{"schemaVersion":"1.0","type":"TechArticle","types":["Article","TechArticle"],"slug":"best-open-weight-llms-september-2026-d4ujd","url":"https://zyvop.com/best-open-weight-llms-september-2026-d4ujd","title":"Best Open-Weight LLMs, September 2026","subtitle":"RAM needs, real license terms, API pricing, and benchmarks for GLM-5.3, Kimi K3, Qwen3.8, DeepSeek V4/V4.1, Tencent Hy4, and Gemma 4 - verified against primary sources.","tldr":"Which open-weight LLM should you actually run in September 2026? This guide checks the leading models against primary sources — model cards, license files, and independent trackers — covering RAM requirements, real license terms, API pricing, and where benchmark rankings disagree.","keywords":["self-hosting","AI licensing","open-weight LLMs","llm-benchmarks","local-ai"],"entities":["Anshu Pathak","self-hosting","AI licensing","open-weight LLMs","llm-benchmarks","local-ai","ZyVOP"],"keyTakeaways":["The complete comparison: quality, RAM needed, cost, licenses (verified, in plain English) Last checked: September 20, 2026.","See the changelog in Section 13 for what was corrected.","AI moves fast; re-check anything you plan to build a business on."],"headings":["The complete comparison: quality, RAM needed, cost, licenses (verified, in plain English)","0. The 60-second answer","1. How to read this guide","2. RAM explained in two minutes","The basic formula","The mixture-of-experts (MoE) trap ✅","What quality do you lose? (real example)","RAM vs VRAM vs \"unified memory\"","Long context costs memory too","3. The contenders at a glance","4. RAM cheat sheet: how much memory each model needs","4A. Models you can run at home","4B. Big models (workstation / server)","5. \"What can MY machine run?\"","6. Quality comparison","6A. Independent scores","6B. Head-to-head benchmarks (vendor-reported by Z.ai) ⚠️","6C. Other independent or semi-independent facts","7. Licenses: what you're actually allowed to do","8. Cost comparison","API price per 1 million tokens (input / output)","API vs self-hosting","9. Model profiles","Kimi K3: the biggest and (marginally) broadest"],"outboundLinks":["https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v4-3","https://artificialanalysis.ai/models/comparisons/glm-5-3-vs-kimi-k3","https://artificialanalysis.ai/models/kimi-k3","https://llm-stats.com/","https://benchlm.ai/best/open-source","https://huggingface.co/zai-org/GLM-5.3","https://huggingface.co/moonshotai/Kimi-K3/raw/main/LICENSE","https://github.com/MoonshotAI/Kimi-K3","https://huggingface.co/tencent/Hy4-preview","https://github.com/QwenLM/Qwen3.8","https://huggingface.co/google/gemma-4-31B","https://ai.google.dev/gemma/docs/releases","https://recipes.vllm.ai/deepseek-ai/DeepSeek-V4.1-Flash","https://openai.com/index/introducing-gpt-oss/","https://unsloth.ai/docs/models/glm-5.3","https://unsloth.ai/docs/models/qwen3.8","https://unsloth.ai/docs/models/deepseek-v4","https://unsloth.ai/docs/models/gemma-4","https://www.spheron.network/blog/deploy-glm-5-3-gpu-cloud/","https://thenewstack.io/zai-glm-weights-license/","https://simonwillison.net/2026/Jul/27/kimi-k3/","https://www.kdnuggets.com/why-deepseek-v4-1-flash-is-such-an-exciting-open-model-release","https://apidog.com/blog/what-is-deepseek-v4-1-flash/","https://the-decoder.com/alibabas-qwen-team-releases-qwen-3-8-models-with-open-weights-under-the-apache-2-0-license/","https://technode.global/2026/09/04/tencent-hy4-preview-open-source-model/","https://www.axios.com/2026/07/20/ai-us-china-open-source-kimi","https://techcrunch.com/2026/07/20/openai-is-scared-of-open-weight-models-should-the-us-be/"],"contentText":"The complete comparison: quality, RAM needed, cost, licenses (verified, in plain English) Last checked: September 20, 2026. See the changelog in Section 13 for what was corrected. AI moves fast; re-check anything you plan to build a business on. 0. The 60-second answer You want... Pick Memory you need Smartest open model, period GLM-5.3 or Kimi K3 (effectively tied) 245 GB to 1.5 TB+ (server-class) Best model to run on your own laptop / gaming PC Qwen3.8-27B 16-19 GB (4-bit) Best \"runs on a big Mac / workstation\" DeepSeek V4-Flash-0731 or GLM-5.3-Flash 100-210 GB Cheapest good API GLM-5.3-Flash or DeepSeek V4.1-Flash none (hosted) Cleanest license (do anything) Qwen3.8-27B, Gemma 4, Tencent Hy4 (Apache 2.0) or DeepSeek, GLM-5.3-Flash (MIT) varies Tiny / phone / edge Gemma 4 E2B / E4B 3-6 GB Reality check: the best open models still trail the best closed ones. On Artificial Analysis's Intelligence Index v4.3, the top open models score 44-45 while Claude Fable 5.1 and GPT-6 Astra score 53. But the open models cost a fraction as much per token. 