All open-weights models

Snappy 8B daily driver. Fine for chat; not an agent — fails multi-step tool loops in practice.

Qwen 3.5 9B (Mar 2 2026). Gated DeltaNet hybrid-attention multimodal vision-language model. 262K native context (up to 1M via YaRN), 201 languages, Apache 2.0. MMLU-Pro 82.5, GPQA Diamond 81.7, MathVision 78.9. Fits a 16GB laptop at Q4 (~5.7 GB). Vision: MMBench-EN 90.1, Video-MME 84.5.

Meta's stalwart 8B with 128 K context. Reliable for chat; do not expect agent reliability.

Microsoft's 14 B is a math/reasoning specialist (~56 GPQA). Don't use it for agent loops.

Apache-licensed all-rounder. Sweet spot on a single 24 GB card; not viable for agentic work.

Google's vision-capable 27 B. Strong at OCR + chat; not an agent driver.

Strong general-purpose dense 32 B with thinking variant. Marginal agent — fails multi-step tool calls often.

Solid local coder for autocomplete + small-PR work. MoE design (30 B total / 3 B active) is what gave it the speed-of-3B-dense feel — corrected from legacy 'dense' classification. Now bracketed by Qwen 3.6-27B dense (better quality) and Qwen 3.6-35B-A3B MoE (better speed) — kept for users still on it.

Apr 2026 release. 35B / 3B active MoE — beats Gemma 4-31B on agentic coding, matches Sonnet on most vision tasks. Native 262 K context (extensible to 1 M), ~18 GB at Q4. The new local-coding king under 200 B.

Apr 22 2026. Dense 27B that hits 77.2% SWE-Bench Verified — beats much larger MoEs on coding. Vision-capable, 262 K native context. Best single-24 GB-card coder right now.

26 B total / 4 B active MoE — fast and useful, but only 4 B 'thinking' caps reasoning + agent ceiling.

80 B total, 3 B active per token. Faster than 30B-coder at similar quality. Decent agent in tight loops; struggles on long-horizon work.

Solid local 70 B baseline for chat. Coding / agent quality lags Qwen-Coder, frontier MoEs, and closed models by a wide margin.

Dense 405 B — heavy bandwidth need. Now eclipsed by DeepSeek V4-Pro, Kimi K2.6 and GLM-5.1 on every coding/agent benchmark; kept for chat-only setups with 256+ GB. Q2 fits in 2× DGX Spark cluster or M3 Ultra 256.

Reasoning-distilled Llama 70 B. Best local 70 B for math/logic in early 2025; now lapped by Qwen 3.6-27B dense (better coding) and DeepSeek V4-Flash (better everything else).

Zhipu's MoE 'Air' — ~38% on SWE-rebench, the strongest sub-200B agent model. Still 20+ pts behind frontier MoEs and closed models.

First open-weights model to top SWE-Bench Pro (April 2026). 8-hour autonomous coding agent. MIT.

Z.ai's June 2026 flagship MoE successor to GLM-5.1: 753B total / 39B active, native 1M-token context via IndexShare sparse attention, two thinking-effort levels. MIT-licensed, no benchmarks at launch.

OpenAI's open-weights 120 B (MXFP4). ~33% SWE-rebench resolved. Strong all-rounder; agent ceiling is still frontier MoEs / closed models.

Apr 30 2026. Western 128B dense with vision + 256 K context. 77.6% SWE-Bench Verified; first credible mid-tier open-weight from Mistral in months. Modified MIT.

Apr 12 2026. 230 B / 10 B active MoE — agent-tuned with native multi-agent ("Agent Teams"). 57% Terminal-Bench 2.0, 56% SWE-Bench Pro — strongest sub-300 B open-weights agent.

Jun 1 2026. MiniMax's flagship MoE successor to M2.7 — 428B total / 23B active, 1M context window via MSA sparse attention, native multimodal (image/video/computer-use), open weights. Among the top open-weights reasoning models. SWE-Bench Pro 59.0 (reported), OSWorld-Verified 70.06.

Full open-weights release 2026-07-06 (Apache-2.0, dropping the April preview's custom Tencent license): 295 B / 21 B-active MoE, 256 K context. Model card reports GPQA-Diamond 90.4, SWE-Bench Verified 78, SWE-Bench Pro 57.9.

Excellent reasoning + chat. Strong coding too — but coding SOTA is GLM-5.1 / DeepSeek V3 / Kimi K2.

