
license: MIT

17% of a 24 GB card

#tts #speech #voice-cloning

qwen2.5-vl-32b-instruct

Qwen

Instruction-tuned 32B Qwen2.5-VL vision-language model for visual understanding, agentic use, long-video comprehension and structured outputs.

size: 32B
vram: ~22 GB
license: Apache-2.0

92% of a 24 GB card

#vision #large

hunyuan3d-2-turbo

Tencent Hunyuan

Hunyuan3D-2 with the turbo shape model: full-size 2.0 quality in a handful of steps, using FlashVDM decoding.

size: 1.1B
vram: ~6 GB
license: Tencent Hunyuan Community

25% of a 24 GB card

#3d #image-to-3d #fast

hunyuan3d-2mini-fast

Tencent Hunyuan

Hunyuan3D-2 mini, guidance-distilled: about half the compute per step with nearly the same shapes. The middle ground between mini and mini-turbo.

size: 0.6B
vram: ~5 GB
license: Tencent Hunyuan Community

21% of a 24 GB card

#3d #image-to-3d #fast

hunyuan3d-2mini

Tencent Hunyuan

Hunyuan3D-2 mini turns a single image into a clean 3D mesh. 0.6B parameters — the smallest Hunyuan shape model, and the one to start with.

size: 0.6B
vram: ~5 GB
license: Tencent Hunyuan Community

21% of a 24 GB card

#3d #image-to-3d #mesh

mistral-small3.1

Mistral AI

Mistral Small 3.1 24B: fast general model with image understanding and 128K context, under Apache-2.0.

size: 24B
vram: ~16 GB
license: Apache-2.0

67% of a 24 GB card

#vision #chat #tools

gemma3

Google

Gemma 3: 1B to 27B, with image understanding from 4B up, 128K context and 140+ languages.

size: 12B
vram: ~9 GB
license: Gemma

38% of a 24 GB card

#vision #chat #multilingual

gemma3-4b

Google

Google's multimodal Gemma 3 4B instruction-tuned model: text and image input, text output, 128K context, 140+ languages.

size: 4B
vram: ~5 GB
license: Gemma

21% of a 24 GB card

#chat #instruct #vision

gemma3-27b

Google

Google's multimodal Gemma 3 27B instruction-tuned model: text and image input, text output, 128K context, 140+ languages.

size: 27B
vram: ~19 GB
license: Gemma

79% of a 24 GB card

#chat #instruct #vision

olympiccoder

Hugging Face (open-r1)

OlympicCoder 32B: a reasoning model trained for competitive programming, from Hugging Face's open-r1 project.

size: 32B
vram: ~22 GB
license: Apache-2.0

92% of a 24 GB card

#code #reasoning #competitive

mxbai-embed-large

Mixedbread AI

A state-of-the-art large embedding model for English retrieval and semantic search.

size: 335M
vram: ~2 GB
license: Apache-2.0

8% of a 24 GB card

#embedding #rag #search

phi4-mini

Microsoft

Phi-4-mini 3.8B: strong maths and reasoning for its size, 128K context and function calling, under MIT.

size: 3.8B
vram: ~3 GB
license: MIT

13% of a 24 GB card

#small #math #tools

smolvlm2-2.2b-instruct

Hugging Face

Lightweight multimodal model that processes video, image and text inputs to generate text; about 5.2GB of GPU RAM for video inference.

size: 2.2B
vram: ~4 GB
license: Apache-2.0

17% of a 24 GB card

#vision #small

nomic-embed-text-v2-moe

Nomic AI

Multilingual mixture-of-experts text embedding model (475M total, 305M active) supporting about 100 languages with Matryoshka embeddings.

size: 475M
vram: ~2 GB
license: Apache-2.0

8% of a 24 GB card

#embedding #small

qwen2.5-vl

Alibaba Qwen

Qwen2.5-VL: reads documents, charts, screenshots and photos, and can point to things in an image.

size: 7B
vram: ~6 GB
license: Apache-2.0

25% of a 24 GB card

#vision #ocr #documents

hunyuan3d-2

Tencent Hunyuan

The full-size Hunyuan3D-2.0: a 1.1B shape model plus a paint model that textures the mesh from the same image.