1. How to read this guide \"Open-source\" vs \"open-weight.\" Almost every model here is open-weight: you can download and run it, but the training data and sometimes the license are not fully open. Several licenses have revenue conditions (see Section 7). Confidence labels used below Label Meaning ✅ Confirmed by 2+ independent sources, or by the primary source itself (model card, license file, vendor docs) ⚠️ Single source, or vendor-reported (the maker measured it themselves) ❓ Sources disagree (explained in Section 10) Two rules that prevent most misreadings of this guide: Don't compare scores across trackers. Artificial Analysis, llm-stats and BenchLM use different scales and different underlying evals. A 44 on one is not comparable to a 53 on another. Don't compare scores across index versions. Artificial Analysis showed GLM-5.3 at 60 in August and 45 in September because the index itself was rebuilt (v4.3). Honest limits. I did not run these models myself. Benchmark scores come from the model makers or from third-party trackers. Memory numbers come mostly from Unsloth's run-locally guides (a team that publishes compressed versions), model cards, and community tests. Where they disagree, I say so. 2. RAM explained in two minutes The basic formula Memory needed is roughly number of parameters × bytes per parameter, plus 10-20% overhead, plus extra for long conversations. Precision Bytes per parameter 27B-model example 16-bit (BF16) 2 ~54-56 GB 8-bit 1 ~27-31 GB 4-bit 0.5 ~16-19 GB 2-bit 0.25 ~9-11 GB Smaller precision = less memory but some quality loss. This is called quantization. The mixture-of-experts (MoE) trap ✅ Most giant models here are MoE: a huge total size, but only a slice (\"active parameters\") does the work per word. Active parameters set speed, not memory. Total parameters set memory. Every expert must be loaded, because any word might need any expert. So \"Kimi K3 has only 104B active\" does not mean it fits in 104 GB. It needs memory for all 2.8 trillion parameters. What quality do you lose? (real example) Unsloth measured how often compressed GLM-5.3 picks the same next word as the full-size model (\"top-1 match\"). This is their own metric ⚠️, and these are file sizes, not the memory you need to run them (see the next subsection): Version File size Top-1 match 1-bit (UD-IQ1_S) 216.7 GB 72.6% 2-bit (UD-IQ2_M) 238.6 GB 78.5% 2-bit (UD-Q2_K_XL) 253.9 GB 80.9% 3-bit (UD-Q3_K_XL) 343.0 GB 88.9% 4-bit (UD-Q4_K_XL) 467.3 GB 94.3% 6-bit (UD-Q6_K_XL) 684.4 GB 96.6% Rule of thumb: 4-bit is the sweet spot; 1-2 bit is \"it runs, but expect weaker answers.\" Note on a conflicting figure ❓: The New Stack reported that GLM-5.3's 2-bit quant retains \"about 86 percent\" top-1 accuracy. Unsloth's own published table, which is the source of that claim, shows 78.5% for UD-IQ2_M and 80.9% for UD-Q2_K_XL. This guide uses Unsloth's numbers. RAM vs VRAM vs \"unified memory\" VRAM = graphics-card memory. Fastest. RAM = normal computer memory. Slower, usually more of it. Unified memory (Apple Silicon Macs) = one shared pool, which is why big-memory Macs are popular for local AI. Unsloth's rule of thumb: RAM + VRAM should be about the size of the model file. If not, the model spills to your SSD and gets painfully slow. Long context costs memory too Some models are cheap here, others are not. DeepSeek V4-Flash needs about 10 GB for a full 1M-token conversation ⚠️. DeepSeek V4.1-Flash's new design uses about 0.9 GB for the same length ✅ (890 bytes per token). 3. The contenders at a glance Model Maker Released Total / active params Context Inputs License Kimi K3 Moonshot AI Jul 16 (weights Jul 27) ✅ 2.8T / 104B ✅ ~1M Text + vision Custom \"Kimi K3 License\" GLM-5.3 Z.ai Aug 14 (weights Aug 28) ~744-753B / ~40B ❓ 1M (128K max output) Text Custom \"GLM-5.3 License\" GLM-5.3-Flash Z.ai Aug 2026 320B / 18B 1M (131K max output) Text + images + video ✅ MIT Qwen3.8-2.4T-A95B Alibaba Aug 12, 2026 2.4T / 95B 262K native (~1M via YaRN) Text Qwen3.8-Max License ✅ Qwen3.8-27B Alibaba Aug 2026 27B dense 262K (~1M via YaRN) Text + vision Apache 2.0 ✅ DeepSeek V4-Pro-0813 DeepSeek Aug 13 1.6T / 49B 1M Text MIT DeepSeek V4-Flash-0731 DeepSeek Jul 31 (replaces the April preview) ✅ 284B / 13B 1M Text MIT DeepSeek V4.1-Flash DeepSeek Sep 10 552B backbone + 196B \"Engram\" ≈ 763B / 8B prefill, 16B decode ✅ 1M (384K max output) Text + images MIT Tencent Hy4 preview Tencent Aug 28 770B / 49B ✅ 1M+ Text only Apache 2.0 ✅ Gemma 4 (E2B, E4B, 12B, 26B-A4B, 31B) Google Mar 31 / Apr 2, 2026 (12B added June; all checkpoints refreshed Jul 15) ✅ 26B-A4B has 3.8B active; 31B dense 128K-256K Text, image, video (all sizes); audio on E2B / E4B / 12B ✅ Apache 2.0 gpt-oss-120b / 20b OpenAI Aug 2025 117B / 5.1B and 21B / 3.6B 128K Text Apache 2.0 Older families: Llama 4 (2025) now sits behind the leaders in every ranking I found ✅. It is no longer the \"default\" pick. Re-pull your Gemma 4 weights if you downloaded them before July 15, 2026. Google refreshed every checkpoint that day with Flash Attention 4 support, tool-calling reliability fixes and a new vision configuration, under the same version name. 4. RAM cheat sheet: how much memory each model needs All numbers = total memory (RAM + VRAM, or unified memory) for the model plus a modest amount of context. Sources: Unsloth guides unless noted. 4A. Models you can run at home Model Smallest usable Everyday choice High quality Full precision Gemma 4 E2B 3 GB (QAT) ~4-8 GB ~5 GB (8-bit) ~10 GB Gemma 4 E4B 5 GB (QAT) 5.5-6 GB (4-bit) 9-12 GB (8-bit) ~16 GB Gemma 4 12B 7 GB (QAT) 8 GB (4-bit) 14 GB (8-bit) ~25 GB gpt-oss-20b 16 GB (official) 16 GB n/a n/a Gemma 4 26B-A4B 15 GB (QAT) 18 GB (4-bit) 28 GB (8-bit) ~52 GB Gemma 4 31B 18 GB (QAT) 20 GB (4-bit) 34 GB (8-bit) ~62 GB Qwen3.8-27B 7-8 GB (1-bit), 9-11 GB (2-bit) ✅ 16-19 GB (4-bit) ✅ 23-26 GB (6-bit), 31 GB (8-bit) 56 GB (BF16) gpt-oss-120b ~80 GB (official) ~80 GB n/a n/a Qwen3.8-Flash-Next (125B preview) ⚠️ 75 GB (1-bit) not verified not verified not verified Practical notes: Qwen3.8-27B: the full ladder per Unsloth is 7-8 GB (1-bit) / 9-11 (2-bit) / 12-14 (3-bit) / 16-19 (4-bit) / 23-26 (6-bit) / 31 (8-bit) / 56 (BF16). Unsloth explicitly states that 4-bit runs on a 16-19 GB card such as an RTX 5080 or 4090, or a Mac with 24 GB. Add 1-2 GB of headroom if you enable MTP for faster inference. NVFP4 builds run ~1.5× faster than BF16 on Blackwell cards (RTX 50-series, DGX Spark, B200/B300) in 24 GB of VRAM. Gemma 4 speed (Apple M3 Max, 96 GB): 26B-A4B ran ~56 tokens/sec at 4-bit; the 31B dense ran ~12 ⚠️. The MoE version is much faster because only 3.8B parameters work per word. gpt-oss-120b: OpenAI says it fits in 80 GB ✅. One user reported a 63 GB file on a Mac ⚠️. 4B. Big models (workstation / server) Model Smallest Mid Best practical Full DeepSeek V4-Flash-0731 83 GB (1-bit) 110 GB (3-bit; 103 GB file) ✅ 155 GB (4-bit) to 162 GB (8-bit) ✅ native 167 GB (180 GB w/ 1M context) GLM-5.3-Flash ~100 GB (1-bit) 128-150 GB (3-bit) 162-210 GB (4-bit) 350 GB (8-bit); 650 GB (BF16) GLM-5.3 223 GB (1-bit) ✅ 245 GB (2-bit); 290-360 GB (3-bit) ✅ 372-475 GB (4-bit); 570 GB (6-bit) ✅ 810 GB (8-bit); ~1.5 TB (BF16) DeepSeek V4-Pro-0813 ~520 GB (2-bit community) ~870 GB (4-bit community) ~900 GB+ native ~865 GB weights Qwen3.8-2.4T-A95B 397 GB (1-bit XXXS; needs ~450 GB RAM) ✅ 508 GB (1-bit IQ1_S) 657 GB (2-bit) 2.6 TB (8-bit); 4.9 TB (BF16) ✅ Hy4 preview no verified small build n/a ~770 GB (FP8) ~1.54-1.56 TB (BF16) Kimi K3 610 GB (1-bit) 665-726 GB (1-bit M / 2-bit) 880 GB (2-bit XL) 1.56 TB ✅ DeepSeek V4.1-Flash ⚠️ see note see note see note see note DeepSeek V4.1-Flash note: despite the \"Flash\" name it is not a small model. Community 2-bit files total ~340 GB on disk, of which ~189 GB is the \"Engram\" lookup table that is read from your SSD rather than held in RAM; a 4-bit build is ~483 GB on disk. Mainstream tools (Ollama / LM Studio style) had no verified easy recipe when I checked, and vLLM support landed only on nightly builds from 2026-09-10. One reviewer found no validated setup for 24-512 GB consumer machines. Use the API for now. Cloud GPU example (GLM-5.3, Spheron prices ⚠️): FP8 needs 8× H200 (1,128 GB), about $20.56/hr spot or $38.32/hr on-demand. A 4-bit build fits on 4× H200 (564 GB), about $10.28/hr spot or $19.16/hr on-demand. 5. \"What can MY machine run?