Qwen · 128K tokens context
Compact Qwen3 powerhouse — everyday coding, multilingual analysis, and step-by-step reasoning at fast inference speeds. Uses hybrid think...
Browse the model records in the LITHOVEX catalog. Each profile links to available specifications, licensing, pricing and benchmark records; missing values are identified as unlisted.
Data last updated: 2026-10-16. Confirm current details with model providers.
Compact Qwen3 powerhouse — everyday coding, multilingual analysis, and step-by-step reasoning at fast inference speeds. Uses hybrid think...
Mid-size Qwen3 — balanced speed and intelligence. Strong on coding, analysis, and multilingual chat tasks.
32B dense Qwen3 — excels at complex coding, mathematics, and deep multi-step reasoning. Top performer in its size class.
MoE architecture — 30B total with only 3B active params. Efficient & smart: near 32B quality at near 3B speed.
July 2025 Qwen3 48B instruct — updated instruction following, improved coding, and enhanced reasoning.
Flagship Qwen3 — 235B total, 22B active. World-class reasoning, elite coding, 100+ languages. Rivals GPT-4o on most benchmarks.
July 2025 updated Qwen3-235B — improved instruction following and enhanced capabilities over the base release.
Thinking-mode Qwen3-235B — always uses extended chain-of-thought. Best for math olympiad problems, scientific reasoning, and agentic tasks.
Latest Qwen3 vision-language model — understands images, charts, diagrams, and screenshots. Great for UI analysis and image-to-code.
Largest Qwen vision model — 235B with extended thinking. Elite for complex visual reasoning, scientific diagrams, and multimodal research.
Next-gen Qwen Coder — agentic software engineering, repository-level code understanding, function calling, and autonomous bug fixing.
FP8-quantized Qwen3 Coder Next — near-identical quality at lower cost and faster inference.
MoE code specialist — 30B total, 3B active. Excellent for agentic coding workflows, debugging, and multi-step programming tasks.
Largest Qwen Coder — 480B params, 35B active, 256k context. The flagship for full-codebase agentic software engineering.
FP8 version of the flagship Qwen3 Coder 480B — same elite capabilities at reduced inference cost.
Qwen3 Next 80B MoE — strong general-purpose model with 80B knowledge compressed into 3B active params.
Thinking-mode Qwen3 Next 80B — extended chain-of-thought reasoning for math, science, and complex analysis.
Previous-gen Qwen2.5 7B — reliable, well-tested, excellent for general instructions, code, and structured output.
Dense 72B Qwen2.5 — previous flagship, excellent across coding, writing, math, and analysis tasks.
Dedicated code model trained specifically on code. Best fast option for programming, bug fixing, and code review.
State-of-the-art code model from Qwen2.5 — leads HumanEval and coding benchmarks for 32B class.
Vision-language model — understands images, charts, documents, and screenshots. Great image-to-code and visual Q&A.
Expert-level vision-language model — understands complex images, charts, and documents with deep reasoning capabilities.
Qwen3.5 9B — updated general-purpose model with improved instruction following and creative capabilities.
Qwen3.5 27B — well-rounded model excelling at creative writing, coding, and analytical tasks.
MoE Qwen3.5 35B — 35B total, 3B active. Excellent general-purpose capabilities at MoE efficiency.
Large MoE Qwen3.5 — 122B total, 10B active. Top-tier reasoning and writing at remarkable efficiency.
Massive Qwen3.5 MoE — 397B total, 17B active. Near-GPT-4o quality at significantly reduced inference cost.
Groundbreaking open reasoning model — rivals OpenAI o1 on math, science, and coding. Transparent chain-of-thought reasoning.
May 2025 update to DeepSeek-R1 — improved reasoning depth, better code generation, and enhanced benchmark scores.
DeepSeek flagship chat & code model — top-tier on coding benchmarks, long context tasks, and instruction following.
March 2024 DeepSeek-V3 checkpoint — improved coding and reasoning over the original V3 release.
DeepSeek-V3.1 — next iteration with improved instruction following, better coding, and stronger general capabilities.