size: 1.1B
vram: ~6 GB
license: Tencent Hunyuan Community

25% of a 24 GB card

#3d #image-to-3d #texture

deepseek-r1

DeepSeek

Qwen2.5-Math-7B fine-tuned on reasoning data generated by DeepSeek-R1.

size: 7B
vram: ~6 GB
license: MIT

25% of a 24 GB card

#reasoning

deepseek-r1-distill-qwen-14b

DeepSeek

Qwen2.5-14B fine-tuned on reasoning data generated by DeepSeek-R1.

size: 14B
vram: ~10 GB
license: MIT

42% of a 24 GB card

#reasoning

deepseek-r1-distill-qwen-32b

DeepSeek

Qwen2.5-32B fine-tuned on DeepSeek-R1 reasoning data; card reports it outperforms OpenAI o1-mini across various benchmarks.

size: 32B
vram: ~21 GB
license: MIT

88% of a 24 GB card

#reasoning #large

deepseek-r1-distill-llama-8b

DeepSeek

Llama-3.1-8B fine-tuned on reasoning data generated by DeepSeek-R1.

size: 8B
vram: ~6 GB
license: MIT

25% of a 24 GB card

#reasoning

phi4

Microsoft

14B dense decoder-only Transformer trained on synthetic and filtered public data, built for high-quality reasoning in constrained environments.

size: 14B
vram: ~10 GB
license: MIT

42% of a 24 GB card

#chat #instruct

llama3.3-70b

Meta

Meta's 70B multilingual instruction-tuned text model with 128k context, optimized for multilingual dialogue use cases.

size: 70B
vram: ~44 GB
license: Llama 3.3 Community

1.8× a 24 GB card

#chat #instruct #large

whisper

OpenAI

Whisper — a robust multilingual speech-to-text model for transcription and translation.

size: 1.5B
vram: ~6 GB
license: MIT

25% of a 24 GB card

#audio #speech #transcription

kokoro

hexgrad

Kokoro 82M: a tiny, fast TTS model with good preset voices. It runs on any GPU, or on the CPU.

size: 82M
vram: ~1 GB
license: Apache-2.0

4% of a 24 GB card

#tts #speech #small

qwen2.5-coder-32b

Alibaba Qwen

Code-specific instruction-tuned 32B Qwen2.5-Coder model, improved at code generation, code reasoning and code fixing.

size: 32B
vram: ~21 GB
license: Apache-2.0

88% of a 24 GB card

#code #instruct #large

qwen2.5-coder-14b-instruct

Qwen

Instruction-tuned 14B code model with improved code generation, reasoning and fixing, and long-context support up to 128K tokens.

size: 14B
vram: ~10 GB
license: Apache-2.0

42% of a 24 GB card

#code #instruct

smollm2-1-7b-instruct

Hugging Face

1.7B compact instruct model trained on 11T tokens with SFT and DPO; supports rewriting, summarization and function calling.

size: 1.7B
vram: ~3 GB
license: Apache-2.0

13% of a 24 GB card

#chat #instruct #small

stable-diffusion-3.5

Stability AI

Stable Diffusion 3.5 Large — a powerful 8B text-to-image model with improved typography and photorealism.

size: 8B
vram: ~16 GB
license: Stability AI Community

67% of a 24 GB card

#image #text-to-image #photorealism

llama3.2

Meta

Meta's 3B multilingual instruction-tuned text model, optimized for dialogue, agentic retrieval and summarization tasks.

size: 3B
vram: ~3 GB
license: Llama 3.2 Community

13% of a 24 GB card

#chat #instruct #small

llama3.2-1b

Meta

Meta's 1B multilingual instruction-tuned text model, optimized for dialogue, agentic retrieval and summarization tasks.

size: 1B
vram: ~2 GB
license: Llama 3.2 Community

8% of a 24 GB card

#chat #instruct #small

qwen2.5-coder

Alibaba Qwen

Code-specific instruction-tuned 7B Qwen2.5-Coder model, improved at code generation, code reasoning and code fixing.

