Open-weights AI models in 2026
A practical evidence catalogue for downloadable models. Current quality claims require comparable versioned runs; missing scores are never invented.
TL;DR
There is no universal winner. Open-weights means downloadable weights in this catalogue — not an OSI-approved open-source licence. Method AIS-V2 treats every result as a versioned model × harness × settings × task record and does not convert a frozen legacy composite into a current universal quality rank. Unscored models stay unscored.
Quality claims follow Method AIS-V2. Commercial relationships cannot change rank, eligibility, evidence quality, or publication timing — money never buys rank. Submitting or claiming a listing is free and does not skip the editorial queue.
Reviewed / legacy snapshot
catalogue profile9 modelsSaved editorial catalogue snapshot. These profiles are not a current universal quality rank.
| Name | Best for | Documented price notes | Evidence status | Links |
|---|---|---|---|---|
| DeepSeek V4 FlashDeepSeek | High-concurrency, low-latency coding assistance and batch page generation on a tight budget | Peak $0.44 in / $1.32 out per 1M; off-peak exactly half at $0.22/$0.66; cache-hit input $0.014 peak / $0.007 off-peak. Peak hours 01:00-04:00 and 06:00-10:00 UTC Mon-Fri. | catalogue profile | Evidence profileCompare |
| DeepSeek V4 Pro 0813DeepSeek | Cost-controlled agentic backend and full-stack work, and teams that need MIT-licensed weights for compliance | DeepSeek direct peak: $1.32 in / $3.96 out per 1M; off-peak exactly half at $0.66/$1.98; cache-hit input $0.044 peak / $0.022 off-peak. Peak hours 01:00-04:00 and 06:00-10:00 UTC Mon-Fri. OpenRouter blended: $0.5795 in / $1.738 out, cache read $0.01932. | catalogue profile | Evidence profileCompare |
| GLM-5.2Z.ai | Teams that want an MIT-licensed, self-hostable model with good design taste and instant responses | $1.40 in / $4.40 out per 1M with $0.26 cached input (Z.ai); Alibaba resale $1.40/$4.40 in Singapore and $1.10/$3.851 in China, Germany, US and Hong Kong; `glm-5.2-us` $1.40/$4.4 | catalogue profile | Evidence profileCompare |
| GLM-5.3Z.ai | Agentic coding teams that want near-frontier terminal/SWE performance with open weights and low cost | $1.40 in / $4.40 out per 1M with $0.26 cached input (Z.ai docs; cached-input storage free for a limited time). The GLM Coding Plan uses points, with 50% off-peak discount outside 14:00-18:00 UTC+8 on weekdays. | catalogue profile | Evidence profileCompare |
| Tencent Hy3Tencent | High-volume, cost-capped code and content generation where design quality is not critical | $0.14 in / $0.55 out per 1M (Artificial Analysis) - the cheapest model in this set | catalogue profile | Evidence profileCompare |
| Tencent Hy4 previewTencent | Cost-driven teams and self-hosters who want a permissively licensed frontier-scale coding model | $0.834 in / $2.501 out per 1M with $0.042 cache hits (Tencent newsroom and OpenRouter agree) | catalogue profile | Evidence profileCompare |
| Kimi K3Moonshot AI | Design-led teams and self-hosters who want the best visual-preference model with open weights | $3 in / $15 out per 1M with $0.30 cached input (Moonshot); Alibaba resells at $2.827/$14.133 in most regions and $3/$15 in Singapore/US international; OpenRouter median $2.55/$12.75 with $0.256 cache read | catalogue profile | Evidence profileCompare |
| Qwen3.8-27BQwen | Teams that need a self-hostable, screenshot-capable frontend model on their own hardware | Hosted $0.50 in / $3.00 out per 1M (Singapore/international); $0.424/$1.696 in China (Beijing); a third-party comparison lists $0.425/$2.55; self-hosting has no token price | catalogue profile | Evidence profileCompare |
| Qwen3.8-Flash-NextQwen | Budget-conscious frontend teams and self-hosters who want frontier-adjacent UI quality | Hosted qwen3.8-flash $0.15 in / $0.47 out per 1M (Singapore/international; $0.113/$0.382 other regions); cache read $0.016, cache write $0.20; self-hosting has no token price | catalogue profile | Evidence profileCompare |
Columns are live catalogue profile fields only: name, best-for (job or model.bestFor), documented priceNote, and evidence or provisional status. Missing notes stay “Not documented”. These tables are not quality ranks.
Tracked catalogue
tracked · not ranked37 entriesCatalogue-discovery entries. They are listed for inspection, not ranked, and they are excluded from the comparison table above.
Best for, all entries: Catalogue discovery; compare availability, context, and public token pricing before benchmark review.
Meta 8
Mistral 8
Qwen 8
DeepSeek 7
Z.ai 6
FAQ
Does open-weights mean OSI open source?
No. Downloadable weights are not automatically an OSI-approved software licence. This catalogue records open-weights as a documented weights-availability fact. Check each model’s licence on its evidence profile.
Is there a number-one open-source model?
No. This page does not publish a universal quality rank. Current quality claims require comparable versioned model × harness × settings × task records. Missing scores stay missing.
What is the difference between reviewed models and tracked catalogue entries?
Reviewed / legacy-snapshot models come from the saved editorial catalogue snapshot. Tracked models are catalogue-discovery entries without a published quality rank. Tracked does not mean poor quality, and reviewed does not mean a current universal winner.
Is the comparison table a ranking?
No. The table lists documented profile facts for the reviewed set only: name, best-for, documented price notes, and evidence status. Catalogue has not published an evidence-dated order for this guide, so the table is not an editorial rank.