Terminus edition of DeepSeek-V3.1 — enhanced for agentic workflows, tool use, and complex multi-step tasks.
Experimental V3.2 — leading SWE-bench scores. Frontier-level coding agent. Not stable release but exceptional performance.
Tiny R1 distillation — astonishing reasoning at 1.5B params. Perfect for local deployment and fast prototyping.
Fast chain-of-thought reasoning in a compact 7B package. Excellent math, logic, and code at low latency.
Powerful R1 distillation in a 32B package — matches GPT-4-class performance on math, science, and complex reasoning.
Llama-based R1 distillation — fast reasoning using Llama-3 architecture backbone. Broadly compatible.
Large Llama-based R1 distillation — powerful reasoning across math, science, code, and complex analysis.
Ultra-compact Llama 3.2 — 1B params with 128k context. Ideal for classification, extraction, and simple tasks at maximum speed.
Meta's widely-used 8B open model with 128k context. Great general assistant, widely compatible, strong instruction following.
Strong 70B Llama with 128k context — great for complex instruction following, writing, and coding tasks.
Latest Llama-3.3 70B — exceptional instruction following, coding, and multilingual capabilities across 8 languages.
Original Llama-3 8B — battle-tested, widely supported, great for general chat and simple coding tasks.
Original Llama-3 70B — powerful open-weight model for complex reasoning, writing, and coding. 8k context.
Llama-4 Scout — MoE model with 10M token context. Unmatched for massive documents, full codebases, and ultra-long tasks.
Llama-4 Maverick — 128-expert MoE, 1M context, multimodal. Top performance on coding, reasoning, and visual tasks.
Safety classifier — detects unsafe content in LLM inputs/outputs across 13 harm categories. Use for content moderation.
Ultra-compact edge model — runs on phones/devices. Best for simple Q&A, quick summaries, on-device applications.
Google's Gemma-3 27B — excellent coding assistant with strong reasoning, multilingual support, and 128k context.
Large Gemma-4 MoE — 268B total, 4B active. Google's open model with multimodal capabilities and strong reasoning.
Google's largest open model — 318B. State-of-the-art on reasoning, coding, knowledge tasks, and multimodal understanding.
GLM-5.1 with FP8 quantization — advanced reasoning model with strong Chinese/English bilingual performance.
GLM-5 flagship — powerful reasoning model with elite Chinese/English bilingual capabilities and agentic tool use.
Lightweight GLM-4.5 variant — fast responses for everyday tasks with excellent Chinese language support.
Full GLM-4.5 — flagship Zhipu model with deep Chinese/English reasoning and agentic capabilities.
Vision-capable GLM-4.5 — understands images alongside text. Strong Chinese visual tasks and document analysis.
GLM-4.6 — strong Chinese/English bilingual model with tool use, function calling, and agentic capabilities.
Vision-capable GLM-4.6 — understands images in Chinese and English. Great for Chinese visual tasks.
Fast vision GLM-4.6 variant — quick image understanding for high-throughput visual applications.
GLM-4.7 — next iteration with improved reasoning, better code generation, and enhanced bilingual performance.
FP8-quantized GLM-4.7 — near-identical quality to GLM-4.7 at faster inference and lower cost.
Dense 32B GLM-4 — strong performer for coding, math, and complex Chinese/English reasoning tasks.
Phone/device automation specialist — designed for mobile UI interaction, app control, and device automation tasks.
Compact Cohere model — optimized for RAG, document search, and grounded generation at fast inference speeds.
Cohere Command R — powerful RAG and retrieval-focused model with excellent grounded generation and citation.
Cohere's enterprise flagship — 111B params, 256k context. Elite for long documents, agentic tasks, and enterprise RAG.
Command A with extended reasoning — combines Cohere's enterprise features with chain-of-thought reasoning.
Command A specialized for translation — enterprise-grade multilingual translation across 23+ languages.
Command A with vision — processes images with enterprise-grade grounded generation. Great for document analysis.
Cohere's multilingual specialist — state-of-the-art performance in 23 languages. Ideal for global content generation.
Multilingual vision model — understands images in 23 languages. Unique cross-lingual visual capability.
