all-MiniLM-L6-v2
Sentence-Transformers · Sentence-Transformers all-* · all-MiniLM-L6-v2, all-MiniLM-L12-v2, all-mpnet-base-v2 | Open source | 23 M | 256 | embedding | Open weights | Apache-2.0 | 2 Mar 2022 | Fast 384-dimension sentence embeddings on CPU, used in the Field Guide's runnable examples
The date is the Hugging Face repository date. The card says the model was trained during the 2021 Hugging Face community week but gives no day. Inputs beyond 256 word pieces are truncated. |
Apple Foundation Models (AFM 3)
Apple · Apple Foundation Models · AFM on-device (about 3B), AFM server (PT-MoE), AFM 3 Core (3B), AFM 3 Core Advanced (20B sparse), AFM 3 Cloud, AFM 3 Cloud Pro | Apple | — | — | multimodal, text, vision | Closed | Proprietary (Apple Intelligence, Foundation Models framework) | 17 Jul 2025 | On-device generation inside iOS and macOS apps through the Swift Foundation Models framework
Date is the 2025 tech report. Apple discloses about 3 B parameters for the on-device model and 20 B with 1 to 4 B active for AFM 3 Core Advanced (June 2026). The cloud models and context window are undisclosed. The third generation was built with Google. No weights are published. |
BGE-M3
BAAI · BGE · bge-m3, bge-large-en-v1.5, bge-base-en-v1.5, bge-small-en-v1.5 | Open source | 568 M | 8 k | embedding | Open weights | MIT | 30 Jan 2024 | One open multilingual embedding model that returns dense, sparse and multi-vector outputs
Parameter count derived from the XLM-RoBERTa-large backbone and checkpoint size, since the card does not state it. 1024 dimensions, more than 100 languages. Multilingual-E5-large (Microsoft, MIT, 512 tokens) is the usual alternative. |
BLOOM
BigScience · BLOOM · 560m, 1b1, 1b7, 3b, 7b1, 176B | Hugging Face | 176 B | 2 k | text, code | Open, restricted | BigScience RAIL License v1.0 | 11 Jul 2022 | Historical reference for the first large open multilingual model, trained in public on Jean Zay
46 natural and 13 programming languages. The RAIL licence carries use restrictions, so it is open weights rather than open source. BLOOMZ is the instruction-tuned variant. |
Claude Fable 5.2
Anthropic · Claude 5.2 · Claude Fable 5, Claude Fable 5.1, Claude Fable 5.2, Claude Mythos 5.2 | Anthropic | — | 1 M | multimodal, text, vision, code | Closed | Proprietary (API terms) | 9 Nov 2026 | Demanding reasoning and long-horizon agent runs once Opus 5.5 at high effort falls short
Parameter count undisclosed. The newest Claude model at the time of writing. Thinking cannot be switched off. Requires 30-day data retention, so it is unavailable under zero-retention agreements. Claude Mythos 5.2 is the same model for verified organisations. |
Claude Haiku 5.5
Anthropic · Claude 5.5 · Claude Haiku 4.5, Claude Haiku 5.5 | Anthropic | — | 1 M | multimodal, text, vision | Closed | Proprietary (API terms) | 7 Oct 2026 | High-volume classification, routing and extraction where latency and cost dominate
Parameter count undisclosed. 128 k max output. Tiered pricing: cheaper for prompts up to 100 k tokens. |
Claude Opus 5.5
Anthropic · Claude 5.5 · Claude Haiku 5.5, Claude Sonnet 5.5, Claude Opus 5.5 | Anthropic | — | 1 M | multimodal, text, vision, code | Closed | Proprietary (API terms) | 22 Sept 2026 | Agentic coding and long document work where Anthropic's docs recommend starting
Parameter count undisclosed. 128 k max output. Adaptive thinking is always on and depth is set with an effort level. Anthropic's model overview names it the default for most workloads. |
