Models

Hugging Face's smallest SmolLM2 chat model at 135M parameters, in a full-precision build. Blazing fast and featherweight — good for basic completion and as the reference point in a quantization comparison.

The 360M SmolLM2 quantized to 4-bit for a tiny footprint. Handles short instructions and simple Q&A while staying extremely light on memory.

The same 360M model at full precision — larger and a bit more robust than the 4-bit build, and the natural reference for seeing how much quantization costs at this size.

A community 1.1B model on the Llama architecture, pretrained on roughly 3 trillion tokens. A classic lightweight baseline for on-device chat.

Google's smallest Gemma 3 instruction-tuned model. Surprisingly coherent for its size and well suited to short, snappy chat on modest hardware.

Meta's 1B Llama 3.2, built for edge and on-device use. Solid at summarization and simple instruction-following with a very small memory footprint.

The full-precision build of Meta's 1B model — the heavier, higher-fidelity reference to compare against the 4-bit version.

Alibaba's 0.5B Qwen2.5. Tiny but punches above its weight, with notably strong multilingual ability for the size.

The full-precision counterpart of the 0.5B Qwen2.5 — heavier but more faithful, ideal for a side-by-side quantization test.

A code-specialized 0.5B Qwen tuned on programming data. Handles basic autocomplete and snippet generation right in the browser.

The full-precision build of the 0.5B coder model — the reference quality for comparing how quantization affects code generation at tiny scale.

Microsoft's early ~1.3B research model trained on "textbook-quality" synthetic data. A landmark small model that showed strong reasoning-per-parameter.

The newer Qwen3 generation at 0.6B, with optional "thinking" mode. A clear step up from Qwen2.5 at the same size.

The full-precision Qwen3 0.6B — the higher-fidelity reference for the quantization face-off against the 4-bit build.

A capable small generalist with strong multilingual support and reliable instruction-following — a good default for quick everyday tasks.

A code-focused 1.5B Qwen, useful for autocomplete, short functions, and explaining snippets without leaving the browser.

The largest model in Hugging Face's SmolLM2 line — the strongest of the family, with solid general chat for a sub-2B model.

Qwen3 at 1.7B with an optional reasoning mode, offering a meaningful quality bump over the previous generation at the same size.

Google's Gemma 2 at 2B — a well-rounded small chat model known for punching above its size on general tasks.

The original first-generation Gemma 2B instruct model — a lightweight, dependable baseline from Google.

Meta's 3B edge model — strong at summarization and instruction-following, and one of the most popular small Llamas for on-device use.

A capable 3B generalist that balances quality and speed well for in-browser chat and light reasoning.

Among the strongest sub-7B models available, with hybrid thinking. A standout choice when you want the most capability that still fits the Mini tier.

Microsoft's compact Phi-4 (~3.8B), tuned for reasoning and math. Delivers notably strong logical performance for its size.

The classic 7B workhorse — reliable general chat with function-calling support and a large, mature ecosystem behind it.

A strong 7B generalist with good multilingual reach and reasoning — one of the most well-rounded models that runs locally.

One of the best open coding models at 7B, with competitive code generation and debugging — a genuine local coding assistant.

Meta's hugely popular 8B all-rounder, from the long-context Llama 3.1 line. A dependable general-purpose local model.

Qwen3 at 8B with hybrid thinking — excellent quality for a model that still runs in the browser, strong at reasoning and code.

Google's 9B Gemma 2 — among the better mid-size open chat models, with a noticeable quality lift over the smaller Gemmas.

A newer Qwen generation at 9B — the strongest in-browser model that still fits a typical high-memory laptop, with refined reasoning and coding.

Meta's older 13B Llama 2 chat model. Now legacy, but the largest "Small" option and a useful point of comparison against newer, smaller models.

Meta's 70B flagship-class open model, aggressively quantized to 3-bit (~31GB VRAM) so it can run locally. Approaches the quality of hosted models, but only on machines with very large memory.

IBM's tiny hybrid-architecture Granite — enterprise-grade reliability at one of the lowest prices in the catalog.

Meta's tiny 1B Llama — one of the smallest hosted models, for classification and simple text at minimal cost.

Mistral's smallest Ministral — a 3B edge-class model that is quick and inexpensive for lightweight everyday tasks.