\" Your machine Best fit Also works 8 GB RAM laptop Gemma 4 E2B Gemma 4 E4B (tight), Qwen3.8-27B at 1-bit 16 GB Gemma 4 26B-A4B (QAT, tight), gpt-oss-20b Gemma 4 12B, E4B, Qwen3.8-27B at 2-3 bit 24 GB GPU / 24-32 GB Mac Qwen3.8-27B (4-bit) ✅, Gemma 4 31B Gemma 4 26B-A4B at 8-bit 64 GB Qwen3.8-27B at 8-bit or BF16 Gemma 4 31B at 8-bit 96-128 GB Mac / workstation DeepSeek V4-Flash-0731 (3-bit), gpt-oss-120b, Qwen3.8-Flash-Next (1-bit) GLM-5.3-Flash 1-2 bit (needs ~112 GB) 192 GB DeepSeek V4-Flash (4-8 bit), GLM-5.3-Flash (3-4 bit) 256 GB Mac GLM-5.3 (2-bit, 245 GB) GLM-5.3-Flash (4-bit) 512 GB GLM-5.3 (3-4 bit), Qwen3.8-2.4T (1-bit, 397 GB file / ~450 GB RAM) 8× 80 GB GPUs (640 GB) Kimi K3 (1-bit, 610 GB, very tight), GLM-5.3 (6-bit, 570 GB) 8× H200 / B200 (1.1-1.5 TB) GLM-5.3 FP8, Hy4 FP8, DeepSeek V4-Pro, Kimi K3 (2-bit) 8× B300 (~2.3 TB) Kimi K3 full precision Above 128 GB, running these locally is possible but slow. For most people the API is the practical route to the top models. 6. Quality comparison 6A. Independent scores Artificial Analysis Intelligence Index v4.3 (Sept 7) ✅ (independent tester) Model Score Claude Fable 5.1 (closed, reference) 53 GPT-6 Astra (closed, reference) 53 GLM-5.3 (max) 45 Kimi K3 (max) 44 GLM-5.3-Flash 42 Qwen3.8 2.4T A95B 40 DeepSeek V4 Pro 0813 (max) 36 Read the reasoning effort, not just the model name. ✅ Artificial Analysis scores each reasoning effort separately, and the spread is large: Kimi K3 scores 44 at max effort but around 30-34 at low effort. Every number in the table above is the model's max-effort score. A deployment that runs these models at low or medium effort — which is what most cost-conscious setups do — will not see these numbers. The top two open models are 8-9 index points behind the top closed models. DeepSeek V4.1-Flash reportedly scores ~40 ⚠️. llm-stats \"overall\" leaderboard (Sept 20) ⚠️: Kimi K3 53.0 → GLM-5.3 52.7 → Atria Dawn Preview (Shanghai AI Lab, 744B MoE, new preview) 52.3 → Qwen3.8 Max 52.0 = DeepSeek V4-Pro-0813 52.0. Closed reference: GPT-6 Astra 59.6, Fable 5.1 55.7. These are not the same scale as the Artificial Analysis numbers above — do not read Kimi K3's \"53.0\" here as equal to Fable 5.1's \"53\" there. BenchLM (Sept 18) ❓: ranks Qwen3.8 Max first (73.2) ahead of GLM-5.3 (66.9). This disagrees with Artificial Analysis. One likely reason: Alibaba's hosted \"Max\" differs from the downloadable 2.4T release — the open version is text-only and thinking-always-on, while hosted Max adds vision, optional thinking and built-in tools. Treat this ranking cautiously. 6B. Head-to-head benchmarks (vendor-reported by Z.ai) ⚠️ Z.ai chose these benchmarks, so expect a flattering picture. Nobody outside Z.ai had fully reproduced them when I checked. Every figure below was verified against Z.ai's published table. Benchmark (higher = better) GLM-5.3 Kimi K3 DeepSeek V4-Pro-0813 Qwen3.8-Max Closed reference Terminal Bench 2.1 88.2 88.3 87.9 86.6 GPT-5.6 Sol 88.8 Terminal Bench 3.0 (hard shell tasks) 28.3 17.4 n/a n/a GPT-5.6 Sol 34.6, Fable 5 33.7 DeepSWE v1.1 (fix real GitHub issues) 66.9 67.5 62.7 56.6 GPT-5.6 Sol 72.7 NL2Repo 58.0 58.0 61.1 55.9 Opus 4.8 69.7 SWE-Marathon v1.1 42.5 48.1 n/a n/a Opus 4.8 48.8 Toolathlon Verified (tool use) 73.0 76.5 74.1 72.5 Opus 4.8 76.2 CyberGym (finding vulnerabilities) 84.5 80.0 83.3 78.5 Fable 5 83.8 HLE with tools 62.5 59.8 60.0 56.2 GPT-5.6 Sol 64.5, Fable 5 63.9 Reading this simply: GLM-5.3 and Kimi K3 trade blows. GLM leads on the hardest terminal tasks and cyber; K3 leads on some bug-fixing and tool-use tests; DeepSeek V4-Pro even wins one. Nobody wins everything. 6C. Other independent or semi-independent facts GDPval-AA v2 (real-world agent work): Artificial Analysis measured Kimi K3 at 1,668 Elo ✅. Z.ai's own table puts GLM-5.3 at 1,769 and K3 at 1,682 ⚠️ — a small discrepancy in K3's number between the two sources, worth noting before quoting either. AA-Omniscience (factual accuracy): Kimi K3 20, GLM-5.3 14 ✅. K3 is more accurate on facts. Kimi K3 tops the Frontend Code Arena (web UI generation) ✅ and posts 93.5% on the GPQA Diamond science test ⚠️. DeepSeek V4-Flash-0731 was a large jump over the April preview ⚠️ (DeepSeek-reported): DeepSWE 7.3 → 54.4, CyberGym 38.7 → 76.7, Terminal Bench 2.1 61.8 → 82.7. Tencent's blind eval for Hy4 ⚠️: 163 Tencent experts rated 203 engineering tasks; Hy4 preview scored 2.99 / 4.00, narrowly ahead of GLM-5.3 and Kimi K3. It is an internal, vendor-run evaluation. Tencent also reports GPQA Diamond 92.3 and SWE-bench Multilingual 82.9 ⚠️. Cost per task: similar index scores can come at very different costs, as Artificial Analysis itself points out. GLM-5.3-Flash and GPT-5.6 Terra both score 42, but GLM-5.3-Flash costs about 18% as much per task ($0.25 vs $1.40) ✅. 