Claude Sonnet 4.5
Anthropic · Claude 4.5 · Claude Haiku 4.5, Claude Sonnet 4.5, Claude Opus 4.5 | Anthropic | — | 200 k | multimodal, text, vision, code | Closed | Proprietary (API terms) | 29 Sept 2025 | Reading code written against the 2025 Claude API, with budget_tokens extended thinking
Parameter count undisclosed. Deprecated on 30 September 2026 and retires on 30 November 2026 on the Claude API, with Claude Sonnet 5.5 as the recommended replacement. Claude Sonnet 4 (May 2025) was retired on 15 June 2026. 64 k max output. |
Claude Sonnet 5.5
Anthropic · Claude 5.5 · Claude Haiku 5.5, Claude Sonnet 5.5, Claude Opus 5.5 | Anthropic | — | 1 M | multimodal, text, vision, code | Closed | Proprietary (API terms) | 28 Sept 2026 | Everyday coding, agents and enterprise work at half the Opus price
Parameter count undisclosed. 128 k max output. Priced at 2 and 10 USD per million input and output tokens at the time of writing. |
CLIP ViT-L/14
OpenAI · CLIP · ViT-B/32, ViT-B/16, ViT-L/14, ViT-L/14@336px, RN50, RN101 | OpenAI | 400 M | 77 | multimodal, vision, embedding | Open weights | MIT | 1 Jan 2022 | Zero-shot image classification and joint image-text embeddings
Context is the 77-token limit of the text encoder. OpenAI's model card dates CLIP to January 2021 and the ViT-L/14 weights to January 2022 without a day, so the release date here is the first of that month. Parameter count is the Hugging Face estimate of about 0.4 B. |
Codestral 22B
Mistral AI · Codestral · Codestral 22B v0.1, Codestral 25.01 (API), Codestral 25.08 (API) | Mistral | 22 B | 32 k | code | Open, restricted | Mistral AI Non-Production License | 29 May 2024 | Fill-in-the-middle code completion in an editor, for research or non-production use
Trained on more than 80 programming languages. Later Codestral versions (25.01 with 256 k context, 25.08) are API-only with undisclosed sizes. |
DeepSeek-R1
DeepSeek · DeepSeek-R1 · R1, R1-Zero, R1-0528, Distill-Qwen-1.5B, Distill-Qwen-7B, Distill-Llama-8B, Distill-Qwen-14B, Distill-Qwen-32B, Distill-Llama-70B | Open source | 671 B | 128 k | text, code | Open weights | MIT | 20 Jan 2025 | Open reasoning traces and the distilled sizes that run on one GPU
MoE: 671 B total, 37 B active. The R1-0528 update (28 May 2025) is also MIT. Distilled Llama variants inherit the Llama licence. |
DeepSeek-V3
DeepSeek · DeepSeek-V3 · DeepSeek-V3-Base, DeepSeek-V3, DeepSeek-V3-0324 | Open source | 671 B | 128 k | text, code | Open, restricted | DeepSeek Model License (code MIT) | 26 Dec 2024 | The MoE and multi-token-prediction design that set the 2025 open-weights agenda
MoE: 671 B total, 37 B active. The checkpoint is 685 B on disk because it includes the multi-token prediction module. The original weights are under the DeepSeek Model License, which permits commercial use. From V3-0324 (March 2025) onward the weights are MIT. |
DeepSeek-V4.1-Flash
DeepSeek · DeepSeek-V4 · V4-Flash (284B-A13B), V4-Pro (1.6T-A49B), V4.1-Flash, V4.1-Flash-1017 | Open source | 552 B | 1 M | multimodal, text, vision, code | Open weights | MIT | 10 Sept 2026 | DeepSeek's newest open model at the time of writing, multimodal with 1 M context
552 B backbone plus a 196 B sparsely accessed memory, 763 B on disk, with 8 to 16 B active depending on mode. Causal encoder-decoder trained from scratch on 45 T multimodal tokens. DeepSeek-V4-Pro (1.6 T total, 49 B active, MIT) was previewed in April 2026. |
