Open-weight models

Qwen 3 flagship MoE. 235B total / 22B active per token, with built-in 'thinking' and 'non-thinking' modes that trade speed for reasoning depth at inference…

Qwen3-0.6B is the smallest dense model in Alibaba's Qwen3 generation, supporting a 32K-token context and dual-mode operation that toggles between explicit…

Mid-tier Qwen 3 MoE. 30B total / 3B active. The Q4_K_M weights (about 18.6 GB) fit on a single 24GB card. The sweet spot of the Qwen 3 lineup for prosumer…

Coding-specialist Qwen 2.5. The [technical report](https://arxiv.org/abs/2409.12186) lists HumanEval 92.7, against 92.1 for GPT-4o-2024-08-06. The default…

Dense Qwen 3 32B. Best dense open-weight model in its size class at release; pairs nicely with a single RTX 5090 or 4090.

Qwen 3 at the 8B scale. Direct head-to-head against Llama 3.1 8B on most benchmarks; usually wins on coding and structured output.

Qwen3-1.7B is the mid-tier dense model in Qwen3, sharing the same hybrid thinking architecture and 32K context as the 0.6B but with ~3x the parameters for…

14B Qwen 3. Fits on 12GB cards at Q4. Strong default for users with a single mid-range GPU.

Qwen2-VL 2B Instruct is Alibaba's compact vision-language model with native dynamic-resolution image handling and multimodal RoPE (M-RoPE) for video and…

The community-default small Qwen prior to Qwen 3. Still widely used because of mature ecosystem support.

14B Qwen 2.5. Sweet spot for 16GB VRAM. Many production deployments still on this version.

Dense 32B Qwen 2.5. Strong daily-driver on 24GB cards prior to Qwen 3 32B.

Qwen team's reasoning-focused experimental release. Visible chain-of-thought in its output. Precursor to Qwen 3's thinking mode.

Dense 27B Qwen model with a vision encoder, published on Hugging Face in August 2026 under Apache 2.0. The model card lists 64 layers and a native context of…

Compact Qwen 3 for edge and laptop deployment. Outperforms many 7B models from prior generations.

The flagship of Qwen 2.5. Workstation-tier; needs 48GB+ VRAM for usable inference.

Qwen 3.6 27B dense (not MoE) with Multi-Token Prediction. Sits between the 14B and 35B-A3B as a "single dense model with MTP throughput acceleration." Targets…

Qwen 3.5 2B base with supervised fine-tuning on Turkish instruction-following data. Recent community fine-tune (early 2026) that bridges Qwen 3.5's strong…

Consumer-tier Qwen 2.5 VL. 7B + vision. Fits 8GB cards; the smallest practical multimodal Qwen.

Qwen 2.5 vision-language flagship at 72B. Strong on document understanding + multi-image queries. Qwen License.

Largest Qwen 2.5 Math. Datacenter-tier math specialist; eclipsed by R1 distills for general reasoning.

Compact Qwen 2.5 Coder. Sweet spot for laptop autocomplete and small refactor agents.

Smallest Qwen 2.5 Coder. Targets edge / autocomplete on integrated GPUs and Apple Silicon laptops.

Compact Qwen 2.5. The 1.5B Apache-2.0 baseline.

Smallest Qwen 2.5. Apache 2.0; phone / Pi-class deployment target.

Smallest Qwen 2.5-VL. Edge-deployable VLM with strong document Q&A.

Mid-edge Qwen 2.5. Note: 3B variant uses the Qwen Research License (not Apache 2.0), which is non-commercial.

Qwen 3 family embedding model. Apache 2.0 with strong multilingual coverage.

CodeQwen 1.5 — Qwen Coder predecessor. Superseded by Qwen 2.5 Coder for new deployments.

Qwen 2 vision-language predecessor to Qwen 2.5-VL. Apache 2.0 with strong document Q&A.

Qwen 2.5 fine-tuned for math problem-solving with chain-of-thought and tool-integrated reasoning.