Moonshot AI's flagship — 1T MoE param model with exceptional agentic coding, reasoning, and software engineering.
September 2025 Kimi-K2 — updated with improved agentic performance and enhanced coding capabilities.
Kimi-K2 with extended thinking — chain-of-thought for deep math, science, and complex problem solving.
Faster Kimi variant — strong K2 capabilities at reduced inference cost. Great for agentic coding tasks.
MiniMax M1 reasoning model — linear attention architecture enabling very long context at scale with strong reasoning.
MiniMax M2 — improved reasoning and coding over M1 with better instruction following and general capabilities.
MiniMax M2.1 — incremental improvement over M2 with better coding performance and enhanced reasoning.
MiniMax M2.5 with 1M context — one of the longest context models available. Excellent for massive document analysis.
Latest MiniMax flagship — M2.7 with 1M+ context, frontier reasoning, and strong coding capabilities.
Microsoft/OpenAI's open-source GPT model — 20B params optimized for chat, Q&A, and coding tasks.
Microsoft/OpenAI's large OSS model — 120B params with strong performance on coding and reasoning benchmarks.
Safety-focused 20B model — content moderation, policy enforcement, and safety classification tasks.
Function-calling specialist — excellent at JSON mode, tool use, and structured outputs. Best for agentic workflows.
Community creative writing model — fine-tuned Llama-3-8B for storytelling, roleplay, and creative fiction.
Community creative powerhouse — fine-tuned Llama-3-70B for storytelling, character creation, and creative writing.
Community creative 8B model — focused on expressive writing, emotional depth, and character immersion.
Fully open-source OLMo-3 — transparent training data, weights, and code. Ideal for research and reproducibility.
WizardLM-2 8x22B MoE — excellent at following complex instructions, creative writing, and complex reasoning tasks.
Tiny model router — classifies requests and routes to appropriate models. Use for intent classification and request routing.
Baidu's ERNIE 4.5 21B MoE — efficient Chinese language model with broad knowledge and coding capabilities.
Baidu's flagship ERNIE — 300B MoE with elite Chinese language capabilities and broad reasoning.
Massive vision-language ERNIE — 424B MoE with vision capabilities. Elite Chinese multimodal understanding.
Southeast Asian language specialist — optimized for Thai, Vietnamese, Indonesian, Malay, Filipino, and more.
Qwen-based SEA-LION — stronger baseline with Qwen2.5 backbone for Southeast Asian languages and code.
Community instruction-tuned model — general-purpose chat and writing assistant.
Tiny Aya model — ultra-fast multilingual responses for simple tasks and high-throughput applications.
Tiny global Aya — multilingual quick responses with global language coverage at maximum speed.
Tiny Aya Earth — lightweight multilingual model for simple grounded tasks at minimal cost.
Tiny Aya Fire — fast multilingual model focused on creative and expressive responses.
KAIST AI's 70B instruction model — strong reasoning and academic task performance.
European multilingual LLM — optimized for EU languages with strong performance across 24 European languages.
Hebrew-specialized LLM with thinking mode — best model for Hebrew language tasks with chain-of-thought reasoning.
Meta's compact 8B chat model with 128k context — fast, broadly compatible, and a solid everyday assistant for general Q&A and instruction following.
Mistral's polished 7B instruct model — efficient European open-weight chat with strong reasoning per parameter and native function-calling.
Microsoft's tiny 3.8B Phi-3 — punches far above its weight on reasoning and code, ideal for low-latency assistants and on-device deployment.
Google's open Gemma-2 9B instruct — derived from Gemini research, excellent at concise answers, summaries, and lightweight chat.
Community-trained OpenChat 3.5 — outperforms ChatGPT-3.5 on MT-Bench at just 7B params. Friendly, conversational tone.
Zephyr 7B Beta — helpful assistant fine-tuned from Mistral-7B, known for clear, well-structured answers and great instruction following.
TII Abu Dhabi's open Falcon 7B — reliable general-purpose chat trained on 1.5T multilingual tokens. Great baseline assistant.
Meta's BART-Large fine-tuned on CNN/DailyMail — gold-standard for fluent abstractive summarization of news articles and long text.