Devstral Small
Mistral AI · Devstral · Devstral Small 1.0 (2505), Devstral Small 1.1 (2507), Devstral Small 2 (2512), Devstral 2 (2512) | Mistral | 24 B | 128 k | code, text | Open weights | Apache-2.0 | 21 May 2025 | Local agentic coding with tools such as OpenHands on a single 24 GB GPU
Fine-tuned from Mistral Small 3.1 with the vision encoder removed, built with All Hands AI. Magistral Small (June 2025, 24 B, Apache-2.0) is the reasoning sibling. |
DINOv2
Meta · DINO · ViT-S/14 (21M), ViT-B/14 (86M), ViT-L/14 (300M), ViT-g/14 (1.1B) | Meta | 1.1 B | — | vision, embedding | Open weights | Apache-2.0 | 17 Apr 2023 | Self-supervised image features for linear probes, retrieval and dense prediction without labels
Parameter count is the ViT-g/14 teacher. The smaller sizes are distilled from it. DINOv3 (August 2025, up to 6.7 B) is stronger but under the custom DINOv3 License. |
Falcon-H1-7B-Instruct
Technology Innovation Institute (TII) · Falcon-H1 · 0.5B, 1.5B, 1.5B-Deep, 3B, 7B, 34B | Open source | 7.6 B | 262 k | text, code | Open, restricted | Falcon-LLM License | 20 May 2025 | A hybrid Transformer and Mamba architecture with long context at small sizes
The Falcon-LLM License is Apache-2.0 based with an acceptable-use policy, hence open-restricted. Falcon 3 (December 2024, 1B to 10B, 32 k context) is the previous pure-Transformer family. |
Florence-2-large
Microsoft · Florence-2 · Florence-2-base (0.23B), Florence-2-large (0.77B) | Microsoft | 770 M | — | vision, multimodal | Open weights | MIT | 15 Jun 2024 | One small model for captioning, detection, segmentation and OCR via text prompts
Release date is the first Hugging Face commit. The card gives only the November 2023 paper date. Text context length undisclosed. Fine-tuned -ft variants exist for both sizes. |
FLUX.1 [dev]
Black Forest Labs · FLUX.1 · FLUX.1 [schnell] (Apache-2.0), FLUX.1 [dev], FLUX.1 [pro] (API only) | Open source | 12 B | — | vision | Open, restricted | FLUX.1 [dev] Non-Commercial License | 31 Jul 2024 | The most widely fine-tuned open image model of 2024 and 2025
Rectified flow transformer. Date is the Hugging Face repository date under the maker's organisation. FLUX.1 [schnell] has the same 12 B size under Apache-2.0. FLUX.2 [dev] (25 November 2025, 32 B) succeeded it under the FLUX Non-Commercial License. |
Gemini 2.5 Flash
Google · Gemini 2.5 · Gemini 2.5 Flash-Lite, Gemini 2.5 Flash, Gemini 2.5 Flash Image | Google | — | 1.0 M | multimodal, text, vision, audio, video | Closed | Proprietary (API terms) | 17 Jun 2025 | The low-cost multimodal tier that many 2025 RAG and agent examples target
Parameter count undisclosed. Deprecated at the time of writing, with API access ending on 1 December 2026. Knowledge cutoff January 2025. |
Gemini 2.5 Pro
Google · Gemini 2.5 · Gemini 2.5 Flash-Lite, Gemini 2.5 Flash, Gemini 2.5 Pro | Google | — | 1.0 M | multimodal, text, vision, audio, video, code | Closed | Proprietary (API terms) | 17 Jun 2025 | Reading 2025 benchmarks and tutorials built on the Gemini 2.5 generation
Parameter count undisclosed. Stable release date, preview builds appeared in March 2025. Deprecated at the time of writing, with API access ending on 1 December 2026 and requests rerouted to a Gemini 3.8 Flash model. 65 k max output. |
Gemini 4 Carbon
Google · Gemini 4 · Gemini 4 Hydrogen, Gemini 4 Argon, Gemini 4 Carbon | Google | — | 1.0 M | multimodal, text, vision, audio, video, code | Closed | Proprietary (API terms) | 10 Nov 2026 | Google's strongest reasoning model for multimodal and long-context work
