Models
A reference catalog of the AI models we use, recommend, or compare in the playbook — what they're for, where to try them, when we last verified each entry.
Showing 103 of 103 models.
GPT-6.1 Terra
A fast, efficient GPT-6 model for visual reasoning, precise computer use, and clear, natural writing. Priced between GPT-6 Luna and GPT-6.1 Sol.
Cohere Rerank 5
Cohere's flagship reranker family (rerank-v5.0-pro, rerank-v5.0-fast). The standard second-pass model after a vector or BM25 retrieval — multilingual, 32K per-doc context, calibrated relevance scores, and optional Context Focus that returns the relevant passages inside each document.
MiniMax M3.1 Flash
MiniMax's low-cost model for agents, coding and computer use, at $0.15 in / $0.60 out per MTok. API-only for now; MiniMax says the weights will be released later.
FLUX 3 Image Klein
Black Forest Labs' small open-weights FLUX 3 image model — one 5.6B model for text-to-image, photo editing and multi-reference generation with up to 10 reference images, built to run on consumer graphics cards. Supersedes FLUX.2 [dev] as the open FLUX model to start with.
Nemotron 3.5 Bolt 74B-A9B
NVIDIA's local agentic model, packed into a 24GB download so it runs on a single high-end consumer GPU (32 GB+) — text, image and video input with a 256k context. A compressed version of the 120B-A12B Nemotron 3.5 Super VL, NVIDIA's open data-centre model released the same day.
Gemini 4 Argon
Google's flagship reasoning model and the first of the Gemini 4 generation — its most capable model for complex coding, agentic, math, and science work. Takes over the Pro tier from Gemini 3.1 Pro Preview.
Claude Haiku 5.5
Anthropic's fastest, cheapest tier — built for high-volume, latency-sensitive work such as classification, extraction, routing, and subagent tasks. Succeeds Claude Haiku 4.5.
EmbeddingGemma 2
Google's open multimodal embedding model — puts text, code, images, video and audio into one shared 768-dimension search space. 740M parameters under Apache 2.0, designed by Google to run on laptops and phones, with 100+ languages and an 8K context.
Mistral Large 4
Mistral AI's new flagship (v26.10, API ID mistral-large-4) — a general-purpose multimodal model that Mistral describes as open-weight, with weights not yet published. Succeeds Mistral Large 3. EU-jurisdiction provider.
Nano Banana 2.1
Google's current image generation and editing model in the Gemini API (model ID gemini-nano-banana-2.1). Improves visual quality, prompt adherence, multi-turn character consistency, and text rendering, and replaces gemini-3.1-flash-image.
Cohere Embed 5
Cohere's flagship embedding family (embed-v5.0-pro, embed-v5.0-fast). Multimodal (text, images, mixed text+image documents such as PDFs) with 128K context, 100+ languages, Matryoshka dimensions and int8/binary embeddings — useful for cost-sensitive RAG.
GPT-6.1 Sol
Near-Astra performance for complex coding, computer use, and professional work at a lower cost than GPT-6 Astra. The GPT-6 generation's frontier workhorse; it supersedes GPT-6 Sol (released 2026-09-22, one week earlier).
Claude Sonnet 5.5
Anthropic's mid-tier workhorse — the best combination of speed and intelligence in the current lineup, at the same $2/$10 price as Claude Sonnet 5, which it succeeds.
ElevenLabs Eleven v4
ElevenLabs' flagship expressive TTS (eleven_v4), with eleven_v4_turbo as its real-time variant (~100ms median inference). The benchmark for natural-sounding speech and voice cloning across 90+ languages.
Claude Opus 5.5
Anthropic's recommended starting point for most workloads — long-running agentic coding and knowledge work, at a lower price than Claude Opus 5. Succeeds Claude Opus 5 as the Opus-tier flagship.
GPT-6 Luna
OpenAI's most efficient model for focused, high-volume tasks — the cheap tier of the GPT-6 generation. OpenAI's named replacement for GPT-4o, GPT-4o mini, GPT-4.1, GPT-4.1 mini and GPT-5.4 nano.
Grok 4.7
SpaceXAI's current frontier model (formerly xAI), succeeding Grok 4.6 at the same price. Built for long-running agents and coding, with text and image input, a 500k context and a May 2026 knowledge cutoff.
MinerU
An open-source tool that wraps document-parsing models to turn PDFs, scans, Word and PowerPoint files into clean Markdown or JSON, keeping tables, formulas and reading order. Version 4.0 adds a local document library and agent-friendly reading with stable page and block citations.