Google's PEGASUS trained on the XSum dataset — produces extremely concise one-sentence summaries that capture the gist of a document.
BART-Large fine-tuned on the SAMSum dialogue dataset — purpose-built for summarizing chat conversations, meetings, and Slack threads.
Meta's WMT19 winning English-to-German translator — production-grade quality, especially strong on news and formal writing.
Helsinki-NLP's MarianMT trained on the OPUS corpus — fast, lightweight English-to-Chinese translation that runs well on CPU.
Helsinki-NLP MarianMT for English-to-Spanish — natural, fluent translations of everyday text, articles, and short documents.
Helsinki-NLP MarianMT for English-to-French — high-quality formal and conversational translation with proper accents and grammar.
Helsinki-NLP MarianMT for English-to-Russian — handles Cyrillic morphology and produces natural, grammatical translations.
T5-base fine-tuned to fix grammar, spelling, and punctuation — a Grammarly-style cleanup pass for sentences and short paragraphs.
Flan-T5 Large tuned for full sentence rewriting — corrects grammar while preserving meaning, useful for ESL writing and email polish.
Cardiff NLP's RoBERTa fine-tuned on 124M tweets — classifies text as positive/neutral/negative. Best-in-class for social-media sentiment.
Multilingual BERT trained to predict 1–5 star product-review ratings across English, Dutch, German, French, Spanish, and Italian.
DistilBERT fine-tuned on the SST-2 movie-review benchmark — fast, lightweight binary positive/negative sentiment classifier.
Alibaba's Qwen2.5 specialized for mathematics — excels at arithmetic, algebra, calculus, and step-by-step word problems with chain-of-thought.
Llama-3 8B extended to a 1M-token context window by Gradient AI — read entire books, codebases, or massive log files in one prompt.
Nous Research's Hermes 2 Pro — Llama-3 tuned for creative writing, role-play, and structured JSON/function-call output. Great storyteller.
DeepSeek's compact code model trained on 2T tokens of code — strong at completion, refactoring, and bug fixes across 80+ languages.
BigCode's StarCoder2 15B — trained on the Stack v2 with 600+ programming languages. Great for code completion and repository-level tasks.
Meta's Code Llama 7B Instruct — Llama-2 specialized for code generation, with strong infill (FIM) support and 16k context.
BigCode's 1.1B SantaCoder — small, fast multi-language code model (Python/Java/JS) ideal for inline IDE completion at low latency.
Mistral's flagship 22B code model — speaks 80+ languages, excellent at code generation, debugging, fill-in-the-middle, and tests.
Original BigCode StarCoder 15B trained on permissively-licensed GitHub code — strong at completion, docstring generation, and refactoring.
WizardLM's Evol-Instruct tuned StarCoder — exceptionally good at explaining code, writing test cases, and producing well-commented snippets.
Defog's SQLCoder 7B — generates accurate SQL from natural-language questions over your database schema. Beats GPT-4 on SQL benchmarks.
BigCode's tiny 1B StarCoder base — quick shell-command suggestions, regex generation, and small one-line code completions.
Salesforce BLIP-Large — describes images in natural English. Reliable workhorse for alt-text, image search, and content moderation.
ViT vision encoder paired with GPT-2 — fast, lightweight image captioning ideal for batch tagging and accessibility workflows.
Surya's high-accuracy OCR — extracts text with bounding boxes from scanned pages, screenshots, and photos in 90+ languages.
Transformer-based YOLO — detects and labels objects in images using the COCO 80-class taxonomy. Compact and fast.
Meta's DETR with a ResNet-50 backbone — end-to-end transformer object detector. Robust on cluttered scenes and complex layouts.
Peking University's real-time DETR — sub-30ms object detection at SOTA accuracy. Great for live video and edge inference.
Meta's Segment Anything Model — click any point or draw a box, and SAM produces a precise pixel-level mask of that object.
Depth Anything V2 — produces accurate monocular depth maps from a single 2D image. Useful for AR, 3D, and background blur effects.
Swin Transformer-based super-resolution — upscales real-world photos 4× while restoring detail, sharpening edges, and reducing noise.