7. Licenses: what you're actually allowed to do Model License Plain-English summary Qwen3.8-27B Apache 2.0 ✅ Do almost anything, including commercial use. Gemma 4 Apache 2.0 ✅ Same. No usage caps, no revenue thresholds. Hy4 preview Apache 2.0 ✅ Same, at frontier scale. gpt-oss Apache 2.0 + OpenAI usage policy ✅ Permissive. DeepSeek V4 / V4.1 MIT ✅ Do almost anything. GLM-5.3-Flash MIT ✅ Do almost anything. GLM-5.3 Custom GLM-5.3 License ✅ Individuals and most companies can run, fine-tune and deploy it. Companies that host the model and earn over $10B in any 12 months must pass a Z.ai security review before commercial use. Notably, GLM-5.1 and 5.2 were MIT; 5.3 is not. Kimi K3 Custom Kimi K3 License ✅ Free for most use. Model-as-a-Service providers whose aggregate revenue exceeds $20M over any consecutive 12 months must enter a separate agreement with Moonshot before any commercial use. Services above 100M monthly active users or $20M monthly revenue must prominently display \"Kimi K3\" in the UI. One source reports revenue sharing of up to 30% for large inference providers ⚠️. Moonshot consistently calls this \"open weight,\" not \"open source.\" Qwen3.8-2.4T-A95B Qwen3.8-Max License ✅ Not Apache 2.0 — a separate custom license shipped in the same announcement as the Apache-2.0 27B. Grants use, modification, hosting, fine-tuning and redistribution, with two conditions: companies running a Model-as-a-Service or \"AI Work Assistant\" business above $50M revenue over any consecutive 12 months must obtain a separate license from Qwen, and products above 100M monthly active users or $20M monthly revenue must display the model name in the UI. Bottom line: if you are a normal developer or small company, all of these are usable. The conditions matter mainly if you plan to resell the model as a hosted service at scale. Policy risk (not a license issue) ✅ reported, ⚠️ unresolved: After Kimi K3's July release, Axios reported (July 20) that the administration was reviving efforts to restrict Chinese AI models in the US on cybersecurity grounds, and that the Commerce Department had previously considered adding Chinese AI labs to the Entity List. Politico separately reported that Commerce would not take that step. Critics including David Sacks argued a ban would hand the market to the leading closed labs; Anthropic's Dario Amodei publicly stated that Anthropic has never backed a ban on open-weight models. Attribution of the push to specific US labs is contested, and nothing had been enacted when I checked. If you are in a regulated industry or work with government, check your legal position. 8. Cost comparison API price per 1 million tokens (input / output) Model Price Notes GLM-5.3-Flash $0.15 / $0.50 ✅ Z.ai list price. The 50% launch promo ($0.075 / $0.25) ended Sept 9, 2026; some routes lagged the change. MIT-licensed. DeepSeek V4.1-Flash $0.15 / $0.60 off-peak; $0.30 / $1.20 peak ✅ GA Sept 10. Model id deepseek-flash. Third-party hosts list ~$0.22 / $0.66. DeepSeek V4-Flash-0731 ~$0.22 / $0.66 Prices have changed several times ⚠️ DeepSeek V4-Pro-0813 $0.66 / $1.98 off-peak; $1.32 / $3.96 peak Superseded: V4-Pro API requests were rerouted to V4.1-Flash on Sept 14 at 04:00 UTC ✅. Do not plan around V4-Pro as a separately served endpoint. Hy4 preview $0.834 / $2.501 ✅ Tencent Cloud TokenHub / OpenRouter GLM-5.3 $1.40 / $4.40 Third-party hosts can be cheaper Qwen3.8 (2.4T / Max) ~$2 / $6 ✅ QwenCloud and Novita; other hosts $1.65-2 input Kimi K3 $3 / $15 ✅ The premium open option; flat at any context length Claude Fable 5.1 (closed, reference) $10 / $50 GPT-6 Astra (closed, reference) $10 / $50 Open models are roughly 3× to 100× cheaper on output tokens than the top closed models, but they are also less capable. Prices vary by provider and change often. API vs self-hosting For the big models, the API almost always wins on cost. One analysis found that renting an 8× H100 setup (~$25.60/hr) rarely beats DeepSeek's very cheap Flash API ⚠️. The main reason to self-host is privacy or control, not savings. 