DeepSeek V4.1 Flash
DeepSeek's current flagship and its only actively developed API model — a natively multimodal MoE that DeepSeek says beats V4 Pro, under MIT. The 2026-10-17 update (DeepSeek-V4.1-Flash-1017) sharpened coding and agentic work and now serves the deepseek-flash API ID on web, app and API.
Suno v6
Suno's flagship music generation model, developed with Warner Music Group, BMG and Believe. Produces full songs (vocals + instrumentation) from text, audio, images or video, with plain-language section edits, mashups and single-lyric changes — the most polished consumer music AI.
GPT Image 2.5
OpenAI's current image generation and editing line, shipped as two models: Sunburst (`gpt-image-2.5-sunburst`) for workflows where editing precision matters most, and Flare (`gpt-image-2.5-flare`) for fast, high-quality everyday generation. Supersedes GPT Image 2.
GPT-6 Astra
OpenAI's most capable model, built for the hardest end-to-end work — computer use, browsing, software engineering, and scientific analysis. The GPT-6 generation's flagship, sitting above the GPT-5.6 tiers.
Gemini 3.8 Flash
Google's most intelligent Flash model — engineered for long-horizon software engineering, autonomous agents, and complex enterprise workflows at Flash-tier speed and cost. The current head of the Gemini Flash line.
Claude Fable 5.1
Anthropic's most capable widely released model — the top tier above Opus 5, built for long-running agentic coding, knowledge work, and research. Succeeds Claude Fable 5.
Claude Mythos 5.1
Anthropic's restricted frontier model — Claude Fable 5.1's specifications and pricing with safeguards lifted in select domains. Invitation-only through Project Glasswing, succeeding Claude Mythos 5.
Gemini Omni Flash
Google's current video generation model in the Gemini API (model ID gemini-omni-1.1-flash), for fast generation, editing, keyframe interpolation, and extension. Google's named replacement for the Veo 3.1 line, which shut down on 2026-10-22.
GLM-5.3-Flash
Z.ai's efficient GLM-5.3 tier under plain MIT — a 320B sparse model with 18B active, natively multimodal on input, at roughly a tenth of frontier pricing. The most permissively licensed model at this capability level.
DeepSeek V4 Pro
DeepSeek's flagship general-purpose MoE model. Broadly competitive with the strongest proprietary models, released under MIT at open-weights cost.
MiniMax Music 3
MiniMax's open-weight song generator. Give it lyrics and a description and it composes, arranges and sings a complete song up to five minutes long, as 32 kHz stereo audio.
LTX-2.5
Lightricks' open-weights video model: makes video with synchronised sound from text or an image, with multi-shot scenes that stay consistent across cuts and native 4K HDR output. The top trending image-to-video model on Hugging Face in October 2026.
Muse Glimmer 30B
Meta's current open-weight model and the practical successor to the Llama line for local deployment. An agentic model built for tool use and multi-step task completion, and Meta's first language model under a standard OSI-approved licence.
Grok Imagine Image 2.0
SpaceXAI's image generation and editing model. Ranked #3 on the Arena text-to-image leaderboard as of 2026-09-04, and edits from up to five reference images in a single request.
Qwen3.8 27B
Alibaba's current open-weight workhorse — a 27B dense model under Apache 2.0, natively multimodal, and the model most commonly run on a single 24GB consumer GPU. Supersedes the Qwen3.6 generation.
MiniMax H3
MiniMax's open-weights video model: makes 4–15 second clips with native stereo sound from text, images, video and audio references, with stable dialogue in 11 languages. Strong on instruction following, text and brand rendering, and motion transfer from a reference video.
GPT Live Transcribe
OpenAI's realtime transcription model — the streaming half of the pair replacing GPT-4o Transcribe. Built for live captioning and voice agents, with tunable latency and keyword hints.
GPT Transcribe
OpenAI's current hosted transcription model and the named replacement for GPT-4o Transcribe. Priced per minute of audio rather than per token, for batch and file-based transcription.
Kimi K3
The largest open-weights model published to date — 2.8T total parameters with 104B active — and natively multimodal across text, images and video. Frontier-adjacent performance with downloadable weights.
Gemini 3.5 Flash-Lite
Google's fastest and most cost-effective Gemini tier — the high-throughput option for classification, extraction, and routing. Google's named replacement for Gemini 3.1 Flash-Lite.
