Models
Commercial and open-weight models — same template for both, so it stays comparable and easy to update as versions change. Every row shows when it was last checked; models move fast, don't trust a stale row.
| Model | Type | Released | Context | Self-host? | Strength | Verdict | Source |
|---|---|---|---|---|---|---|---|
| Claude (Anthropic) Anthropic |
commercial | 2026 · generation 5 | 200k | no | Four tiers (Haiku 4.5 / Sonnet 5 / Opus 5 / Fable 5) — from cheap-and-fast to complex agentic code. Good cost/quality balance at the Sonnet tier. | The reasonable default if you don't care about depending on the cloud and want the best quality/cost balance per level. | source ↗ verified 2026-08-21 |
| ChatGPT / GPT-5.6 + Codex (OpenAI) OpenAI |
commercial | 2026 · GPT-5.6 family | 128k | no | Codex is no longer a separate product — it's a unified GPT-5+Codex stack. The current code model is GPT-5.3-Codex, with a Max variant for long agentic work and a cheaper Mini. | Strong specifically for long-format/agentic code tasks via Codex; regular ChatGPT for everything else. | source ↗ verified 2026-08-21 |
| Gemini (Google) Google DeepMind |
commercial | 2026 · 3.x family | 1000k | no | 3.1 Pro and 3 Deep Think for frontier reasoning; 3.5/3.6 Flash for agents and code at lower cost; context of up to 1M tokens in the top-tier variants. | The option with the largest context window if you need to feed huge documents all at once. | source ↗ verified 2026-08-21 |
| Qwen3.6 (Alibaba) Alibaba (Qwen Team) |
open-weight | April 2026 | 262k | yes | The family with the most active release cadence in H1 2026. The dense 27B variant is the sweet spot for real self-hosting: native 262K context, multimodal, strong performance in code. | The best practical option to run locally today if you have 1-2 consumer GPUs — it's what this site runs in production. | source ↗ verified 2026-08-21 |
| Kimi K3 (Moonshot AI) Moonshot AI |
open-weight | July 2026 | 1048k | no | The largest open-source model released to date, with hybrid attention (Kimi Delta Attention) and an always-on reasoning mode. ~1M token context. | Open doesn't always mean self-hostable — this is the example: the license is free, the required hardware isn't. | source ↗ verified 2026-08-21 |
| DeepSeek V4 (DeepSeek) DeepSeek |
open-weight | April 2026 (Preview) | 1000k | no | Architectural reset to hybrid attention — the strongest efficiency story in the open-weight group on compute-per-token despite its size. | Best capability/compute ratio in the open-weight batch, but real self-hosting only if you have your own multi-GPU infrastructure. | source ↗ verified 2026-08-21 |
Local hosting notes
Claude (Anthropic)API/cloud only — no open weights.
ChatGPT / GPT-5.6 + Codex (OpenAI)API/cloud only.
Gemini (Google)API/cloud only (the open sibling is Gemma, not Gemini).
Qwen3.6 (Alibaba)The dense 27B variant runs on 1-2 consumer GPUs (Q4) — the same size as Sol, this site's primary model.
Kimi K3 (Moonshot AI)Open weights, but 2.8T parameters is datacenter scale — it doesn't fit on consumer hardware even if it's "open-weight".
DeepSeek V4 (DeepSeek)V4-Flash (the small variant) still asks for ~172GB VRAM at Q4 — out of reach for a single consumer GPU, feasible on a dedicated multi-GPU rack.