9. Model profiles Kimi K3: the biggest and (marginally) broadest Strengths: top factual accuracy, front-end and UI generation, tool use, native vision, 1M context. Weaknesses: enormous (1.56 TB), priciest open API, custom license with revenue clauses. Moonshot's own guidance calls for 64+ accelerators. Memory: 610 GB minimum (1-bit), 1.56 TB full. Next up: a successor is reportedly in post-training, but no model card, weights or API existed as of Sept 20 ⚠️. GLM-5.3: the best all-round open model right now Strengths: top on Artificial Analysis by a hair (45 at max effort), strongest on hard agentic coding and cyber tasks (vendor-reported), cheaper than K3, 1M context. Weaknesses: new non-MIT license, benchmark claims mostly self-reported, weights only released Aug 28. Memory: 245 GB for a usable 2-bit (fits a 256 GB Mac); ~470 GB for near-full quality (4-bit). Cyber note: Z.ai delayed the weights two weeks for safety review because of the model's strong vulnerability-finding ability, and published a ledger of 2,436 findings across 269 open-source projects attributed to the GLM series ⚠️. GLM-5.3-Flash: the smart budget pick Strengths: MIT license, third place among open models on Artificial Analysis (42), $0.15 / $0.50 API, natively multimodal (text, image, video), hybrid sparse + linear attention that cuts KV cache ~4.4× versus GLM-5.3. Weaknesses: still needs ~100 GB+ of memory locally — this is datacenter self-hosting, not laptop self-hosting. Memory: 100 GB (1-bit) to 210 GB (4-bit). Qwen3.8-27B: the best \"actually runs on my computer\" model Strengths: Apache 2.0, vision, 262K context (extendable toward 1M via YaRN), fits a 24 GB GPU or a 24 GB Mac at 4-bit. Qwen says it outperforms the larger Qwen3.7-Plus overall and especially on real-world coding and office workflows ⚠️. Weaknesses: can't match the giants on the hardest tasks. Memory: 16-19 GB (4-bit); as little as 9-11 GB at 2-bit if you accept the quality hit. Gotcha: some community GGUFs ship a chat template that truncates multi-turn chats. Use a build with a corrected template ⚠️. Qwen3.8-2.4T-A95B: the second-largest open model Strengths: frontier-scale, Alibaba's first Max-class open release, independent index score 40. Weaknesses: huge; custom license with a $50M revenue condition (Section 7); the open version is text-only and thinking-always-on, unlike the hosted Max. Memory: 397 GB (1-bit XXXS, needs ~450 GB RAM) to 4.9 TB (BF16). DeepSeek V4 family: best value and the cleanest license V4-Flash-0731: 284B / 13B, the most runnable \"big\" model (83-162 GB). The July 31 build is the official release that replaced the April preview, and it was a large jump on agentic coding. V4-Pro-0813: 1.6T / 49B, needs ~500-870 GB. Its Artificial Analysis score (36) is the lowest of the leaders, and its API now routes to V4.1-Flash. V4.1-Flash (Sept 10): new Causal Encoder-Decoder architecture with vision, 552B backbone plus a 196B Engram memory store, 8B active on prefill and 16B on decode, 890 bytes per token of KV cache, trained on 45T multimodal tokens. DeepSeek says it beats V4-Pro on coding and agentic work ⚠️. Very cheap via API; hard to run locally. License: MIT across the board. Tencent Hy4 preview: the most permissive frontier-scale model Strengths: 770B / 49B active, Apache 2.0, 1M+ context, FP8 build shipped alongside BF16, mirrored on Hugging Face, ModelScope, GitCode and CNB. Weaknesses: it's a preview, text-only, needs an 8-GPU-class server (16× B200 or 8× B300 for the full model), no verified small build. Memory: ~770 GB (FP8). Curiosity: Tencent says Hy4 participated in its own development, autonomously optimizing its inference stack for a 31.8% throughput gain ⚠️. Gemma 4: the local-hardware family Sizes: E2B, E4B, 12B, 26B-A4B (MoE), 31B (dense). Apache 2.0, 140+ languages, up to 256K context. Multimodal across the board: text, image and video on every size; native audio on E2B, E4B and 12B. Sweet spot: 26B-A4B gives strong quality at ~18 GB and runs fast (only 3.8B active). Memory: 3 GB (E2B) to ~20 GB (31B, 4-bit). gpt-oss-120b / 20b: the Western permissive option (older) Strengths: Apache 2.0, well supported, 20b runs in 16 GB. Weaknesses: released August 2025, so it is a year behind the leaders. Also worth watching Atria Dawn Preview (Shanghai AI Lab): a 744B MoE agentic preview, ranked #10 overall on llm-stats at 52.3 ⚠️. MiniMax M3 / H3: reported at 428B and ~465B (omni-modal) sizes; M3 now appears in Artificial Analysis's tracked models, but I found only single-source detail on the specs, so they stay out of the tables ⚠️. 10. Verification log: where sources disagreed Question What I found What I used Kimi K3 active parameters 104B (Moonshot docs, Unsloth, OpenRouter). One table said ~50B; older Kimi models had 32B. 104B Kimi K3 full weight size 1.56 TB (Unsloth; Simon Willison; Hugging Face). One guide called 594 GB the \"full\" download. 