Parameter count undisclosed. The context figure is a unified input plus output window. Gemini 4 Argon (13 October 2026) is the mid tier and Gemini 4 Hydrogen (12 November 2026) the small one. Introductory pricing ends 30 November 2026. |
Gemini Embedding (gemini-embedding-001)
Google · Gemini Embedding · gemini-embedding-001, gemini-embedding-2 | Google | — | 2 k | embedding | Closed | Proprietary (API terms) | 14 Jul 2025 | Hosted text embeddings with Matryoshka output from 128 to 3072 dimensions
Parameter count undisclosed. Text only. Scheduled shutdown on 14 May 2028 with the multimodal gemini-embedding-2 (8 k input) as replacement. The open EmbeddingGemma (308 M) is the on-device alternative. |
Gemma 3 27B
Google DeepMind · Gemma 3 · 270M, 1B, 4B, 12B, 27B | Google | 27 B | 128 k | multimodal, text, vision | Open, restricted | Gemma Terms of Use | 12 Mar 2025 | A single-GPU open model with image input and 140 languages under the Gemma terms
The 1 B and 270 M sizes are text only with 32 k context. Knowledge cutoff August 2024. Superseded by Gemma 4 in April 2026 but still widely deployed. |
Gemma 4 31B
Google DeepMind · Gemma 4 · E2B, E4B, 12B, 26B A4B, 31B | Google | 30.7 B | 256 k | multimodal, text, vision, video | Open weights | Apache-2.0 | 2 Apr 2026 | The first Gemma under Apache-2.0, with configurable thinking and 256 k context
Dense 31 B flagship. The 26B A4B sibling is an MoE with 25.2 B total and 3.8 B active. E2B and E4B are 128 k context with native audio input. Knowledge cutoff January 2025. |
GLM-4.6
Z.ai (Zhipu AI) · GLM-4.5 · GLM-4.5-Air (106B-A12B), GLM-4.5 (355B-A32B), GLM-4.6 | Open source | 355 B | 200 k | text, code | Open weights | MIT | 29 Sept 2025 | An MIT-licensed MoE aimed at agentic coding, with a 106 B Air variant for smaller hosts
MoE: 355 B total, 32 B active. Date is the Hugging Face repository date, since the maker's blog did not render for automated checks. GLM-4.5 (July 2025) had 128 k context. |
GPT-4o
OpenAI · GPT-4 · gpt-4o, GPT-4o mini, gpt-4o-audio-preview, gpt-4o-realtime-preview, gpt-4o-transcribe | OpenAI | — | 128 k | multimodal, text, vision, audio | Closed | Proprietary (API terms) | 13 May 2024 | Understanding the generation of chat and vision models that most 2024 and 2025 tutorials assume
Parameter count undisclosed. 16 k max output. Marked deprecated on the API model page at the time of writing, with the May 2024 snapshot shut down. Knowledge cutoff October 2023. |
GPT-6 Astra
OpenAI · GPT-6 · GPT-6 Astra, GPT-6 Astra Pro, GPT-6 Sol, GPT-6 Luna, GPT-6.1 Sol, GPT-6.1 Terra | OpenAI | — | 1.1 M | multimodal, text, vision, code | Closed | Proprietary (API terms) | 3 Sept 2026 | The hardest reasoning and agentic work where cost matters less than quality
Parameter count undisclosed. Listed first under flagship models on the OpenAI API docs. 922 k max input, 128 k max output. Reasoning is built in, so there is no separate o-series successor. Knowledge cutoff April 2026. |
GPT-6.1 Terra
OpenAI · GPT-6 · GPT-6.1 Sol, GPT-6.1 Terra | OpenAI | — | 1.1 M | multimodal, text, vision, code | Closed | Proprietary (API terms) | 26 Oct 2026 | Everyday API work with vision and computer use at a lower price than the Astra tier
Parameter count undisclosed. The newest OpenAI text model at the time of writing. 128 k max output. Knowledge cutoff June 2026. |
gpt-oss-120b