OCR 4.1
Mistral's current OCR service, powering its Document AI stack. Adds native paragraph-level bounding boxes, structural block labels, and block-level confidence scores over OCR 3.
Kimi K2.7 Code
Moonshot's coding-agent model built on Kimi K2.6 — a 1T-parameter mixture-of-experts (32B active) tuned for long, real-world software tasks, using about 30% fewer thinking tokens than K2.6. Open weights, but sized for a multi-GPU server.
PaddleOCR 3.7
Baidu's open-source OCR toolkit. Version 3.7 ships PP-OCRv6, a classic text-detection and recognition pipeline in tiny, small and medium sizes, alongside PaddleOCR-VL-1.6, a 0.9B vision-language model for full document parsing into Markdown or JSON.
Chatterbox
Resemble AI's MIT-licensed open text-to-speech family. Multilingual V3 (500M) clones a voice in 23 languages; Turbo (350M) and Nano (110M) are English-only models for fast voice agents, with [laugh]-style tags. Every output is watermarked.
Luma Ray3.2
Luma AI's flagship video model (behind the Dream Machine app). Ray3.2 adds multi-keyframe control and professional HDR/EXR output on top of Ray3's reasoning-driven, 3D-aware generation.
Stable Audio 3.0
Stability AI's open-weight audio model. Generates music tracks, sound effects, and audio loops from text prompts — the open alternative to Suno/Udio. (Small/medium tiers open-weighted; large is API-only.)
Midjourney V8.1
Midjourney's flagship image generator. V8.1 is its fastest model yet (2K HD by default); strong artistic quality and a distinctive aesthetic, bound to its Discord and web product, no public API.
Mistral Medium 3.5
Mistral's frontier-class multimodal model for agentic and coding use cases. It sits between the largest flagship tier and the lighter Small line while remaining open-weight.
Qwen3.6 35B-A3B
Alibaba's mid-size open-weight mixture-of-experts from the Qwen3.6 generation — 35B total / 3B active parameters, Apache 2.0, natively multimodal (text + vision) with up to ~1M-token context. Still the current Qwen MoE at this size: Qwen3.8 has no MoE this small.
VoxCPM2
OpenBMB's open text-to-speech model — 30 languages, voice cloning from a short clip, and Voice Design that invents a new voice from a written description. Apache 2.0, so a business can ship it without a licence fee.
Gemma 4 31B
Google's flagship open-weight Gemma 4. Natively multimodal (text + vision); supersedes Gemma 2 with significantly improved capabilities and a larger size. Smaller Gemma 4 sizes for laptops and single GPUs are a separate entry.
Gemma 4 (E2B / E4B / 12B / 26B-A4B)
The smaller Gemma 4 sizes under Apache 2.0 — E2B and E4B for phones and laptops, 12B and the 26B-A4B mixture-of-experts for a consumer graphics card. All read text, images and video; E2B, E4B and 12B also take audio. Thinking mode is built in.
Mistral Small 4
Mistral's hybrid open model that unifies instruct, reasoning, and coding in a single efficient line. It is the current small generalist flagship in Mistral's open lineup.
Qwen3.5 9B / 4B
Alibaba's newest small open models — dense 9B (headline) and 4B under Apache 2.0, with image input, 262K context and thinking on by default. Qwen3.6 and Qwen3.8 ship nothing smaller than 27B, so this is still the current small Qwen.
Kling 3.0
Kuaishou's video generation model. 3.0 adds native audio, multimodal input, and up to 15-second clips; strong on human motion and physical realism.
Voxtral Mini 4B Realtime
Mistral's open live-transcription model. It writes text as people speak, under half a second behind, in 13 languages — accuracy close to offline transcribers, small enough to run on one graphics card.
MiniCPM-o 4.5
OpenBMB's 9B open model that sees, hears and talks at the same time — full-duplex voice and video conversation small enough to run on a Mac. Speech is English and Chinese only.
Qwen3-Coder-Next
Alibaba's open-weight coding-agent model for local development — an 80B mixture-of-experts with only 3B active per token, 256K context, and built to drive agent tools such as Claude Code, Cline and Qwen Code.
GLM-OCR
Z.ai's small open OCR model for complex documents — under 1B parameters, built on the GLM-V architecture, and strong on tables, formulas, seals and code-heavy pages. Its SDK adds layout analysis for full document parsing.
Qwen3-ASR
Alibaba's open speech-recognition family. The 1.7B model transcribes 30 languages and 22 Chinese dialects and detects the language itself; a separate ForcedAligner-0.6B adds word-level timestamps.