Coding-specialized Qwen 2.5 at 14B. The 16GB-VRAM tier coding model — fits comfortably with 8K context.

Meta's 2025 flagship MoE model. 109B total parameters with only 17B active per forward pass and a 10-million-token maximum context that Meta says was evaluated…

Meta's small flagship. Strong general reasoning, 128K context, broad multilingual. The default first try for most local-AI use cases on consumer hardware.

Late-2024 refresh of the 70B Llama line. Meta says that for some applications it approaches the performance of Llama 3.1 405B, at about one-sixth the parameter…

Lightweight 3B for edge and laptop deployment. Runs comfortably on 8GB VRAM.

The 70B sibling of Llama 3.1 8B. Strong generalist reasoning with 128K context, popular base for agentic fine-tunes (Hermes 3, Nemotron). Mostly superseded by…

NVIDIA's HelpSteer2-tuned Llama 3.1 70B. Topped Arena Hard at release. The pre-Nemotron-3 NVIDIA reference open weights.

First-party multimodal Llama. Accepts images alongside text for VQA, document understanding, and chart reading. Runs on 12GB+ VRAM.

TinyLlama-1.1B-Chat-v1.0 is a 1.1B Llama-2-architecture model pretrained on 3 trillion tokens and chat-tuned on UltraChat and UltraFeedback. It was one of the…

NVIDIA's top open reasoning model in the Llama 3.1 lineage. Server-tier; trained for groundbreaking reasoning accuracy on agentic workloads.

Meta's high-end Llama 4 sibling — 128 experts MoE built for performance over efficiency. Multilingual strength is its standout. Effectively a server-tier…

Smallest of the Nemotron reasoning trio. Derived from Llama 3.1 8B; NVIDIA says it fits on a single RTX GPU.

True edge-tier Llama. Runs on a phone or Raspberry Pi. Useful for classification, simple summarization, and on-device agents.

Llama 3 8B Instruct fully fine-tuned on a 30GB Turkish dataset. YTU CE COSMOS group's most-downloaded Llama variant. GGUF builds available — drops into Ollama…

Turkish-tuned chat model released by Trendyol, Turkey's largest e-commerce platform. Built on Llama 2 7B, fine-tuned on 180K instruction sets with LoRA. The…

The 90B vision Llama. Best-in-class first-party multimodal open weight at the time of release. Workstation-class only.

YTU CE COSMOS's Llama 3 8B Turkish instruction-tuned variant. Follow-up to the original Turkish-Llama-8b that uses the Llama 3 base instead of Llama 2 — better…

Base (non-chat) variant of Trendyol's 7B Turkish LLM. The chat sibling is the more popular pick; this base version is for operators building their own…

RefinedNeuro RN TR R2 is an Apache-2.0 Llama-family 8B model distributed on Hugging Face and Ollama. It is measured alongside R1 to compare same-size…

RefinedNeuro RN TR R1 is an Apache-2.0 Llama-family 8B reasoning model distributed on Hugging Face and Ollama. It is included in the local sweep as a compact…

Llama 3.2 multimodal at 11B. Consumer-tier multimodal predecessor to Llama 4 Scout.

Llama 3.2 multimodal at 90B. Datacenter-tier predecessor to Llama 4 Maverick. Strong visual reasoning.

Phind's CodeLlama-derived coder at 34B. Older release; retained for historical / continuity value. Newer Qwen Coder lineage has surpassed it.

Google's flagship dense Gemma 4. Targets the 24GB single-GPU sweet spot.

Gemma 3 270M is the smallest member of Google's Gemma 3 family, a 270-million-parameter text-only model designed for on-device deployment and task-specific…

Pre-Gemma-4 flagship. Multimodal (4B+ variants), 128K context, 140 languages. Strong daily driver on 24GB cards.

Edge-class Gemma 4. The E stands for effective parameters: 4.5B effective and 8B in total, because of per-layer embeddings.

12B Gemma 3. Fits on 12GB consumer cards. Multimodal.