Original 1 T MoE — superseded by K2.5 / K2.6 for agents. Q2_K_XL ~250 GB fits 4× DGX Spark or M3 Ultra 512 GB.

Same architecture as K2; long-horizon agent training + native image input. Strong open-weights agent — close to closed frontier.

Apr 2026 release. 66.7% Terminal-Bench 2.0 — top open-weights agent. Native INT4, 262 K context, native image + video input.

Jun 2026 release — code-specialized variant built on K2.6 with ~30% fewer thinking tokens for the same task. Same 1 T total / 32 B active MoE (384 experts top-8 + 1 shared), 256 K context, Modified MIT. Internal Moonshot benchmarks only (Kimi Code Bench v2 62.0, MCP Atlas 76.0, MCP Mark Verified 81.1) — standard SWE-Bench / TB2 cells stay null until third-party leaderboards land.

May 2026 release. 1.02 T / 42 B active omnimodal MoE — #1 on OpenRouter by weekly token volume in May 2026 (~4.92 T tokens/week; the Xiaomi MiMo family reached ~13% of all OpenRouter token traffic, up from zero a year earlier, per OpenRouter May 2026 rankings). 1 M context, 384 experts top-8. Reported agentic-task parity with Claude Opus 4.6 per vendor blog; confirmed by first-party model card benchmarks (SWE-Pro 57.2, TB2 68.4, tau3-bench 72.9 — ~ties Opus 4.6 on SWE-Pro, ahead on TB2/tau3).

671 B MoE — superseded by V4 for agents. Q2 (~220 GB) fits 4× DGX Spark; Q4 needs 512 GB / 8×H100.

Apr 2026 release. 284 B / 13 B active, 1 M-token context. ~57% Terminal-Bench. The agent-grade DeepSeek that fits a quad Pro 6000 build (or Mac Studio M3 Ultra 256) at Q4.

Apr 2026 release. 1.6 T / 49 B active, 1 M-token context. 67.9% Terminal-Bench — top open-weights agent, matches closed frontier on most agentic tasks. Q2 ~430 GB.

Feb 24 2026. 122B / 10B-active hybrid Gated-DeltaNet MoE — first Qwen open-weights with native vision. On a single 128 GB DGX Spark the best first-party recipe is DFlash block-speculative decode (INT4 AutoRound + FP8 experts, vLLM 0.23, 262K ctx): ~59 tok/s general single-stream decode / ~81 tok/s on real agentic tool-call traffic (NVIDIA DevForum #374328, Jun 2026). NVFP4 underdelivered on this model; the older MTP-2 / v2.1 stack floored at ~51. Strong long-context generalist; its 27B-dense sibling out-codes it on LCB.

Feb 17 2026 flagship. 397B / 17B-active hybrid GDN-MoE — was the open-weights coding king pre-Qwen 3.6 (SWE-V 76.4, LCB v6 83.6, AIME26 91.3). First Qwen open-weights with native vision. ~25 tok/s on a 256 GB M3 Ultra with offload.

May 2026 release. 310 B / 15 B active omnimodal MoE — mid-tier sibling of MiMo V2.5-Pro. 256 experts top-8, 1 M context, MIT. Same Xiaomi MiMo family that reached ~13% of all OpenRouter token traffic by May 2026 (up from zero a year earlier).

Apr 2 2026. Google's dense flagship open-weights — 30.7 B params, 256 K context, hybrid local+global attention with toggleable <think> reasoning. Strong on AIME 2026 (89.2) and GPQA Diamond (84.3) for its weight class.

Jun 3 2026. Google's encoder-free 12B dense — unified decoder-only transformer with no separate vision/audio encoder; raw patches + audio waveforms project directly into embedding space. 256K context, 140+ langs, native multimodal (text/image/audio/video), Apache 2.0. Runs on a 16GB laptop (~8-9GB Q4). Strong for its size: AIME 77.5, GPQA 78.8, MMLU-Pro 77.2, LCB 72.0.

Jun 10 2026. Google's experimental text-diffusion Gemma — 25.2 B MoE (3.8 B active) that denoises 256-token blocks in parallel: 1,000+ tok/s on H100, 700+ on a 5090, and it runs on 18 GB cards quantized. Trades benchmark quality for ~4× generation speed vs Gemma 4 26B-A4B.