ACE-Step 1.5
An open music model that writes full songs, with vocals, from 10 seconds to 10 minutes long, in 50+ languages, and can make covers or re-generate parts of a track. MIT-licensed, and its makers say the output is cleared for commercial use.
Qwen3-TTS
Alibaba's open text-to-speech family. The headline 1.7B CustomVoice model speaks 10 languages in nine preset voices you can steer with plain-language instructions; sibling models clone a voice from 3 seconds of audio or design one from a description.
VibeVoice-ASR
Microsoft's open long-form transcription model. It takes up to an hour of audio in one pass and returns who spoke, when, and what they said — transcription, speaker labels and timestamps together, in 50+ languages.
Voyage 4
Voyage AI's flagship embedding model (voyage-4-large). Top of the RTEB retrieval leaderboard; the embedding service Anthropic recommends for Claude RAG workloads.
Qwen-Image-Edit-2511
Alibaba's open photo-editing model: describe the change in plain words and it edits the image, including combining several photos into one. The newest Apache 2.0 Qwen editing model — the later Qwen-Image-2.1 is research-only.
Runway Gen-4.5
Runway's flagship video generation model. Strong creative-tooling ecosystem (motion brush, camera control, style transfer); the production tool of choice for many video creators.
Devstral Small 2 (24B)
Mistral's small open-weight coding-agent model — 24B parameters under Apache 2.0, 256K context, light enough for a single RTX 4090 or a 32 GB Mac. Built to explore codebases, edit many files and drive coding agents.
Ministral 3 14B
The largest model in Mistral's Ministral 3 family. It is built for local deployment with strong text and vision performance on diverse hardware.
Z-Image-Turbo
A fast 6B open image model from Alibaba's Tongyi-MAI lab. A distilled version of Z-Image that makes a photoreal image in 8 steps and renders English and Chinese text accurately, on a 16 GB graphics card, under Apache 2.0.
FLUX.2 [pro]
Black Forest Labs' commercial flagship (FLUX.2 [pro]). The closed top tier when you need commercial-use rights and maximum quality; FLUX.2 [dev] is the open-weight sibling.
Qwen3-Omni
Alibaba's open model for live voice conversation. It listens, sees images and video, and answers in text or natural speech in real time — an open alternative to the voice modes of ChatGPT and Gemini, but it needs data-centre hardware.
Topaz Gigapixel
Industry-standard desktop application for photo upscaling and restoration. The default tool for archival work, photo restoration, and print-sized enlargements where fidelity to the original matters.
gpt-oss (20b / 120b)
OpenAI's open-weight reasoning models under Apache 2.0. The headline gpt-oss-20b runs on a 16 GB machine; gpt-oss-120b fits one 80 GB data-centre GPU. Both offer low / medium / high reasoning effort, tool use and agentic coding.
Codestral 25.08
Mistral's current code-completion model, released at the end of July 2025. It is tuned for low-latency fill-in-the-middle and high-frequency code generation tasks.
Wan 2.2
Alibaba's open video model family from the Wan team, under Apache 2.0. The 5B TI2V model makes 720p, 24fps video from text or an image on a single 24 GB graphics card; the larger A14B Mixture-of-Experts models handle text-to-video and image-to-video at higher quality.
Qwen3-Embedding (0.6B / 4B / 8B)
Alibaba's open text-embedding models in three sizes under Apache 2.0 — 0.6B is the headline for self-hosted search, with 4B and 8B for higher quality. 100+ languages including programming languages, 32K context, and adjustable vector size.
Qwen3-Reranker (0.6B / 4B / 8B)
Alibaba's open rerankers under Apache 2.0 — the second-pass model that re-scores search results so the best matches rise to the top. Three sizes (0.6B, 4B, 8B), 100+ languages, 32K context, and task instructions you can customise.
SeedVR2
ByteDance Seed's open restoration and upscaling model for video and still images. Restores and upscales in a single diffusion step, so it is far faster than multi-step diffusion upscalers while adding more real detail than older GAN upscalers.
NVIDIA Parakeet
NVIDIA's STT family. parakeet-tdt-0.6b-v2 tops the HF Open ASR leaderboard for English; parakeet-tdt-0.6b-v3 adds 25-language multilingual support. Very fast on NVIDIA hardware via NeMo.