YTU's Turkish-tuned Gemma 2 9B model. It has 180 Hugging Face likes as of November 2026. Based on Turkish-Gemma-9b-v0.1 and adapted for multi-step reasoning in…

Gemma 2 2B Instruct is Google's instruction-tuned 2B model from the Gemma 2 generation, trained with knowledge distillation from larger Gemma models. It…

Trendyol LLM Asure 12B is a Gemma 3 based multimodal instruct model for Turkish and English business workflows. The public Ollama build used in local testing…

Mid-size Gemma 2. Strong chat quality with a different training mix from Llama family.

4B Gemma 3 for edge. Multimodal.

Smallest Gemma 4. Designed for phones and Raspberry-Pi-class hardware.

Gemma 4 26B MoE (4B active params) pruned from 128 to 101 experts per layer and Turkish-tuned with LoRA. After pruning the GGUF holds about 20.4B parameters.

YTU Turkish Gemma 9B v0.1 is a Gemma 2 based Turkish instruction model from the YTU CE COSMOS ecosystem. The benchmarked Ollama tag is an alibayram GGUF…

Text-only Gemma 3 for phones and IoT.

Medical-specialist Gemma fine-tune. Trained on de-identified medical literature and imaging. Use is governed by the HAI-DEF terms.

Coding-specialist Gemma. Decent FIM completion. Now mostly historical with Qwen 2.5 Coder dominating.

3B-parameter visual document retriever built on PaliGemma-3B using a ColBERT-style late-interaction objective. Encodes a PDF page as a grid of patch…

PaliGemma 2 — Gemma 2 base + SigLIP vision encoder. Designed for fine-tuning on specific vision tasks.

Mid-tier PaliGemma 2 fine-tuning base. Better baseline for complex vision tasks.

DeepSeek's April 2026 frontier flagship. 1.6T total / 49B active MoE with hybrid Compressed Sparse Attention + Heavily Compressed Attention. 1M context window.…

Open reasoning model that closed the gap with frontier proprietary reasoners. Visible chain-of-thought, MIT license, and a family of distilled smaller variants.

The cost-efficient sibling of V4-Pro. 284B total / 13B active MoE, same hybrid CSA+HCA attention, same 1M context. The published FP4 + FP8 checkpoint is 160…

Reasoning distillation onto Llama 3.3 70B. Best-in-class open-weight reasoner you can actually fit on a workstation.

32B distill — fits on a single 24GB card with reasoning capability. Best price-per-thinking-token combo for prosumers.

DeepSeek's flagship MoE — 671B total / 37B active. Server-tier, but the smaller R1 distills make this lineage approachable.

Smallest practical R1 distill. Reasoning on a 6GB GPU.

14B reasoning distill. Fits on 12GB cards.

MoE coding specialist — 16B total / 2.4B active. Fast on 12GB cards.

DeepSeek-V2-Lite-Chat is a 15.7B-total, 2.4B-active mixture-of-experts chat model using DeepSeek's Multi-head Latent Attention (MLA) architecture for KV-cache…

Full DeepSeek Coder V2. 236B total / 21B active MoE coder.

DeepSeek's first MoE — 16B / 2.8B active. Older model retained for ecosystem-context value as the base of the V2/V3 lineage.

R1 reasoning distilled into a Llama 3 8B base. Smaller R1 distill; useful when 32B is too heavy. Reasoning quality is meaningfully below the 32B distill but…

Smallest R1 distill. Surprisingly capable reasoning at 1.5B for its size class; right pick when you need reasoning AND edge deployment.

DeepSeek V2.5 — merged V2 chat + Coder. Pre-V3 baseline; 21B active MoE.

New 24B Mistral Small model under Apache 2.0. Competitive with Llama 3.3 70B at one-third the size for many tasks.

Joint Mistral/NVIDIA release with native 128K context and a new Tekken tokenizer. Strong multilingual; popular fine-tune base.

