Model repository

Up to 1M token context

Strong reasoning, coding, and long-document work; Claude Code and the Agent SDK make it a first pick for agentic workflows. The Claude 5 generation landed through mid-2026: Fable 5 (Jun 9) is the most capable widely released tier, Sonnet 5 (Jul 1) brings near-Opus quality at Sonnet cost, and Opus 5 (Jul 24) targets complex agentic coding. Fable, Opus and Sonnet all take 1M tokens of context; Haiku 4.5 is the fast, cheap tier at 200K. A fourth model, Mythos 5, shares Fable 5’s specs but is invitation-only for defensive cybersecurity work — you cannot sign up for it. From Aug 2026 Claude also marks its own output: an invisible watermark is woven into generated text by fixing the choice between equally good next words with a hidden key, and generated images carry signed C2PA provenance metadata. Two caveats matter more than the feature. Code is largely unmarked — where the output has to be exact there are no free word choices to hide a signal in — and a positive detection only means Claude was likely involved somewhere, not that a human did not write it. Anthropic says a detection API is planned. Claude Fable 5.1 (Sep 2026, model ID `claude-fable-5-1`) succeeds Fable 5 as the top tier, on Anthropic’s own platforms plus AWS, Google Cloud and Microsoft Azure; same 1M context and per-token rates as Fable 5, but prompt-cache reads got a steep price cut that meaningfully lowers cost for agentic work leaning on cache reuse. Two safeguard changes loosen restrictions specifically for defensive use: cybersecurity safeguards now permit identifying vulnerabilities for patching rather than exploiting them, and biology safeguards fire far less often on ordinary elementary-biology and clinical questions — both figures are Anthropic’s own, without independent measurement. A new anti-distillation measure blocks manually editing Claude’s prior context on new API accounts, closing a technique used to extract a smaller model’s training signal from Claude’s outputs. Claude Mythos 5.1 mirrors Fable 5.1 with those two safeguards relaxed further, reached only through Anthropic’s cyber and life-sciences trusted-access programs — the life-sciences track is US-organizations-only for now — the same invitation-only shape as Mythos 5.

128K+ token context

Broadest ecosystem and tooling. Multimodal (text, vision, audio). GPT-5.6 (Jul 2026) splits into three tiers: Sol (hardest problems), Terra (high-volume business tasks), Luna (fast, low-cost everyday work) — limited preview first, general rollout following. From Aug 2026 Luna is rolling out as the default model for ChatGPT’s Free and Go plans, which also gain the “Think” button for harder questions.

Up to 1M token context

Very long context windows — useful for whole-codebase or many-document questions. Tight Google Cloud / Workspace integration. The Jul 2026 Flash refresh added 3.6 Flash (workhorse tier, ~17% fewer output tokens), 3.5 Flash-Lite (high-throughput, low cost), and 3.5 Flash Cyber (security model, government/partner-only). In the Gemini API, 3.6 Flash is now the default model for Managed Agents, and managed agents run on free-tier projects — a cheap way to try agent work without adding billing. Gemini 3.7 Flash followed three weeks later (Aug 2026), pitched almost entirely at coding and agent work: 1,048,576 tokens of input across text, image, video, audio and PDF, and up to 65,536 out. Google’s gains-over-3.6-Flash figures are its own, with no independent testing behind them, and the context window is in the API model page rather than the launch post.

500K token context

xAI’s Aug 2026 frontier release, aimed at long-running agents, coding, and interactive or visual work. Takes text and images and returns text, with four reasoning-effort levels from low to xhigh. Reachable through the xAI API, Grok Build, Cursor, OpenRouter, Vercel and Cloudflare; the weights are closed and no license is published. Two things worth knowing before quoting it: the benchmark table on the announcement is xAI’s own — competitor figures included, drawn from those vendors’ published cards rather than a common test — and the context window and modality details appear only in the developer docs, not the launch post.

Many sizes (1B–400B+)

The reference open-weight family. Huge fine-tune ecosystem; most local-inference tutorials assume Llama.

Meta’s proprietary coding-model line, separate from the open-weight Llama family. Version 1.2 (Aug 2026) is a coding-focused update and powers Muse Code, a terminal coding agent in beta for macOS and Linux; both are available through the Meta Model API. Worth knowing about, but hold it loosely: Meta has published no model size, context window, license or pricing, and its capability claims have no third-party testing yet.