Bria AI's RMBG 1.4 — best-in-class background removal, producing clean transparent PNG cutouts of people, products, and objects.
LLaVA-NeXT extended to video — analyses short clips frame-by-frame and answers natural-language questions about what happens.
Microsoft X-CLIP — zero-shot video classification. Match a video to any natural-language label without retraining the model.
Meta's TimeSformer fine-tuned on Kinetics-400 — recognises 400 human actions in video clips (running, cooking, sports, etc.).
Alibaba's ModelScope T2V 1.7B — generates short video clips from text prompts. The original open-source text-to-video model.
Cerspense Zeroscope V2 — produces 576×320 watermark-free videos from text. A community favourite for short artistic clips.
Meta's VideoMAE — masked autoencoder for video understanding. Foundation for action recognition and video classification tasks.
VideoMAE fine-tuned on Kinetics-400 — accurate human-action recognition across 400 daily-life and sports activities.
Google's FILM — synthesises smooth in-between frames from two photos. Turns a burst of stills into buttery-smooth slow-motion video.
Video-LLaVA v1.5 — unified vision-language model that captions short videos and answers questions about visual events over time.
OpenAI's flagship Whisper Large V3 — state-of-the-art multilingual speech-to-text. Highly accurate transcription across 99 languages.
OpenAI's Whisper Medium — balanced multilingual transcription model. Faster than Large with only a small accuracy trade-off.
Distil-Whisper — 6× faster than Whisper Large V3 and 49% smaller, with within-1% transcription accuracy. English-focused.
Meta's Wav2Vec2 trained on 960h of LibriSpeech — fast, lightweight English speech-to-text. Great when you need low-latency transcription.
Suno's Bark — generative text-to-audio that produces realistic speech, music, ambience, and even non-verbal sounds like laughter or sighs.
Meta's Massively Multilingual Speech — natural-sounding English text-to-speech. Part of a 1000+ language TTS family.
Microsoft SpeechT5 — unified speech model for text-to-speech with speaker embedding support. Generate audio in any reference voice.
ESPnet VITS trained on LJSpeech — produces clear, natural single-speaker English TTS. A research-grade open baseline.
Coqui XTTS v2 — clones any voice from just 6 seconds of audio. Multilingual TTS in 17 languages with cross-lingual voice transfer.
MyShell OpenVoice — instant voice cloning with fine-grained control over emotion, accent, rhythm, and style. 1-shot voice replication.
MIT Audio Spectrogram Transformer fine-tuned on AudioSet — classifies 527 sound events (music, speech, traffic, animals, etc.).
Meta's MusicGen Small — generates short pieces of music from a text prompt or melody hum. The fast tier of Meta's music model family.
Stability AI's Stable Audio Open — generates up to 47 seconds of music, sound effects, or ambience from a text description.
Microsoft SpeechT5 voice conversion — transforms a recording so it sounds like it was spoken by a different target voice.
Pyannote's speaker diarization pipeline — identifies who spoke when in a multi-speaker recording (meetings, interviews, podcasts).
Microsoft TrOCR for printed text — transformer OCR model that extracts text from clean printed documents, books, and screenshots.
Microsoft TrOCR fine-tuned on handwriting — reads handwritten notes, journals, and forms with strong accuracy across cursive and print.
Microsoft LayoutLMv3 — combines text, layout, and image understanding for parsing forms, receipts, invoices, and structured documents.
Naver Clova's Donut — OCR-free document understanding. Answers natural-language questions about scanned PDFs, receipts, and forms.
Impira's LayoutLM Document QA — extract field values (e.g. invoice total, vendor, date) from PDFs by simply asking a question.
Microsoft Table Transformer — finds tables inside PDFs and document images and isolates them for downstream extraction.
Donut fine-tuned on RVL-CDIP — classifies document images into 16 types (letter, invoice, resume, scientific paper, etc.).
Meta's Nougat — converts academic PDFs into clean Markdown, preserving math equations, tables, and inline code blocks.
Microsoft LayoutLMv2 — earlier-generation document AI well-suited for invoice parsing, key-value extraction, and form filling.