Mistral's 2024 multimodal model, which Mistral now marks as deprecated and no longer maintained. 12B parameters, vision + text, Apache 2.0. Good document and…

Mistral's coding-specialist. Strong fill-in-the-middle for IDE autocompletion. Personal/research use only.

The reference 7B from Mistral. Apache 2.0 with native function calling. Mature ecosystem.

Mistral's 123B dense model from July 2024. Open weights but restricted commercial license — research and non-commercial only.

Kumru 2B is a compact Turkish text-generation model from VNGRS. The Hugging Face config reports a Mistral-family architecture with an 8K context window, and…

Turkcell LLM 7B v1 is an Apache-2.0 Turkish text-generation model built on a Mistral architecture. The measured Ollama artifact uses a RefinedNeuro GGUF…

Mistral Turkish v2 is a public Ollama-distributed Turkish Mistral variant. The upstream Hugging Face repository was not publicly accessible during intake, so…

Malhajar Mistral 7B Turkish is an Apache-2.0 Mistral 7B Instruct v0.2 Turkish fine-tune. The benchmarked Ollama tag is a koezgen quantized distribution of the…

Mistral's Mamba (state-space) architecture coding model. Linear inference cost — the architectural alternative to attention-based coding models. Apache 2.0.

Ministral 8B with sliding-window attention and 128k context. Research license — Mistral 7B v0.3 is the commercial alternative.

Cohere's flagship — RAG-tuned, multilingual. Open weights but non-commercial.

Cohere's mid-tier — RAG and tool use. Non-commercial license.

Command R7B (December 2024) is Cohere's smallest model in the Command R family, an 8B-parameter dense transformer with 128K context, trained for…

Cohere's August 2024 Command R+ refresh. RAG-optimized; non-commercial license. Strong tool-calling and citation discipline.

Cohere's multilingual Aya at 32B. Covers 23 languages; strongest open-weight multilingual model in late 2024 — Apache-2.0 alternative is Qwen 2.5 32B but Aya…

Granite 3.1 2B Instruct is IBM's 2B-parameter dense instruct model with a 128K context window, post-trained for enterprise tasks including RAG, function…

IBM Granite at 2B. Apache 2.0 enterprise-friendly small model with safety tuning.

IBM Granite 3.3. Iterative refresh of 3.2 — same architecture; improved reasoning and instruction following. Apache 2.0.

IBM's enterprise-tuned 8B. Apache 2.0. Strong on enterprise-shaped tool-calling and structured output. Watson + RHEL ecosystem alignment.

Granite 3.0 8B — IBM's enterprise-tier baseline. Apache 2.0.

Microsoft's Phi-4 14B trained on synthetic textbook-quality data. Punches above weight on reasoning and math; MIT licensed.

Reasoning-focused fine-tune of Phi-4. Visible chain-of-thought, competitive with much larger models on math and STEM benchmarks.

Compact 3.8B Phi for edge deployment. 128K context. Strong reasoning per parameter.

Multimodal Phi 3.5. Document and chart understanding at edge size. MIT licensed.

Falcon 3 3B Instruct is TII's 3-billion-parameter instruct model from the Falcon 3 family, supporting English, French, Spanish, and Portuguese with a 32K…

Falcon 3 mid-size from TII. Permissive Falcon license; multilingual focus.

TII's Mamba (state-space) architecture model. Linear inference cost; the architectural alternative to attention-based models.

TII's Falcon 3 at the 10B tier. Supports English, French, Spanish and Portuguese, with a context length of up to 32K.

NousResearch's Hermes fine-tune of Llama 3.1 8B. Stronger system-prompt adherence, JSON output, role-play, and agent steering than the base Llama.

Hermes 3 at 70B. Workstation-tier agent-tuned model.

Nous Research's Hermes 3 fine-tune of Llama 3.2 3B. Strong general-instruction following at the 3B tier.

Eric Hartford's Dolphin fine-tune of Mistral Small 3 — uncensored, function-calling, agent-friendly.

