Best Open-Weight LLMs, September 2026

The complete comparison: quality, RAM needed, cost, licenses (verified, in plain English)

Last checked: September 20, 2026. Revision 2 — 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

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

Two rules that prevent most misreadings of this guide:

1. 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.

2. 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.

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):

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

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

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)

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?"

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)

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.

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

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)

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

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

1. "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).

2. "I need it cheap, at volume." → GLM-5.3-Flash or DeepSeek V4.1-Flash.

3. "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).

4. "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).

5. "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.

6. "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:

1. 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.

2. 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.

3. GLM-5.3-Flash pricing corrected to Z.ai's list of $0.15 / $0.50, with the expired launch promo noted.

4. DeepSeek V4-Pro reroute upgraded from single-source to confirmed: Sept 14, 04:00 UTC, model id deepseek-flash.

5. 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.

6. 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.

7. Cross-tracker warning added to Section 1: llm-stats, BenchLM and Artificial Analysis are not on the same scale.

8. Policy section rewritten around the actual reporting (Axios, Politico, TechCrunch) and the contested attribution to specific labs.

9. Two unverifiable figures removed — the Opus 5 GDPval Elo and Tencent's Terminal-Bench / DeepSWE numbers — and logged in Section 10.

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.

Published via ZyVOP — Write once in Markdown, auto-backup to GitHub, and syndicate to Dev.to, Medium & Hashnode in 1 click.