Open-Source LLM Leaderboard for Self-Hosters: August 2026

By AIFoss · 10 min read
llmleaderboardselfhostedopensourceai

TL;DR: The models topping the August 2026 benchmark leaderboards — GLM-5.2, Kimi K3, Inkling — are datacenter hardware plays you will rent, not run. The models you can actually self-host got dramatically better this month: Qwen3.8-27B and Meta’s Muse Glimmer 30B both landed under clean Apache 2.0 and both fit a 24GB card. Benchmark rank and self-hostability are different leaderboards; this article is the second one.

Benchmark king24GB sweet spotBudget pick
Best forRenting via API or RunPodA single RTX 3090/4090 at home12–16GB cards
ModelGLM-5.2 (MIT, 744B MoE)Qwen3.8-27B / Muse Glimmer 30B (Apache 2.0)Gemma 4 12B or 26B-A4B QAT (Apache 2.0)
The catch~241GB at its smallest usable quantDense models — slower than the old MoE champQAT quants only; naive Q4 conversion loses accuracy

Honest take: If you have one 24GB card, run Qwen3.8-27B and stop reading leaderboards for a quarter. Everything above it on the charts needs hardware you don’t own.

Every month a new “best open-source LLM” ranking makes the rounds, and every month it commits the same sin: sorting by benchmark score as if a 744-billion-parameter MoE and a 27B dense model are choices the same person is making. For self-hosters they are not. This is the leaderboard rebuilt around the question that actually matters: what is the best model you can run on the hardware you have, under a license that lets you use it?

Everything below reflects the state of play as of late August 2026. Model churn in this space is brutal — the release cadence piece explains why — so treat the specific picks as dated the moment a new weight drop lands.

What changed in August 2026

Three releases this month reshuffled the deck, and two of them matter far more for home labs than the headline model does.

Qwen3.8-27B (Apache 2.0, weights Aug 14). Alibaba shipped the open weights on Hugging Face under plain Apache 2.0 — no MAU cap, no attribution clause, none of the Tongyi Qianwen license terms some expected. It’s a dense 27B with native image and video input and 262K context, about 56GB in BF16, which puts a Q4_K_M build in the ~16–17GB range: comfortable on a 24GB card with room for real context, and squeezable onto 20GB. Ollama shipped day-one support in v0.32.12. This is the new default answer at the enthusiast tier.

Meta Muse Glimmer 30B (Apache 2.0, weights Aug 10). After going closed with Muse Spark — covered in what self-hosters should do now — Meta’s first open release since Llama 4 arrived with a surprise: plain Apache 2.0, not a Llama-style community license with a 700M-MAU cap. It’s a dense 30B multimodal model tuned specifically for agent workloads, ships with its own speculative-decoding drafter, and Meta coordinated day-one support across llama.cpp, Ollama, MLX, and vLLM. A quantized build runs on a single 24GB GPU. Whether it beats Qwen3.8-27B on your workload is genuinely open — the agent tuning is real, but independent benchmarks are still thin.

GLM-5.3 (released Aug 14 — weights still pending). Z.ai’s follow-up to GLM-5.2 launched API-first, with open weights promised roughly two weeks after launch pending safety evaluation — Z.ai tied the delay partly to the model’s unusually strong cybersecurity and vulnerability-finding performance. As of this writing the full-size weights have not landed and the license is unconfirmed; MIT is the pattern from 5.1 and 5.2, but don’t assume it carries over. When the weights drop, expect the same story as 5.2: a datacenter model you rent, not run.

The license map, August 2026

Benchmarks change monthly; licenses change your legal exposure. The current map, each verified against the actual license file at review time:

ModelLicenseSelf-host verdict
Qwen3.8-27B, Qwen3.6-35B-A3BApache 2.0Clean — commercial use unrestricted
Muse Glimmer 30BApache 2.0Clean — no Llama-style MAU cap
Gemma 4 familyApache 2.0Clean (first Apache Gemma generation)
Inkling 975B / Inkling-Small 276BApache 2.0Clean
Codestral 2 (22B)Apache 2.0Clean
GLM-5.1 / GLM-5.2MITClean
DeepSeek V4-Flash / V4-ProMITClean
Kimi K2.7 / K3Modified MIT / “Kimi K3 License”Fine at home; attribution + revenue clauses above $20M/mo or 100M MAU
Llama 4 Scout / MaverickLlama Community License700M-MAU cap, attribution, EU multimodal restriction — not OSI
MiniMax M3MiniMax Community LicenseNon-commercial — check before any business use

The pattern that held all year still holds: the Chinese labs (Qwen, DeepSeek, Z.ai) and now Meta’s Superintelligence Labs ship the cleanest licenses, while “open” releases with asterisks cluster elsewhere. The full background is in the license shootout.