Eric Hartford's Dolphin fine-tune at 3B. Less-censored than the base Llama; popular for unconstrained-generation use cases.

The MoE model that introduced the 8-experts pattern to the open-weight world. 47B params total, 13B active. Still a viable workhorse on 36GB+ setups.

The bigger Mixtral. 141B total / 39B active. Strong general model, workstation-tier deployment.

01.AI's 34B model. Solid bilingual EN/ZH performance, Apache 2.0.

01.AI's coding specialization at 9B. Apache 2.0; positioned as a lighter alternative to Qwen 2.5 Coder for the 16GB tier.

OLMo 2 1B Instruct is AllenAI's 1.48-billion-parameter instruct model (named 1B) from the April 2025 OLMo 2 release, post-trained with RLVR on math. It is…

AI2's fully-open 13B. Apache 2.0; full training data + checkpoints + recipes published. The reproducibility-first model in the 13B class.

OpenBMB's edge-optimized 4B. Apache-2.0 license; designed for phone deployment. Strong reasoning per parameter.

Multimodal MiniCPM at 8B. Vision + text; strong on document Q&A for the size class.

InternLM 2.5 mid-size chat. Code is Apache-2.0; the weights allow free commercial use after a license application; strong on math and Chinese.

Shanghai AI Lab's open-research line. InternLM 3 at 8B; strong on Chinese-language tasks.

Zhipu's GLM-4 at 9B. Strong on Chinese-language tasks; tool-calling format slightly different from OpenAI convention.

GLM-4 with vision encoder. Strong on Chinese document Q&A; commercial use is free after registering with the maker.

LG AI's edge-tier EXAONE. Strong Korean / English. Research-only license.

LG AI Research's flagship Korean-ecosystem model. Instruction-tuned bilingual English and Korean model. License blocks commercial use without LG agreement.

580M-parameter end-to-end OCR-2.0 model: a vision encoder paired with a Qwen-based decoder, trained specifically for general OCR including math formulas (LaTeX…

DeepSeek's multimodal 7B. Decoupled visual encoding for understanding vs generation — different from typical VLM design.

Tencent's frontier MoE. 389B total / 52B active. License permits commercial use with restrictions on companies above MAU thresholds.

Fully-open coding model — training data + recipes published. Released under INF's permissive model license, which allows commercial use, with verifiable…

RWKV 7 'Goose' at 1.5B. Linear-time inference architecture (constant memory regardless of context). Apache 2.0.

Medical / biomedical fine-tune of Llama 3 70B. Strong on USMLE and clinical-knowledge benchmarks; right pick when domain-specific medical depth matters more…

all-MiniLM-L6-v2 is a 22M-parameter sentence-transformers embedder producing 384-dim vectors with a 256-token context, built on a 6-layer cut of Microsoft's…

12B-parameter rectified-flow transformer for text-to-image, guidance-distilled from the FLUX.1 [pro] teacher. Among the most-liked models on Hugging Face…

Nomic Embed Text v1.5 is a 137M-parameter English embedding model with an 8192-token context window, trained with Matryoshka Representation Learning so the…

BGE Large EN v1.5 is the 335M-parameter English flagship from BAAI's FlagEmbedding family, producing 1024-dim embeddings with a 512-token context window.…

82M-parameter StyleTTS2-derived TTS that went viral in early 2025; its card says it delivers comparable quality to larger models. Apache-2.0 weights, dozens of…

BGE Reranker v2-M3 is a 568M-parameter cross-encoder built on XLM-RoBERTa-large by BAAI, designed for second-stage reranking with an 8192-token context over…

Coqui's flagship multilingual voice-cloning TTS — clones a speaker from a 6-second reference clip and synthesizes in 17 languages with cross-lingual transfer.…

all-mpnet-base-v2 is a 109M-parameter sentence-transformers embedder based on Microsoft's MPNet, producing 768-dim vectors with a 384-token context. Trained on…