1.56 TB. 594 GB is a 1-bit build. Kimi K3 weights date July 26 (US evening) vs July 27 July 27 — same event, time zones GLM-5.3 total size 744B (Z.ai, Unsloth, most coverage) vs ~753B (the BF16 Hugging Face repo and some outlets) ~744-753B, likely counting differences (MTP layer, embeddings) GLM-5.3 2-bit quality The New Stack: ~86% top-1. Unsloth's own table: 78.5% (IQ2_M) / 80.9% (Q2_K_XL). Unsloth's numbers GLM-5.3 license restriction Only hosting providers above $10B revenue need a security review As stated in Section 7 Qwen3.8-2.4T license One outlet described the whole Qwen3.8 release as Apache 2.0. The actual repo ships a separate Qwen3.8-Max License with a $50M MaaS threshold. Qwen3.8-Max License — the Apache 2.0 claim applies only to the 27B \"Run the 2.4T Qwen in 17 GB of RAM\" One tracker claimed this. Unsloth's table says 397 GB minimum; 17 GB is the 27B figure. Ignored the claim Qwen3.8-27B minimum RAM One guide said 32 GB is the practical minimum for 4-bit. Unsloth says 4-bit runs on 16-19 GB of VRAM or a 24 GB Mac. Unsloth's figures Qwen3.8-Max: open or not? The downloadable release is the 2.4T-A95B build (text-only, thinking-on); the hosted \"Max\" adds vision, optional thinking and tools Treated as two related but different things AA index: 60 vs 45 Different index versions (v4.2-era vs v4.3) Used the latest (v4.3) Kimi K3 GDPval-AA v2 Elo AA: 1,668. Z.ai's table: 1,682. Reported both DeepSeek V4.1-Flash size 552B (backbone) vs 763B (backbone + Engram) Reported both DeepSeek V4-Flash release date April 23, 2026 (preview) vs July 31, 2026 (0731) July 31 is the official build; April was a preview GLM-5.3-Flash 4-bit memory Unsloth 162-210 GB vs a community MLX build at ~224 GB Different builds; gave the range Gemma 4 release date \"2026\" in most coverage; Mar 31 announcement / early April weights, 12B added June, all checkpoints refreshed July 15 Gave the full timeline \"Best open model\" lists naming Llama 4 Several SEO-style pages still lead with Llama 4 or 2025 models Treated as out of date; rankings show it well behind What I could not verify Independent replication of Z.ai's or Tencent's benchmark tables. Claude Opus 5's reported GDPval-AA v2 Elo (~1,855) — I could not confirm this figure in Artificial Analysis's published materials, so it has been removed from Section 6C. Tencent's reported Terminal-Bench 85.4 and DeepSWE 64.3 for Hy4 — I found different vendor figures (GPQA Diamond 92.3, SWE-bench Multilingual 82.9) and could not reconcile them, so the original numbers were dropped. Real-world speed (tokens/sec) on your specific hardware; only a few community measurements exist. Exact memory for very long contexts on Kimi K3, Hy4 and GLM-5.3. License text itself: I relied on published license files, model cards and reputable reporting. Read the LICENSE file for anything commercial. Whether any U.S. restriction on Chinese open-weight models will be enacted. 11. Decision guide \"I just want the best open model and can use an API.\" → GLM-5.3 (cheaper) or Kimi K3 (more accurate on facts, better front-end). \"I need it cheap, at volume.\" → GLM-5.3-Flash or DeepSeek V4.1-Flash. \"I need it private, on my own hardware, with a decent budget.\" → 256 GB Mac: GLM-5.3 (2-bit). 128 GB Mac: DeepSeek V4-Flash-0731 (3-bit). \"I have a normal PC or laptop.\" → Qwen3.8-27B (24 GB GPU or 24 GB Mac at 4-bit) or Gemma 4 26B-A4B (16-24 GB). \"I need the safest license for a product.\" → Qwen3.8-27B, Gemma 4, Hy4 (Apache 2.0) or DeepSeek, GLM-5.3-Flash (MIT). Note that Qwen's large model is not Apache. \"I'm building on a small device or phone.\" → Gemma 4 E2B / E4B. Always: test two or three candidates on your own tasks, at the reasoning effort you actually plan to run. Rankings here disagree with each other, and the differences at the top are small. 