~29.6B params, 131K+ context

Meta’s open multimodal model (Aug 2026), and one of the few released under a plain Apache 2.0 license — genuinely open source, not merely open weights. Takes text and images, outputs text only. Ships as full-precision BF16 plus two 4-bit quantizations sized for 24GB and 32GB consumer GPUs, which is the point: it is aimed at running privately on your own hardware for coding, document analysis and agent work. Meta’s own benchmark numbers are strong (76.0 on SWE-Bench Verified) but are vendor-reported with no third-party testing yet. The model card restricts use to over-18s and advises human-in-the-loop confirmation before an agent takes an irreversible action. Note that Meta’s launch blog and the model card disagree on the specs — the card is the one to trust.

2.6B / 3.1B / 8.3B-A1B params; 128K context, 32K on the vision model

Built for on-device work rather than the frontier: LFM2.5-2.6B and its base checkpoint run on laptops and phones through llama.cpp, MLX, vLLM or ONNX, with a browser demo over WebGPU. Good at tool calling, multi-step workflows and retrieval; Liquid explicitly says it is not the model for agentic coding or knowledge-heavy questions. Read the license before building a business on it — it is Liquid’s own “LFM Open License v1.0”, not an OSI-approved open-source license, and it withholds commercial use from organizations at or above $10M in annual revenue. A vision-language sibling arrived in Aug 2026, LFM2.5-VL-3B: a 400M SigLIP2 encoder bolted onto that same 2.6B text backbone, 3.1B parameters in total, images and text in and text out, aimed at reading screens, UIs and documents on-device. Check its context window rather than assuming it inherits one — it is 32,768 tokens, a quarter of the text model’s, and Liquid recommends it for single-turn, low-latency work rather than long-context reasoning. Same license, same revenue restriction. The largest member is LFM2.5-8B-A1B, a mixture of experts with 8.3B total parameters but only about 1.5B active per token, 128K context, text in and text out — Liquid points it at agentic workflows, tool use and structured output on device, so it is the one to reach for when the 2.6B is not enough. Liquid also publishes DSpark, a set of small draft models for speculative decoding, where a lightweight model proposes tokens that the real model then verifies: output quality is unchanged by construction and only decoding speed moves. Three checkpoints pair with targets of the same name — LFM2.5-1.2B-Instruct-DSpark (~296M), LFM2.5-2.6B-DSpark (~328M) and LFM2.5-8B-A1B-DSpark (~328M) — in safetensors and GGUF, under that same LFM Open License v1.0. Treat the speedups as hardware-specific rather than one headline number: Liquid’s own measurements peak at 3.18x for the 8B-A1B on an H100 and 2.63x for the 2.6B on an M4 Max MacBook, with no third-party testing, and the draft cards state no context window of their own.

Dense + Mixture-of-Experts variants

European lab with both hosted API and open weights. Mixtral popularized MoE for open models — good efficiency per parameter.

MoE up to 1.7T; R-series adds reasoning traces

Open-weight models with near-frontier reasoning at low cost. R-series shows its chain-of-thought — useful for studying how reasoning models work. V4-Flash-0731 (Jul 2026) is a 304B-parameter MoE released under a plain MIT license — genuinely open source, not just open weights — and independent testing puts it near the top of the value-for-intelligence rankings. V4-Pro-0813 (Aug 2026) is the much larger sibling — 1.7T parameters, text in and text out — and is also plain MIT. Be careful what else you believe about it: DeepSeek published no announcement, no spec sheet and no benchmark table, neither checkpoint states a context window on its model card, and the benchmark figures circulating for it have never come from the vendor.