Amazon's low-cost multimodal Nova for high-volume text and image tasks — fast and inexpensive, with the AWS ecosystem behind it.

The cheapest, text-only Nova — built for ultra-low-latency simple tasks at the lowest possible price point.

Google's cheapest and fastest Gemini tier — multimodal and well suited to high-throughput, latency-sensitive tasks.

Google's open Gemma 4 in a mixture-of-experts configuration (~26B total, ~4B active) — efficient open-weight chat at a very low price.

A larger dense Gemma 4 model — strong open-weight quality while staying firmly in the budget tier.

A long-running community favorite for creative writing and roleplay — a Llama-2 merge with a distinctive prose style at a rock-bottom price.

IBM's compact Granite 4.1 — a dependable, enterprise-tuned 8B model for everyday tasks at budget cost.

Meta's proven 8B workhorse — a hugely popular open model with excellent price-performance for simple tasks.

Microsoft's compact Phi-4, hosted in the cloud — punches above its size on reasoning and math for a low cost.

The largest of Mistral's edge-focused Ministral line — strong quality for its size with flat, symmetric pricing.

Mistral's efficient 8B Ministral — fast, capable, and cheap, built for high-volume everyday tasks.

NVIDIA's efficient Nemotron mixture-of-experts (~30B total, ~3B active) — cheap and fast, tuned for cost-effective deployment.

OpenAI's smallest and cheapest GPT-5.4 tier — built for fast, simple tasks where you still want the GPT family's polish.

OpenAI's open-weight 20B mixture-of-experts model — remarkably cheap to run and part of the open GPT-OSS line.

Alibaba's dense 32B Qwen3 — a strong-value generalist that delivers a lot of capability for a budget price.

A code-specialized Qwen3 mixture-of-experts (~30B total, ~3B active) — capable coding assistance at an unusually low price.

An open-weight Qwen3 Coder mixture-of-experts (80B total, ~3B active) with a 262K context — optimized for coding agents and local deployment.

Poolside's compact 33B Laguna model — quick, inexpensive responses with a focus on coding and everyday tasks.

A compact dense Qwen3.6 (~27B) — solid general reasoning and coding at a lightweight price point.

Alibaba's efficient mixture-of-experts model (35B total, ~3B active) — strong general reasoning and coding at budget pricing with fast, low-cost inference.

Tencent's Hunyuan mixture-of-experts model (~13B active) — solid general-purpose quality at a low budget-tier price.

Amazon's newer-generation Lite with added reasoning — multimodal and a clear step up from the original Nova Lite, sitting in the mid tier.

Anthropic's fast, affordable Claude — an excellent balance of speed, quality, and price for high-volume work and quick interactions.

A lightweight model from ByteDance's Seed team, aimed at efficient general-purpose chat at low cost.

DeepSeek's fast, low-cost mixture-of-experts model — surprisingly capable for the price and a strong value-tier choice.

Google's fast workhorse Gemini — strongly multimodal with a very large context window, tuned for speed at scale.

Google's latest fast Gemini — an upgraded workhorse with strong multimodal skills and a very large context window.

Inception's diffusion-based language model — generates tokens in parallel for exceptionally fast responses.

Meta's efficient Llama 4 mixture-of-experts — multimodal with a huge context window, tuned for speed and low cost.

Mistral's small-tier model — efficient, multilingual, and tool-use capable, a strong value pick for everyday work.

NVIDIA's larger Nemotron mixture-of-experts (~120B total, ~12B active) — strong capability at a strikingly low cost.

OpenAI's mid-tier GPT-5.4 — strong reasoning and broad capability at a fraction of the full model's cost.

OpenAI's efficient GPT-5.6 variant — fast, capable everyday performance at the lightest price point in the family.

The pro-grade take on GPT-5.6 Luna — deeper reasoning effort on the same efficient tier and price.

OpenAI's larger open-weight mixture-of-experts model — remarkable capability for the price, anchoring the open GPT-OSS line.

A fast, cost-efficient variant of Qwen3 Coder Plus with a 1M context window — tuned for high-throughput coding agents.

Poolside's 118B Laguna model — capable general reasoning and coding at an aggressive midrange price.

Alibaba's balanced general-purpose Qwen3.7 model — strong reasoning and instruction-following at midrange pricing.