12. Sources Independent trackers and leaderboards Artificial Analysis, Intelligence Index v4.3: https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v4-3 Artificial Analysis, model comparisons: https://artificialanalysis.ai/models/comparisons/glm-5-3-vs-kimi-k3 Artificial Analysis, Kimi K3: https://artificialanalysis.ai/models/kimi-k3 llm-stats leaderboard: https://llm-stats.com/ BenchLM open-weights ranking: https://benchlm.ai/best/open-source Model makers and model cards GLM-5.3 (Hugging Face): https://huggingface.co/zai-org/GLM-5.3 Kimi K3 license: https://huggingface.co/moonshotai/Kimi-K3/raw/main/LICENSE Kimi K3 (GitHub): https://github.com/MoonshotAI/Kimi-K3 Hy4 preview (Hugging Face): https://huggingface.co/tencent/Hy4-preview Qwen3.8 repository: https://github.com/QwenLM/Qwen3.8 Gemma 4 (Hugging Face): https://huggingface.co/google/gemma-4-31B Gemma releases index: https://ai.google.dev/gemma/docs/releases DeepSeek-V4.1-Flash on vLLM: https://recipes.vllm.ai/deepseek-ai/DeepSeek-V4.1-Flash gpt-oss announcement: https://openai.com/index/introducing-gpt-oss/ Memory / hardware guides Unsloth, GLM-5.3: https://unsloth.ai/docs/models/glm-5.3 Unsloth, Qwen3.8: https://unsloth.ai/docs/models/qwen3.8 Unsloth, DeepSeek V4: https://unsloth.ai/docs/models/deepseek-v4 Unsloth, Gemma 4: https://unsloth.ai/docs/models/gemma-4 Spheron, GLM-5.3 GPU guide: https://www.spheron.network/blog/deploy-glm-5-3-gpu-cloud/ News and analysis The New Stack, GLM-5.3 license: https://thenewstack.io/zai-glm-weights-license/ Simon Willison on Kimi K3's license: https://simonwillison.net/2026/Jul/27/kimi-k3/ KDnuggets, DeepSeek-V4.1-Flash architecture: https://www.kdnuggets.com/why-deepseek-v4-1-flash-is-such-an-exciting-open-model-release Apidog, V4.1-Flash GA and the V4-Pro reroute: https://apidog.com/blog/what-is-deepseek-v4-1-flash/ The Decoder, Qwen3.8: https://the-decoder.com/alibabas-qwen-team-releases-qwen-3-8-models-with-open-weights-under-the-apache-2-0-license/ TechNode Global, Tencent Hy4 preview: https://technode.global/2026/09/04/tencent-hy4-preview-open-source-model/ Axios, US policy on Chinese open models: https://www.axios.com/2026/07/20/ai-us-china-open-source-kimi TechCrunch, the open-weight debate: https://techcrunch.com/2026/07/20/openai-is-scared-of-open-weight-models-should-the-us-be/ 13. Changelog Revision 2 — September 20, 2026. Corrections applied after checking the first draft against primary sources: Qwen3.8-27B memory corrected. The full ladder is 7-8 GB (1-bit) / 9-11 (2-bit) / 12-14 (3-bit) / 16-19 (4-bit) / 23-26 (6-bit) / 31 (8-bit) / 56 (BF16). The previous \"11-13 GB (2-bit)\" and \"17-19 GB (4-bit)\" figures were wrong, and the \"32 GB practical minimum\" claim was removed — Unsloth states 4-bit runs on a 24 GB Mac. Qwen3.8-2.4T license resolved. It ships under a named Qwen3.8-Max License, not Apache 2.0 and not an open question: $50M Model-as-a-Service threshold plus a 100M MAU / $20M monthly attribution clause. GLM-5.3-Flash pricing corrected to Z.ai's list of $0.15 / $0.50, with the expired launch promo noted. DeepSeek V4-Pro reroute upgraded from single-source to confirmed: Sept 14, 04:00 UTC, model id deepseek-flash. Modalities corrected. GLM-5.3-Flash accepts video as well as images; Gemma 4 is natively multimodal across all sizes with audio on E2B / E4B / 12B, and its release timeline is now specific. Reasoning-effort caveat added to Section 6A — Artificial Analysis scores are per effort level, and Kimi K3 drops from 44 to roughly 30 at low effort. Cross-tracker warning added to Section 1: llm-stats, BenchLM and Artificial Analysis are not on the same scale. Policy section rewritten around the actual reporting (Axios, Politico, TechCrunch) and the contested attribution to specific labs. Two unverifiable figures removed — the Opus 5 GDPval Elo and Tencent's Terminal-Bench / DeepSWE numbers — and logged in Section 10. Smaller fixes: Engram is 196B (not ~197B), Kimi K3 weights landed July 27, DeepSeek V4-Flash-0731 is the official build replacing the April preview, and the GLM-5.3 quant table is labelled as file sizes rather than memory requirements. This guide reflects publicly available information as of September 20, 2026. It is not legal or financial advice; check each model's license before commercial use.","contentHash":"sha256:f1d0dbd62dbfcd6caf73affef3537f31811e2c1d963677e5fc15850f2b4591b9","authorName":"Anshu Pathak","authorUrl":"https://zyvop.com/author/anshu","authorSameAs":[],"category":null,"tags":["self-hosting","AI licensing","open-weight LLMs","llm-benchmarks","local-ai"],"audience":"Software engineers and developers building applications with self-hosting","tone":"Instructional, practical, code-first","readingTimeMinutes":24,"wordCount":5282,"faqs":null,"primaryTopic":"self-hosting","publishedAt":"2026-09-20T12:24:05.989Z","updatedAt":"2026-09-20T18:07:06.570Z","canonicalUrl":"https://zyvop.com/best-open-weight-llms-september-2026-d4ujd"}