Apr 2025. Meta's 17B-active / 109B-total MoE with a 10 M-token context window — longest open-weights context as of May 2026. Strong for retrieval-heavy and document workflows; average on raw reasoning.

Apr 2025. Meta's 17B-active / 400B-total flagship Llama 4 — 128 experts, 1 M ctx, strong vision. Powers Meta's apps; competitive on MMLU-Pro and GPQA but trails frontier on coding.

Dec 2025. Mistral's first frontier-scale open MoE — 41B active / 675B total, Apache 2.0, 256 K ctx, native vision and function calling. Trained on 3000 H200s; ships at parity with leading instruction-tuned open MoEs.

May 20 2026. Cohere's first fully Apache-2.0 frontier MoE — 218 B total / 25 B active, 128 experts (8 active + 1 shared), native image-text multimodal. 128 K input ctx, 64 K output. Cohere ships a lossless W4A4 quant designed to deploy on 2× H100.

May 2026 release. 197 B / 11 B active MoE — #5 on OpenRouter weekly tokens (~2.73 T) per MACGPU May 2026 analysis. Apache 2.0, 256 K context, MTP-3 speculative-decoding head, 288 experts top-8 + 1 shared. Free tier on OpenRouter driving wide adoption.

Mar 16 2026. Unified successor to Small-3 / Magistral / Devstral — 6.5B active / 119B total MoE with toggleable reasoning, function calling and vision. Apache 2.0, 256 K ctx, 3× throughput vs Small 3.

Dec 2025. Mistral's purpose-built SWE agent — 123B dense, 256 K ctx, 72.2% SWE-Bench Verified, 32.6% Terminal-Bench 2.0. Optimised for Cline / OpenHands / Claude Code style loops. The strongest sub-200B coding agent on open weights.

Mar 2026. NVIDIA's Mamba-2 + Transformer + MoE hybrid — 12B active / 120B total, 1 M ctx, toggleable reasoning, MTP-accelerated. Strong on tool-use and long-doc workflows; commercial-OK license.

Jun 2026. NVIDIA's frontier hybrid Mamba-2 + LatentMoE + attention with MTP — 55 B active / 550 B total, native 1 M ctx (RULER@1M 94.7). SWE-V 71.9, LCB v6 89.0, GPQA 87.0, HLE 26.7. OpenMDW-1.1 (commercial OK).

Legacy 3.5 B SDXL — runs on 8 GB, huge LoRA ecosystem. Beat by FLUX on prompt + photorealism.

Apache-2.0 turbo model — very fast, decent quality, fits 8 GB cards. Smaller community vs SD/FLUX.

Apache-2.0 distilled FLUX, 4-step generation. Best fast permissive option for commercial use.

The 2026 photorealism crown — best lighting, anatomy, text rendering. Non-commercial; commercial license $$.

FLUX.2 small variant — strong prompt-following at SD-class VRAM. Commercial license required.

Black Forest Labs' 2026 flagship. Better prompt + style than FLUX.1 [dev], same VRAM footprint.

Stability's 2024-2026 flagship. Strong style versatility, mature ecosystem (ControlNet, IP-Adapter).

Apache-2.0 — best-in-class for in-image text rendering (English + Chinese), strong style range.

MIT-licensed 17 B — competes with FLUX.1 [dev] on quality without commercial restrictions. Superseded by HiDream-O1-Image (May 2026) on prompt alignment and arena rank.

May 8 2026. 8 B MIT-licensed pixel-space unified transformer (no VAE, no separate text encoder). GenEval 0.90 / DPG 89.83 — highest-ranked open-weights image model at debut. 2048×2048 native.

Real-time-ish on a 4090 — 5 s 720p clip in ~90 s. Fastest open video model; quality below WAN/Hunyuan.

Smallest open video model — 8 GB VRAM, 480p. The 'video on a laptop GPU' option.

MoE: high-noise + low-noise experts. 5 s 720p in ~3-5 min on a 5090; GGUF quants run on 12 GB.

VRAM-hungry but produces 1080p. 16 GB tile-mode possible with WanGP optimizations.

Trimmed to 8.3 B and 14 GB w/ offload. Better instruction-following + structural stability than WAN 2.2.

Original 13 B — best motion realism among open weights. 60-80 GB VRAM at full quality (datacenter).

Permissive + capable, but slower than WAN 2.2. ~8 min for 5 s clip on a 4090.