Many sizes; Qwen3.8 has 262K native context, ~1M extended

Very active family covering chat, coding, vision, and audio, frequently near the top of open-model leaderboards. The Qwen3.8 generation (Aug 2026) is the clearest lesson yet in why “open weights” and “open source” are not the same phrase, because one family shipped both at once. Qwen3.8-27B is Apache 2.0 — genuinely open source — a 27B dense vision-language model taking image, video and text in and returning text, 64 layers, 262,144 tokens of native context extensible to around a million, with an FP8 checkpoint alongside it. The flagship Qwen3.8-2.4T-A95B is not: it carries Alibaba’s own Qwen3.8-Max License, which is not OSI-approved and sends model-as-a-service businesses past a revenue threshold off to negotiate a separate commercial license. That one is a 2.4T-parameter mixture of experts with about 95B active (10 routed experts plus a shared one, from a pool of 512), the same 262K native context, and it is text-only with thinking permanently on. Do not confuse it with the hosted Qwen3.8-Max API, which is the version that adds vision input, a non-thinking mode and 1M context by default. On the 27B, thinking is on by default but can be turned off per request and its effort level is adjustable; independent hands-on testing suggests it overthinks simple prompts at the default setting. An Aug 2026 preview, Qwen3.8-Flash-Next, sits outside the Qwen3.8 line proper — Qwen itself calls it an experimental early look at the architecture behind Qwen4. It is a vision-language mixture of experts: 125B total parameters with about 6B active per token (10 routed experts plus 1 shared, from a pool of 512, alongside a 51B n-gram embedding table and 4B of multi-token-prediction weights), 262,144 tokens of native context extensible to about 1M, taking text, image and video in and returning text. Its license, read off the model card, is Alibaba’s own “Qwen Community License 1.0” — distinct from both the 27B’s Apache 2.0 and the flagship’s Qwen3.8-Max license, and, like the flagship’s, not OSI-approved: it requires a separate commercial license for model-as-a-service or AI-assistant businesses, plus prominent model-name attribution once a product built on it passes 100M monthly users or $20M in monthly revenue.

K3: 2.8T total / ~104B active MoE, 1M context

Near-frontier open-weight family from China; K3 (Jul 2026) debuted #3 on Artificial Analysis. API and chat are live, and the K3 weights are now published on Hugging Face. Note the license: they are released under Moonshot’s own “Kimi K3 License”, not an OSI-approved open-source license — read the terms before building on it commercially.

GLM-5.2: 753B mixture of experts, 1M token context

One of the strongest open releases coming out of China, and a rare case of a large frontier-class model published under a plain MIT license — genuinely open source, not merely open weights. GLM-5.2 is a 753B mixture of experts, text in and text out, with a 1M-token context the model card says holds up over long-horizon work; GLM-5 and GLM-5.1 are published the same way, and each release ships an FP8 checkpoint alongside the full-precision one for cheaper serving. The line is pitched hardest at coding and agent work. Two things to hold loosely: the card does not say how many parameters are active per token, so do not assume the full 753B runs on every one, and the benchmark tables on these releases are Z.ai’s own, without independent replication.

3B / 8B / 30B dense; 128K context, extensible to 512K

IBM’s first reasoning-focused Granite generation (Aug 2026): three dense decoder-only models — 3B, 8B and 30B — each verified on its own Hugging Face card rather than the launch blog, and each genuinely Apache 2.0. Every size supports 128K tokens of context natively, extensible to 512K, and can switch between thinking and non-thinking mode, including a low-effort thinking option for easy questions. Native tool calling follows OpenAI’s function-calling format, and all three cover the same 12 languages (English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, Chinese). The 8B and 30B add a further stage of agentic reinforcement learning inside real sandboxed environments (SWE, terminal and search agents); the 3B’s tool-calling ability comes from ordinary instruction tuning rather than that sandbox stage. IBM’s launch-post benchmark figures are vendor-reported with no third-party testing, so none are quoted here.

Small sizes for local use

Google’s small open models — solid pick for laptop-scale local inference and fine-tuning practice.

Small models (~1–14B)

Research line proving small models trained on curated data can punch above their weight. Good for edge and learning experiments.

3072 / 1536 dims, truncatable

Industry-standard closed option. Matryoshka Representation Learning lets you truncate vectors (e.g. 3072 → 256 dims) to cut vector-DB cost with little accuracy loss.

Mixture-of-Experts architecture

First production-grade MoE embedding model — higher retrieval accuracy than its predecessor at lower serving cost. Strong RAG choice.

8B flagship + ~1B high-throughput variants

Tops retrieval benchmarks like RTEB. 8B version for precision-critical retrieval; hardware-accelerated small variants for high-throughput production.

Multilingual suite

Open-source frontier embedding suite designed for top-tier multilingual semantic search on standard infrastructure.

~570M params, multilingual

Popular for complex RAG pipelines — excels at cross-lingual search and dense retrieval without heavy compute.

v2 uses MoE; runs locally

The easy local option — one `ollama pull nomic-embed-text` and you have embeddings on your own machine for free experimentation.

Multilingual; pairs with Rerank

Strong multilingual retrieval, and Cohere’s Rerank models slot in behind it as a second-stage accuracy boost.

Many sizes + rerankers

Long-running open family that consistently ranks well on MTEB. bge-m3 handles dense, sparse, and multi-vector retrieval in one model.