OpenAI · gpt-oss · gpt-oss-20b, gpt-oss-120b | OpenAI | 117 B | 131 k | text, code | Open weights | Apache-2.0 | 5 Aug 2025 | Self-hosted reasoning on one 80 GB GPU with an OpenAI-style harmony chat format
MoE: 117 B total, 5.1 B active, 128 experts with 4 active per token. Shipped in MXFP4 so it fits one 80 GB GPU. The 20 B version (3.6 B active) fits 16 GB. OpenAI's first open-weight language models since GPT-2. |
Kimi K2 Thinking
Moonshot AI · Kimi K2 · Kimi-K2-Base, Kimi-K2-Instruct, Kimi-K2-Instruct-0905, Kimi-K2-Thinking | Open source | 1000 B | 256 k | text, code | Open, restricted | Modified MIT License | 4 Nov 2025 | A trillion-parameter open agentic model shipped in native INT4
MoE: 1 T total, 32 B active, 384 experts. Date is the Hugging Face repository date. The modified MIT licence adds an attribution clause for very large deployments. Kimi K2 Instruct (July 2025) has 128 k context. |
Llama 3.1 8B
Meta · Llama 3.1 · 8B, 70B, 405B | Meta | 8 B | 128 k | text, code | Open, restricted | Llama 3.1 Community License | 23 Jul 2024 | Local experiments and LoRA fine-tuning on one consumer GPU or a Mac
Base and Instruct variants. Pretraining cutoff December 2023. The licence restricts use by services with more than 700 million monthly users and requires attribution. |
Llama 3.3 70B Instruct
Meta · Llama 3.3 · 70B Instruct | Meta | 70 B | 128 k | text, code | Open, restricted | Llama 3.3 Community License Agreement | 6 Dec 2024 | A dense 70 B model that matched Llama 3.1 405B on Meta's reported benchmarks at a fraction of the cost
Instruct-only release, no base model. Widely served by hosting providers and runs quantised on two 24 GB GPUs. |
Llama 4 Maverick
Meta · Llama 4 · Scout (17Bx16E), Maverick (17Bx128E) | Meta | 400 B | 1 M | multimodal, text, vision, code | Open, restricted | Llama 4 Community License Agreement | 5 Apr 2025 | The larger Llama 4 for multilingual chat and image understanding on a multi-GPU host
MoE: 400 B total, 17 B active, 128 routed experts plus one shared expert. The licence adds EU restrictions on multimodal use. |
Llama 4 Scout
Meta · Llama 4 · Scout (17Bx16E), Maverick (17Bx128E) | Meta | 109 B | 10 M | multimodal, text, vision, code | Open, restricted | Llama 4 Community License Agreement | 5 Apr 2025 | Long-document and image work on a single H100 at int4
MoE: 109 B total, 17 B active, 16 experts. The 10 M context is Meta's claim and most hosts serve far less. Still the newest open Llama at the time of writing, since Meta's 2026 Muse Spark models are closed. Knowledge cutoff August 2024. |
Llama-3.3-Nemotron-Super-49B-v1
NVIDIA · Llama Nemotron · Nano 8B, Super 49B, Ultra 253B | NVIDIA | 49 B | 131 k | text, code | Open, restricted | NVIDIA Open Model License | 18 Mar 2025 | A pruned Llama 3.3 70B derivative that fits one H100 with reasoning switchable by system prompt
Built from Meta's Llama 3.3 70B Instruct by neural architecture search, so the Llama 3.3 Community License also applies. The repository name uses an underscore in 3_3. |
Mistral 7B
Mistral AI · Mistral 7B · Mistral 7B v0.1, Mistral 7B Instruct | Mistral | 7.3 B | 33 k | text | Open weights | Apache-2.0 | 27 Sept 2023 | The classic small open model behind countless fine-tunes and llama.cpp quantisations
Context is the max_position_embeddings in the released config. Attention uses a 4096-token sliding window. Grouped-query attention. |
Mistral Large 4