size: 7B
vram: ~6 GB
license: Apache-2.0

25% of a 24 GB card

#code #instruct

qwen2.5

Alibaba Qwen

Instruction-tuned 7B Qwen2.5 model with improved coding, maths, instruction following and JSON output; 29+ languages.

size: 7B
vram: ~6 GB
license: Apache-2.0

25% of a 24 GB card

#chat #instruct

yi-coder-9b-chat

01.AI

Open-source 9B code chat model with a 128K maximum context length, supporting 52 major programming languages.

size: 9B
vram: ~7 GB
license: Apache-2.0

29% of a 24 GB card

#code #instruct

flux.1-schnell

Black Forest Labs

FLUX.1 [schnell] — a fast, fully open (Apache-2.0) text-to-image model that generates high-quality images in just 1–4 steps.

size: 12B
vram: ~24 GB
license: Apache-2.0

100% of a 24 GB card

#image #text-to-image #fast

flux.1-dev

Black Forest Labs

FLUX.1 [dev] — a 12B text-to-image model with exceptional quality and prompt adherence (non-commercial license).

size: 12B
vram: ~24 GB
license: FLUX.1 [dev] Non-Commercial

100% of a 24 GB card

#image #text-to-image #high-quality

hermes-3-llama-3-1-8b

Nous Research

Generalist Llama-3.1 8B fine-tune with ChatML format, function calling, structured output, roleplay and reasoning improvements.

size: 8B
vram: ~6 GB
license: Llama 3

25% of a 24 GB card

#chat #instruct

sdxl

Stability AI

Stable Diffusion XL — the widely-used open text-to-image model with a huge ecosystem of fine-tunes and LoRAs.

size: 3.5B
vram: ~8 GB
license: OpenRAIL++-M

33% of a 24 GB card

#image #text-to-image #loras

llama3.1

Meta

Meta's 8B multilingual instruction-tuned text model with 128k context, optimized for assistant-like chat.

size: 8B
vram: ~6 GB
license: Llama 3.1 Community

25% of a 24 GB card

#chat #instruct

mistral-nemo-12b

Mistral AI / NVIDIA

12B instruct model trained jointly by Mistral AI and NVIDIA with a 128k context window on multilingual and code data.

size: 12B
vram: ~9 GB
license: Apache-2.0

38% of a 24 GB card

#chat #instruct

gemma2

Google

Google's instruction-tuned Gemma 2 9B, trained on 8T tokens of web documents, code and mathematics; no system prompt support.

size: 9B
vram: ~7 GB
license: Gemma

29% of a 24 GB card

#chat #instruct

deepseek-coder-v2-lite

DeepSeek

DeepSeek-Coder-V2-Lite: a 16B mixture-of-experts coder with a long context. Cheap to run for its size.

size: 16B
vram: ~12 GB
license: DeepSeek License

50% of a 24 GB card

#code #moe #long-context

codestral

Mistral AI

Codestral 22B: Mistral's dedicated code model, trained on 80+ languages. Non-production licence: research and personal use only.

size: 22B
vram: ~15 GB
license: Mistral Non-Production

63% of a 24 GB card

#code #fim #non-commercial

mistral

Mistral AI

Instruct fine-tuned Mistral-7B-v0.3 with extended 32768-token vocabulary, v3 tokenizer and function calling support.

size: 7B
vram: ~6 GB
license: Apache-2.0

25% of a 24 GB card

#chat #instruct

starcoder2

BigCode

Code LLM self-aligned from StarCoder2-15B with a fully permissive, transparent pipeline and no human annotations or distilled data.

size: 15B
vram: ~11 GB
license: BigCode OpenRAIL-M

46% of a 24 GB card

#code #instruct

triposr

Stability AI & Tripo

TripoSR reconstructs a 3D mesh from one image in a single feed-forward pass — under a second on a GPU. Rougher than Hunyuan, but instant.

size: 0.4B
vram: ~6 GB
license: MIT

25% of a 24 GB card

#3d #image-to-3d #fast

nomic-embed-text

Nomic AI

English text embedding model with Matryoshka representation learning for resizable embeddings; requires task instruction prefixes.

size: 137M
vram: ~2 GB
license: Apache-2.0

8% of a 24 GB card

#embedding #small

bge-m3

BAAI