Google DePlot — translates a chart or plot image into a structured table of underlying numbers. Pair with an LLM for chart Q&A.
Microsoft's original LayoutLM — joint text + layout model for understanding forms, expense reports, and visually-rich documents.
Together AI × Nous Research StripedHyena 7B — Hyena/state-space hybrid architecture optimized for ultra-long-context document chat.
Salesforce BLIP-2 with OPT-2.7B — answers natural-language questions about an image (visual question answering).
Surya's text-line detector — locates text regions and signature blocks on scanned pages. Pairs with OCR for full document parsing.
Flan-T5 base fine-tuned on SAMSum — produces concise bullet-point summaries of meeting transcripts and chat conversations.
LLaVA 1.5 7B — open multimodal chat model that can see images. Describe pictures, read screenshots, or analyse charts conversationally.
Alibaba's Qwen2-VL 7B — strong open vision-language model. Reads documents, charts, and UI screenshots with chat-style follow-up.
CogVLM2 19B built on Llama-3 — high-resolution image understanding with strong OCR and detailed visual reasoning capabilities.
IDEFICS2 8B — multimodal chat model that handles arbitrary interleavings of text and images in the conversation.
Salesforce InstructBLIP on Vicuna 7B — strong visual reasoning and detailed image-grounded answer generation.
DeepSeek's 7B model specialized for mathematics — strong on algebra, calculus, and competition problems with step-by-step solutions.
Allen AI UnifiedQA — single model that handles extractive, multiple-choice, yes/no, and abstractive question answering uniformly.
Google Flan-T5 XXL (11B) — instruction-tuned encoder-decoder. Excellent for fact-style Q&A, classification, and zero-shot reasoning.
Llama-3 8B extended to 4M tokens of context by Gradient AI — chat over entire libraries of documents in a single window.
OpenAssistant's flagship open SFT model on Pythia 12B — community-built ChatGPT-style assistant trained on the OASST conversations dataset.
Berkeley Gorilla OpenFunctions v2 — purpose-built for accurate API function calling and tool use, rivaling GPT-4 on the Berkeley Function-Calling leaderboard.
MeetKai Functionary — chat model fine-tuned to reliably emit function-call JSON. Great for tool-using agents and structured output.
Infly INF 34B — bilingual English/Chinese chat model strong at structured data analysis, table reasoning, and business Q&A.
BigCode's instruction-tuned StarCoder2 15B — agentic auto-coding, multi-file refactors, and natural-language-to-code generation.
LITHOVEX flagship — the smartest, most capable model in the lineup. Effortlessly handles code, deep reasoning, writing, math, research, planning, and everyday questions.
LITHOVEX engineering powerhouse — surgical code specialist tuned for code generation, debugging, refactors, and software architecture.
Latest OpenAI flagship (2026). Elite multi-step coding, research, and agentic workflows. Surpasses Claude Opus 4.7 on benchmarks.
Anthropic's most powerful model ever. Deeper reasoning, stronger agentic coding, superior design sense, and sharper analytical precision than Opus 4.7.
Anthropic's newest frontier model — The Narrative Architect & Reasoning Virtuoso. Expert at constructing meaning through narratives, causal chains, systems, and abstractions.
Excels in design, aesthetic quality, and nuanced reasoning. Perfect for agentic coding and complex long-running tasks.
Google's flagship multimodal model. Exceptional reasoning, code generation, and long-context understanding with native multimodal capabilities.
OpenAI GPT-5.4 Pro — professional-grade reasoning and coding with extended context and enhanced accuracy.
Claude Sonnet 4.6 — Anthropic's fast, balanced model. High intelligence with low latency for everyday tasks and production use.
DeepSeek-R1 extended thinking model — chain-of-thought reasoning for math, science, code, and complex analytical problems.
OpenAI o3 — frontier reasoning model with extended chain-of-thought. Top performance on math, science, and complex problem-solving.
Grok 4.1 Fast Reasoning by xAI — high-speed reasoning model with 2M token context window for large codebases and documents.
Moonshot Kimi K2.6 — agentic specialist model with ultra-long context (1M tokens). Exceptional at multi-step tool use and autonomous task completion.