Mistral AI · Mistral Large · Mistral Large 3 (2512), Mistral Large 4 (2610) | Mistral | 1050 B | 1 M | multimodal, text, vision, code | Open, restricted | Mistral Ecosystem Open-Weight License | 6 Oct 2026 | Mistral's flagship for agentic and long-context work, available as open weights for self-hosting
MoE: 1.05 T total, 52 B active, plus a 1.6 B vision encoder. Date is the API public preview. The weights followed in November 2026. The licence is a modified MIT that requires a separate commercial licence for model-as-a-service operators. Mistral Large 3 (675 B, Apache-2.0, December 2025) is deprecated. |
Mistral Small 3.2
Mistral AI · Mistral Small · Mistral Small 3.1 (2503), Mistral Small 3.2 (2506) | Mistral | 24 B | 128 k | multimodal, text, vision | Open weights | Apache-2.0 | 20 Jun 2025 | A dense 24 B multimodal model that fits one RTX 4090 or a 32 GB Mac
An update of Mistral Small 3.1 (17 March 2025) with better instruction following and function calling. Mistral lists the 3.x line as deprecated on its API since Mistral Small 4. |
Mistral Small 4
Mistral AI · Mistral Small · Mistral Small 4 (2603) | Mistral | 119 B | 256 k | multimodal, text, vision, code | Open weights | Apache-2.0 | 16 Mar 2026 | A cheap-to-serve MoE that merges instruct, reasoning and coding in one open model
MoE: 119 B total, about 6.5 B active (Mistral's docs say 6.5 B, the announcement 6 B plus embeddings). Despite the Small name it needs a multi-GPU host at full precision. |
Mixtral 8x7B
Mistral AI · Mixtral · Mixtral 8x7B, Mixtral 8x7B Instruct, Mixtral 8x22B | Mistral | 46.7 B | 32 k | text | Open weights | Apache-2.0 | 11 Dec 2023 | Learning how sparse mixture-of-experts routing works on an open model
MoE: 46.7 B total, 12.9 B active, 2 of 8 experts per token per layer. |
NVIDIA Nemotron 3 Nano 30B-A3B
NVIDIA · Nemotron 3 · Nano 30B-A3B, Super 120B-A12B, Ultra 550B-A55B | NVIDIA | 30 B | 1 M | text, code | Open, restricted | NVIDIA Nemotron Open Model License | 15 Dec 2025 | A fast hybrid Mamba-2 and MoE model for agents and long context on a single GPU
MoE: 30 B total, 3.5 B active, 128 routed experts plus one shared. The Hugging Face default config is 256 k context with 1 M supported. Nemotron 3.5 Super VL (15 October 2026, 120 B total and 12 B active, OpenMDW 1.1 licence) is the newest family member. |
Olmo 3 32B
Ai2 (Allen Institute for AI) · Olmo 3 · 7B, 32B, 7B-Think, 32B-Think, 7B-Instruct, 3.1-32B-Think, 3.1-32B-Instruct | Open source | 32 B | 66 k | text, code | Open weights | Apache-2.0 | 20 Nov 2025 | Fully open science, with training data, code and every checkpoint published alongside the weights
Open source in the full sense, not only open weights. The Think variants expose complete reasoning traces. OLMo 2 (November 2024, 4 k context) is the previous generation. |
OpenAI o3
OpenAI · o-series · o3, o3-pro, o3-mini, o4-mini | OpenAI | — | 200 k | multimodal, text, vision | Closed | Proprietary (API terms) | 16 Apr 2025 | Reading 2025 papers and benchmarks that report o-series reasoning results
Parameter count undisclosed. The last o-series base model. Deprecated on the API docs, succeeded by GPT-5, with o4-mini shut down. 100 k max output. |
OpenELM
Apple · OpenELM · 270M, 450M, 1.1B, 3B | Apple | 3 B | 2 k | text | Open, restricted | Apple Sample Code License | 23 Apr 2024 | Studying a fully reproducible small-model training recipe, with CoreNet training code