244M-parameter Whisper. The smallest Whisper checkpoint considered 'production grade' for non-English audio. Sweet spot for laptops with iGPU/Metal or modest…

74M-parameter Whisper variant — roughly 2x the params of tiny, with WER 12.4% vs 16.9% on LibriSpeech test-other in the [Whisper…

paraphrase-multilingual-MiniLM-L12-v2 is a 118M-parameter multilingual sentence-transformers embedder built on a knowledge-distilled MiniLM-L12, producing…

Smallest member of the Whisper encoder-decoder ASR family (39M params). Trained on 680k hours of weakly supervised multilingual audio. Targets sub-realtime…

Multilingual E5 Large Instruct is a 560M-parameter XLM-RoBERTa-large encoder fine-tuned by Microsoft's intfloat team with task instructions appended to…

mxbai-embed-large-v1 is a 335M-parameter BERT-large English embedder from Mixedbread AI, producing 1024-dim vectors and supporting dimension truncation (the…

12B rectified-flow transformer, timestep-distilled to 1-4 sampling steps, released under Apache-2.0. Same architecture as FLUX.1 [dev] but trades a bit of…

Jina Embeddings v3 is a 572M-parameter multilingual encoder with 8192-token context and five task-specific LoRA adapters (retrieval-query, retrieval-passage,…

Moonshot's long-context, agent-oriented MoE. Optimized for stability under tool use and multi-step coding/planning workflows.

878M-parameter Shape-Optimized vision-language encoder (428M vision tower, 450M text tower) trained with the sigmoid (not softmax) contrastive loss on WebLI.…

Arctic Embed L v2.0 is a 568M-parameter multilingual embedder from Snowflake based on XLM-RoBERTa, producing 1024-dim Matryoshka vectors with an 8192-token…

Jina Reranker v2 Base Multilingual is a 278M-parameter cross-encoder from Jina AI with a 1024-token context, trained on 100+ languages plus code and structured…

2.6B SDXL backbone trained with Adversarial Diffusion Distillation (ADD), producing photorealistic 512px images in a single forward pass. Designed for…

756M-param distilled Whisper-large-v3 with the decoder shrunk from 32 to 2 layers. 6.3x faster than the teacher and within 1% WER on long-form audio, per the…

SmolLM2-135M-Instruct is the smallest instruction-tuned model in Hugging Face's SmolLM2 family, a 135M-parameter Llama-architecture model trained for on-device…

E5-Mistral-7B-Instruct is a 7.11B-parameter decoder-based embedder fine-tuned from Mistral-7B by Microsoft's intfloat team, producing 4096-dim embeddings with…

GTE ModernBERT Base is a 149M-parameter English embedder built on AnswerDotAI's ModernBERT backbone, producing 768-dim vectors with native 8192-token context…

770M-parameter unified vision foundation model with a DaViT image encoder and BART-style seq2seq decoder. One model, one set of weights — handles captioning,…

31B-parameter Turkish-tuned reasoning model with i1-imatrix quantizations by mradermacher. Designed for step-by-step problem solving in Turkish.

2.5B MMDiT-X with Query-Key Normalization and dual attention blocks in the first 12 transformer layers. Trained for 0.25-2MP output. Positioned as the mid-tier…

mxbai-rerank-large-v2 is a 1.54B-parameter reranker from Mixedbread AI built on Qwen2.5-1.5B, supporting 100+ languages and a context of up to 8K tokens (32K…

VITS-based neural TTS optimized for Raspberry Pi-class hardware. Ships as ONNX checkpoints with 178 voice models across 53 languages as of November 2026.…

InclusionAI's Ring-2.6-1T is a 1 trillion parameter Mixture-of-Experts model with 8 of 256 routed experts active per token, released by Ant Group's AI research…

600M-parameter FastConformer-TDT transducer ASR from NVIDIA NeMo. Topped the Hugging Face Open ASR Leaderboard in 2025 for English, with an average WER of…