A fast model from the Chinese lab StepFun, with multimodal capability and an emphasis on efficiency at low cost.

AI21's hybrid SSM-Transformer (Mamba) model — designed for very long context and efficient throughput on long documents.

Aion Labs' reasoning-focused flagship — a hybrid-architecture model built for deep multi-step reasoning.

Amazon's most capable Nova — built for complex reasoning and multimodal tasks, and used as a teacher model for distilling the smaller Novas.

Amazon's balanced flagship Nova — strong multimodal performance at an attractive price, the workhorse of the Nova family.

Anthropic's balanced flagship — excellent at reasoning, coding, and writing, and a popular default for serious everyday work.

Anthropic's most capable Sonnet-class model — frontier coding and agentic performance with a 1M-token context window.

Cohere's enterprise flagship — optimized for retrieval-augmented generation, tool use, and agentic workflows, and efficient to deploy.

A lighter, cheaper take on Gemini 3.5 Flash — quick multimodal responses at a fraction of the cost.

Meta's Llama 4 mixture-of-experts flagship for open models — strong general and multimodal performance at a very competitive price.

Mistral's coding- and agent-focused model, built specifically for software-engineering workflows and tool use.

Mistral's top general-purpose model — strong multilingual ability and reasoning, competitive with other flagships at a friendlier price.

Mistral's balanced mid-tier model — solid general performance with a strong quality-to-cost ratio.

Mistral's newest medium-tier model — a clear capability step up from Medium 3.1, closing in on flagship quality.

The balanced middle of the GPT-5.6 family — strong all-round quality between Luna and Sol.

The pro-grade take on GPT-5.6 Terra — extra reasoning effort on the balanced mid tier of the family.

Poolside's larger Laguna model — flagship-tier reasoning and coding at a strikingly low price.

Alibaba's large mixture-of-experts flagship (397B total, ~17B active) — top-tier reasoning and coding that rivals other flagships at a competitive price.

Tencent's Hunyuan 3 — flagship-class capability at an aggressively low price point.

Xiaomi's MiMo flagship — strong reasoning and coding performance at a low price, part of China's competitive open-model wave.

Anthropic's creative flagship — tuned for long-form writing, narrative, and nuanced tone at the very top of the range.

Anthropic's most capable model — top-tier reasoning, coding, and agentic ability, and one of the reference points for the current frontier.

DeepSeek's frontier mixture-of-experts at a tiny fraction of Western frontier pricing — a standout value champion for reasoning and code.

Google's frontier multimodal model — an enormous context window and strong performance across reasoning, code, and vision.

An open-weight frontier-adjacent model (released June 2026) pairing elite agentic and coding benchmarks with a 1M-token context and multimodality — at a remarkably low price. Note: this is launch promo pricing; the standard rate is roughly $0.60 / $2.40.

Moonshot AI's general-purpose frontier Kimi — strong reasoning and agentic ability at a sharply competitive price.

Moonshot AI's code-focused frontier model — tuned for agentic coding and long-context software tasks at a competitive price.

Moonshot AI's flagship frontier model — its most capable Kimi yet, aimed squarely at the top Western frontier tier.

NVIDIA's largest Nemotron mixture-of-experts model (550B total, ~55B active) — frontier-grade reasoning at an aggressive price.

OpenAI's flagship GPT-5.5 — top-tier reasoning, coding, and multimodal capability, and a benchmark for the current frontier.

OpenAI's premium GPT-5.5 — extended reasoning effort for the most demanding tasks, at a top-shelf price.

The most capable GPT-5.6 variant — top-shelf reasoning, coding, and agentic performance at the frontier.

The pro-grade take on GPT-5.6 Sol — maximum reasoning effort for the hardest tasks in the GPT-5.6 family.

Alibaba's flagship coding model — a 480B-total / 35B-active mixture-of-experts with agentic function calling and repo-level reasoning.

Alibaba's largest Qwen — frontier-level quality with exceptional coding performance, all at a fraction of Western frontier prices.

xAI's frontier Grok — strong reasoning with a real-time, current-events lean, and competitive value among the top tier.

xAI's frontier Grok — strong reasoning with a real-time, current-events lean at a competitive price for the top tier.

Zhipu/Z.AI's latest GLM frontier model — improved reasoning and agentic coding over GLM-5.1, still keenly priced for the top tier.