Claude Opus 4.6 Pro — Anthropic's high-capability model for complex analysis, writing, and extended agentic coding tasks.
MiniMax M2.7 Pro — high-capability Chinese frontier model with strong multilingual support and professional-grade reasoning.
Zhipu GLM 5 Pro — Chinese frontier model with strong multilingual capabilities and deep reasoning for complex tasks.
Blackbox E2E Encrypted — privacy-first AI model with end-to-end encryption. Ideal for sensitive code, confidential analysis, and private workflows.
Claude Sonnet 4.5 — fast and capable everyday model by Anthropic, balancing speed and intelligence for production use.
OpenAI GPT-5.4 — highly capable general-purpose model with strong reasoning, code generation, and multimodal understanding.
OpenAI GPT-5.4 Nano — ultra-fast lightweight model optimized for speed and efficiency without sacrificing quality.
OpenAI GPT-5.3 Codex — specialized code generation model with deep software engineering expertise and multi-language support.
Google Gemini 3.5 Flash — ultra-fast multimodal model optimized for speed and high-throughput applications.
Google Gemini 3.1 Flash Lite — the most efficient Gemini model, optimized for cost and speed in high-volume use cases.
Google Gemma 4 26B — open-weights model from Google. Excellent for local deployment, fine-tuning, and research with strong general capabilities.
Google Gemma 4 31B — larger open-weights Gemma variant with enhanced reasoning and code capabilities.
Grok 4.3 by xAI — highly capable general model with real-time information access and exceptional reasoning performance.
Grok Build 0.1 — xAI's specialized coding and building model tuned for software engineering and rapid prototyping.
Grok Code Fast by xAI — ultra-fast code-focused model optimized for quick coding tasks, autocomplete, and rapid iteration.
NVIDIA Nemotron 3 Ultra — high-performance open model optimized for enterprise AI, reasoning, and agentic workflows.
NVIDIA Nemotron Nano 30B — compact yet powerful model with strong reasoning and coding in a 30B parameter footprint.
NVIDIA Nemotron Nano 12B VL — compact vision-language model for multimodal tasks with efficient performance.
MiniMax M2.5 — capable multilingual model with strong long-context handling and reasoning for complex tasks.
MiniMax Free — accessible free-tier model with solid multilingual capabilities and everyday task performance.
DeepSeek V4 Flash — ultra-fast variant of DeepSeek V4 for quick reasoning and code generation with minimal latency.
DeepSeek V4 Pro — professional-grade DeepSeek model with enhanced reasoning, coding, and mathematical capabilities.
Meta Llama 3.2 1B — ultra-lightweight open model ideal for on-device inference, edge deployment, and low-resource environments.
Amazon Nova 2 Lite — Amazon's efficient multimodal model optimized for cost-effective reasoning and document understanding.
Amazon Nova Micro — Amazon's smallest, fastest model for text-only tasks. Ultra-low latency for high-volume text processing.
Devstral 2 by Mistral — specialized agentic coding model designed for software engineering agents and automated code workflows.
Pixtral Large by Mistral — powerful vision-language model capable of understanding and reasoning over multiple images with native multimodality.
Ministral 3B by Mistral — compact and efficient model for edge and on-device deployments with solid everyday performance.
Ministral 8B by Mistral — fast and capable mid-size model balancing performance and efficiency for production use.
Mistral Small — Mistral's compact but capable production model for everyday tasks with fast response times.
Pixtral 12B — Mistral's compact multimodal model with vision understanding and image reasoning in a 12B parameter footprint.
Codestral by Mistral — dedicated code model trained specifically for software engineering across 80+ programming languages.
Mistral Medium 3.5 — powerful mid-size model with strong reasoning, code, and vision capabilities at a balanced compute cost.
Morph V3 Fast — ultra-fast code editing model specialized for applying precise code diffs and edits with minimal latency.
Morph Large — Morph's most powerful code editing model for complex code transformations, large file edits, and agentic code workflows.
Trinity Large Thinking — extended chain-of-thought reasoning model for math, science, and hard analytical problems requiring deep step-by-step reasoning.