Turkish-from-scratch language model trained by Ali Safaya (Koç University researcher). Named after the kanarya (Turkish for 'canary'). Trained on a large-scale…

SmolLM2-360M-Instruct is the middle tier of the SmolLM2 instruct family, a 360M-parameter Llama-architecture model with an 8K context. It is shipped with ONNX…

Flow-matching non-autoregressive TTS built on a Diffusion Transformer (DiT) backbone with ConvNeXt text refinement. Trained on the 100K-hour Emilia dataset;…

Smaller Kanarya variant — 750M parameters. Runs on CPU or 4GB GPU comfortably. Useful for low-resource Turkish text classification, embeddings, or completion…

Turkish BART-style sequence-to-sequence model fine-tuned specifically for summarization. Not a chat model — purpose-built for input-document → Turkish-summary…

GPT-2 Large architecture trained from scratch on Turkish. Reference baseline for measuring how much modern instruction-tuned models actually improve on the…

SmolVLM-Instruct is Hugging Face's compact vision-language model built on the Idefics3 architecture, pairing SmolLM2-1.7B-Instruct with a SigLIP-SO400M vision…

LLaMA-architecture 3.78B model fine-tuned as a TTS that emits SNAC audio tokens. Designed for highly expressive, emotion-controllable speech with laughter,…

OpenELM-3B-Instruct is Apple's 3-billion-parameter instruct model using a layer-wise scaled transformer with varying FFN multipliers and KV-head counts across…

SmolLM 2 flagship. Open data + open weights at the edge tier.

Hugging Face's SmolLM 2 at 360M. Apache 2.0; targets phone / Pi-class deployments.

HuggingFace's small-model line at 3B. Apache 2.0. Designed for edge / educational deployments.

BAAI's multilingual embedding flagship. Dense + sparse + ColBERT-style multi-vector. The de-facto open multilingual embedding pick.

InternVL 2.5 mid-tier — Shanghai AI Lab vision-language model with strong document and chart understanding.

InternVL 2.5 flagship. Approaches frontier proprietary VLMs on document and OCR tasks.

Aya 23 at 35B. Built on Cohere's Command-R lineage. Non-commercial.

Cohere's multilingual research model covering 23 languages. CC-BY-NC — research only.

Stability AI's 12B. Stable LM line; commercial use is free under the Stability AI Community License below USD 1M annual revenue. Solid baseline at 12B class.

OpenAI's flagship open speech-to-text model. 100 languages, MIT license. The de-facto open ASR baseline.

Pruned and fine-tuned Whisper Large v3 with 4 decoder layers instead of 32. [OpenAI lists](https://github.com/openai/whisper) ~8x relative speed with a minimal…

NVIDIA's research-grade embedding model. Mistral-7B base. Top of MTEB at release.

Tiny vision-language model. ~1.9B; designed for edge / embedded multimodal use cases. Apache 2.0.

NVIDIA's edge-tier Nemotron. Distilled from Minitron lineage with role-play tuning.

AI2's fully-open post-training recipe applied to Llama 3.1 8B. Open data, open code, open weights.

LLaVA-OneVision unified single-image / multi-image / video VLM on Qwen 2 base.

LLaVA 1.6 on Mistral 7B base. Apache 2.0 vision-language with strong OCR.

Mid-size StarCoder 2. The 8GB-VRAM autocomplete pick.

Jamba flagship at 398B total / 94B active. Frontier hybrid-architecture model with 256k context.

AI21's hybrid Mamba-Transformer MoE. 256k context with the SSM throughput advantage.

Molmo flagship. Apache 2.0 VLM rivaling proprietary models on UI pointing and visual reasoning.

AI2's fully-open VLM. Trained on PixMo dataset; pointing capability for UI grounding.

Tulu 3 at 70B. AI2's fully-open instruct fine-tune — research transparency at scale.

StarCoder 2 flagship. The largest BigCode coder; 16k context with strong fill-in-middle.

BigCode's StarCoder 2 at 3B. Trained on 17 programming languages from The Stack v2.