Parameter count is the largest size. Context is max_context_length from the released config. Date is the first Hugging Face upload. The paper appeared on 22 April 2024. Instruct variants exist for every size. |
Parakeet TDT 0.6B v3
NVIDIA · Parakeet · parakeet-tdt-0.6b-v2 (English), parakeet-tdt-0.6b-v3 (25 languages) | NVIDIA | 600 M | — | audio | Open weights | CC-BY-4.0 | 14 Aug 2025 | Fast open speech recognition with punctuation across 25 European languages
FastConformer encoder with a token-and-duration transducer decoder, run through NVIDIA NeMo. Audio model, so no token context. v2 (1 May 2025) is English only. |
Phi-4
Microsoft · Phi-4 · Phi-4-mini-instruct (3.8B), Phi-4-multimodal-instruct (5.6B), Phi-4 (14B), Phi-4-reasoning (14B), Phi-4-reasoning-vision-15B | Microsoft | 14 B | 16 k | text, code | Open weights | MIT | 12 Dec 2024 | A dense 14 B model trained heavily on synthetic data, strong on maths and reasoning for its size
Release date is the Azure AI Foundry date on the model card. The public Hugging Face upload followed on 8 January 2025. Data cutoff June 2024. |
Phi-4-mini-instruct
Microsoft · Phi-4 · Phi-4-mini-instruct, Phi-4-mini-reasoning, Phi-4-mini-flash-reasoning | Microsoft | 3.8 B | 128 k | text, code | Open weights | MIT | 19 Feb 2025 | On-device and edge chat with function calling, where 128 k context matters more than peak quality
The model card gives February 2025 only. The day is the first Hugging Face commit. 200 k vocabulary with grouped-query attention. The multimodal sibling adds image and audio input at 5.6 B. |
Phi-4-reasoning-vision-15B
Microsoft · Phi-4 · Phi-4-reasoning (14B), Phi-4-reasoning-plus (14B), Phi-4-reasoning-vision-15B | Microsoft | 15 B | 16 k | multimodal, text, vision | Open weights | MIT | 4 Mar 2026 | The newest Phi at the time of writing, pairing Phi-4-reasoning with a SigLIP-2 vision encoder
No Phi-5 existed at the time of writing. Mid-fusion design with up to 3,600 visual tokens per image. Phi-4-reasoning (30 April 2025) is the text-only base with 32 k context. |
Pixtral 12B
Mistral AI · Pixtral · Pixtral 12B (2409), Pixtral Large (2411) | Mistral | 12.4 B | 128 k | multimodal, text, vision | Open weights | Apache-2.0 | 17 Sept 2024 | An early open vision-language model with interleaved images and variable resolution
12 B decoder plus a 0.4 B vision encoder trained from scratch. Marked deprecated by Mistral. Pixtral Large (124 B) is under the Mistral Research License. |
Qwen2.5-7B-Instruct
Alibaba Cloud (Qwen team) · Qwen2.5 · 0.5B, 1.5B, 3B, 7B, 14B, 32B, 72B | Open source | 7.6 B | 131 k | text, code | Open weights | Apache-2.0 | 19 Sept 2024 | The default small open model for multilingual chat, structured output and fine-tuning in 2024 and 2025
The released config is 32 k. The full 131 k needs YaRN. The 3B size is under the Qwen Research License and the 72B under the Qwen License, so only the other sizes are Apache-2.0. Qwen2.5-Coder and Qwen2.5-Math are sibling families. |
Qwen3-235B-A22B
Alibaba Cloud (Qwen team) · Qwen3 · 0.6B, 1.7B, 4B, 8B, 14B, 32B, 30B-A3B, 235B-A22B | Open source | 235 B | 33 k | text, code | Open weights | Apache-2.0 | 29 Apr 2025 | A large open MoE with a thinking switch, served by most hosting providers
MoE: 235 B total, 22 B active, 8 of 128 experts per token. Native context 32 k, 128 k with YaRN. The July 2025 Instruct-2507 and Thinking-2507 updates raise it to 256 k. The dense Qwen3-8B (8.2 B) is the common single-GPU choice. |
Qwen3.8-27B