Gemini 2.5 Flash
Google's speed-optimised tier. Cheap and fast multimodal, with a generous free tier on AI Studio for prototyping.
Llama 4 Maverick
Meta's flagship Llama 4 model — natively multimodal, larger MoE architecture than Scout. The Llama 4 frontier-tier entry.
Llama 4 Scout
Meta's small/efficient Llama 4 variant — natively multimodal MoE architecture. The practical Llama 4 entry point for self-hosted multimodal applications.
Gemini 2.5 Pro
Google's flagship multimodal model. Massive context window and competitive frontier-tier performance, with extended thinking on demand.
Udio Allegro v1.5
Suno's main competitor. Allegro v1.5 is the current model — faster generation with strong genre control and a focus on song-structure quality.
Pika 2.2
Pika's video generation model. 2.2 adds Pikaframes keyframe transitions and 1080p; differentiates with Scene Ingredients (drop in characters/objects across shots) for character consistency across clips.
Kokoro-82M
A tiny open text-to-speech model (82 million parameters) with 54 preset voices in 8 languages. Quality close to much larger models at a fraction of the cost — the most-downloaded TTS model on Hugging Face.
Phi-4
Microsoft's small model trained heavily on synthetic data. Punches above its 14B weight on reasoning and math; MIT-licensed and runs locally.
Llama 3.3 70B
Meta's latest 70B-parameter open-weight model. Reaches frontier-tier performance for English-centric tasks while remaining self-hostable.
Stable Diffusion 3.5
Stability AI's open image-gen family. Three sizes (Large, Large Turbo, Medium) — runs locally on consumer GPUs and supports a massive ecosystem of LoRAs and ControlNets.
Whisper large-v3 Turbo
OpenAI's distilled Whisper variant. ~8× faster than large-v3 with most of the accuracy retained — the practical default for high-throughput STT pipelines.
Llama 3.2 Vision
Meta's open-weight vision-language family. 11B and 90B variants — the practical open-weights vision model for self-hosted multimodal applications.
Llama 3.1 8B
Meta's small open-weight model. Runs on consumer hardware (16GB GPU or modern laptop) and remains a strong default for local-first AI.
Docling
An open-source toolkit from IBM Research that wraps layout, table and OCR models to turn PDFs, Word, PowerPoint, Excel, HTML, images and more into Markdown or lossless JSON for search and RAG. Now hosted by the LF AI & Data Foundation.
BGE Reranker v2
BAAI's open reranker. Apache 2.0 weights, multiple sizes, and the open-weights default for self-hosted RAG pipelines that need a second-pass.
BGE-M3
BAAI's multi-functional embedding model. Supports dense, sparse, and multi-vector retrieval in one model — the strongest open-weights embedding option.
OpenAI text-embedding-3-large
OpenAI's largest embedding model. 3072 dimensions, multilingual, and the default high-quality option for RAG and semantic search.
OpenAI text-embedding-3-small
OpenAI's small embedding tier. 1536 dimensions; the cheap default for most RAG and semantic-search workloads where quality is sufficient.
BiRefNet
A high-quality open background-removal and cut-out model from Peng Zheng and co-authors (lead affiliation Nankai University). Separates the subject from the background with clean edges on hair and fine detail; a high-resolution variant works at 2048×2048.
Magnific AI
The premium 'creative upscaler'. Invents detail at high zoom factors using diffusion priors — the right choice when you want an upscale that adds fidelity rather than just enlarging pixels.
Whisper large-v3
High-accuracy multilingual speech-to-text. Best-in-class for non-English audio; the de-facto open baseline.
Distil-Whisper
Hugging Face's distillation of Whisper (distil-large-v3.5). Faster than Whisper-large-v3-Turbo on long-form audio at small accuracy cost; English-only — pick when you don't need multilingual.
Piper
A fast, local TTS designed for Raspberry Pi-class hardware. Powers most self-hosted voice assistants where Whisper handles input and Piper handles output. Now maintained by the Open Home Foundation under GPL-3.0.
Stable Diffusion x4 Upscaler
Stability AI's diffusion-based 4× upscaler. Trades speed for quality — invents plausible high-frequency detail rather than just sharpening, which suits AI-generated images especially well.
Real-ESRGAN
The de-facto open-weights image upscaler. Battle-tested across years of community use; runs on CPU or GPU, integrates with virtually every local image pipeline.
Vosk
An offline, lightweight speech recognition toolkit. Runs on phones, Raspberry Pi, and embedded devices — the right choice when Whisper is too heavy.