Alibaba Cloud (Qwen team) · Qwen3.8 · Qwen3.8-27B, Qwen3.8-Flash-Next, Qwen3.8-2.4T-A95B | Open source | 27.8 B | 262 k | multimodal, text, vision, video, code | Open weights | Apache-2.0 | 14 Aug 2026 | The newest Apache-2.0 Qwen at the time of writing, dense and natively vision-language with 256 k context
Context extends to about 1 M tokens. The 2.4 T flagship (95 B active, 12 August 2026) is under the custom Qwen3.8-Max License with monthly-user and revenue clauses, so it is open-restricted. Qwen3.5 (February 2026) and Qwen3.6 (April 2026) sit between Qwen3 and this release. |
SAM 2 (Segment Anything Model 2)
Meta · Segment Anything · Tiny (38.9M), Small (46M), Base+ (80.8M), Large (224.4M) | Meta | 224 M | — | vision, video | Open weights | Apache-2.0 | 29 Jul 2024 | Promptable image and video segmentation with open training code
Parameter count is the Large checkpoint. Vision model, so no token context. SAM 2.1 checkpoints (September 2024) are a drop-in improvement. SAM 3 (November 2025, 848 M) adds text prompts but ships under the more restrictive SAM License. |
SmolLM2-1.7B-Instruct
Hugging Face · SmolLM2 · 135M, 360M, 1.7B | Hugging Face | 1.7 B | 8 k | text | Open weights | Apache-2.0 | 31 Oct 2024 | Runnable examples on a laptop CPU, which is why the Field Guide caches the 135M version
Llama-style architecture trained on 11 T tokens, English only. Context from the released config, date from the first Hugging Face commit. Base and Instruct variants at every size. |
SmolLM3-3B
Hugging Face · SmolLM3 · SmolLM3-3B-Base, SmolLM3-3B | Hugging Face | 3 B | 66 k | text, code | Open weights | Apache-2.0 | 8 Jul 2025 | A fully open 3 B model with published training recipe, six languages and a think or no-think switch
Trained at 64 k context and extends to 128 k with YaRN. Hugging Face published the full data mix and training configuration. |
SmolVLM-Instruct
Hugging Face · SmolVLM · 256M, 500M, 2.2B | Hugging Face | 2.2 B | 16 k | multimodal, vision, text | Open weights | Apache-2.0 | 26 Nov 2024 | Vision-language inference in under 6 GB of GPU memory, or on CPU for the 256M size
SmolLM2 backbone with a SigLIP encoder. The 256M and 500M sizes followed on 23 January 2025. SmolVLM2 is a separate later family with video input. |
Stable Diffusion 3.5 Large
Stability AI · Stable Diffusion 3.5 · Medium (2.5B), Large (8.1B), Large Turbo (8.1B) | Open source | 8.1 B | — | vision | Open, restricted | Stability AI Community License | 22 Oct 2024 | Open text-to-image with a licence that is free below 1 M USD annual revenue
Multimodal diffusion transformer. Image model, so no token context. The Hugging Face repository is gated behind licence acceptance. |
text-embedding-3-large
OpenAI · text-embedding-3 · text-embedding-3-small, text-embedding-3-large | OpenAI | — | 8 k | embedding | Closed | Proprietary (API terms) | 25 Jan 2024 | Hosted embeddings for retrieval when you already use the OpenAI API
Parameter count undisclosed. Up to 3072 output dimensions, and the dimensions parameter lets you shorten vectors. Priced per input token. |
Whisper large-v3
OpenAI · Whisper · tiny, base, small, medium, large-v2, large-v3, large-v3-turbo | OpenAI | 1.6 B | — | audio | Open weights | MIT | 6 Nov 2023 | Open speech recognition and translation across 99 languages
Processes audio in 30 s windows, so there is no token context figure. The GitHub repository states MIT while the Hugging Face card metadata says Apache-2.0. large-v3-turbo (809 M) is a faster distilled decoder. |