The benchmark leaderboard vs. the self-hosting leaderboard

Here’s the top of the charts by SWE-bench-class coding benchmarks, with the number that actually gates you — the smallest usable quantized footprint:

ModelHeadline scoreSmallest usable quantRuns on
Kimi K3 (2.8T MoE)88.3 Terminal-Bench 2.1 (vendor)~594GB (1-bit GGUF)Multi-node datacenter only
DeepSeek V4-Pro (1.6T MoE)80.6% SWE-bench Verified~800GB Q4Datacenter only
DeepSeek V4-Flash (284B MoE)79.0% SWE-bench (vendor)~96–103GB (1–3-bit)128GB-RAM server w/ GPU offload
Inkling 975B77.6% SWE-bench Verified (vendor)~270GB disk (1-bit)512GB DDR5 offload rig
Qwen3.6-35B-A3B73.4% SWE-bench Verified~21GB Q4_K_MOne 24GB card
GLM-5.2 (744B MoE)62.1% SWE-bench Pro (tops the chart)~241GB smallest usableCloud / multi-GPU

Read the last two rows together and you get the only insight this article needs to deliver: the gap between the datacenter king and the best single-card model is a handful of benchmark points, at roughly a thousand times the hardware cost. GLM-5.2 tops SWE-bench Pro; Qwen3.6-35B-A3B gets you within striking distance on a card you can buy used for $700. (SWE-bench Pro and SWE-bench Verified are different, harder-vs-easier suites — don’t compare the percentages directly. The point survives the caveat.)

Kimi K3 deserves its own asterisk: the largest open-weight release ever is still effectively unrunnable at home — mainline llama.cpp didn’t support its KDA architecture as of late August, so even the 594GB 1-bit GGUF needs the Unsloth fork and a ~610GB RAM+VRAM floor. The Kimi K3 self-hosting guide has the full hardware reality check.

The tier list: best model per hardware class

This is the part to bookmark. Picks weight license cleanliness, quant quality, and tooling support — not just benchmark rank.

8–12GB VRAM (RTX 3060, 4060): Gemma 4 12B in its QAT build (~6.6–7GB, gemma4:12b-it-qat on Ollama). Multimodal, 256K context, Apache 2.0. One trap: use the official QAT tags or Unsloth’s UD quants — naive Q4_0 conversion of a QAT checkpoint measurably loses accuracy, as covered in the Gemma 4 QAT guide.

16GB VRAM (RTX 4060 Ti 16GB, 4080): Gemma 4 26B-A4B QAT (~15GB) for general use; Codestral 2 at ~13GB Q4_K_M if the workload is code. Qwen3.8-27B squeaks in at aggressive quant levels, but you’ll sacrifice context headroom.

24GB VRAM (RTX 3090, 4090): This tier finally moved after months of stasis. Qwen3.8-27B Q4_K_M (~16–17GB) is the new default: dense, multimodal, 262K context, Apache 2.0, day-one Ollama support, and room left for real KV cache. Muse Glimmer 30B is the pick if your workload is agentic — tool-calling loops, computer use — since that’s literally what Meta tuned it for. The outgoing champ, Qwen3.6-35B-A3B (~21GB Q4_K_M), remains the throughput king: its 3B-active MoE design decodes far faster than either dense newcomer, so keep it if tokens-per-second is your constraint.

48GB (dual 3090s): Qwen3.6-72B-class dense models at Q4, or DeepSeek V4-Flash at 1-bit if you have 128GB+ system RAM to offload experts into — the DeepSeek V4 self-hosting guide covers the -ot expert-offload flags that make it usable.

96GB+ (RTX PRO 6000, Mac Studio 128GB): Inkling-Small 276B at 2-bit (~88GB) is the frontier-adjacent option; DeepSeek V4-Flash at 3-bit (~103GB) is the better-supported one. Above this you’re building a server, not a home lab — at which point renting a RunPod pod by the hour beats owning depreciating silicon for all but constant workloads. For choosing the card itself, runaihome.com’s GPU guides cover the hardware side; for wiring any of these models into Cursor, Cline, or Continue.dev as a coding backend, see aicoderscope.com.

When NOT to trust this leaderboard

Three honest caveats. First, most scores above are vendor-reported — Kimi K3, Inkling, and DeepSeek V4-Flash numbers have no independent replication yet, and vendor benchmarks systematically flatter. Second, benchmark rank predicts your workload poorly: Muse Glimmer will likely beat higher-scoring models on agent loops because it was tuned for them, and Qwen3.6’s MoE speed advantage matters more than 2 points of SWE-bench in interactive use. Third, this snapshot decays fast — GLM-5.3 weights are due within days of publication, and Inkling-Small’s coding benchmarks were still unpublished at review time. Run your own eval on your own tasks before committing hardware money to any of it.

FAQ

What’s the best open-source LLM I can run on a 24GB GPU in August 2026? Qwen3.8-27B (Apache 2.0) at Q4_K_M is the best all-rounder — multimodal, 262K context, ~16–17GB. Choose Muse Glimmer 30B for agent workloads, or stay on Qwen3.6-35B-A3B if decode speed matters more than raw quality.

Is GLM-5.2 worth self-hosting since it tops SWE-bench Pro? For almost everyone, no. Its smallest usable quant is ~241GB — that’s a multi-GPU server or a rented cloud pod. The MIT license is clean, but the practical way to use it is via API or an hourly RunPod rental, not owned hardware.

Which August 2026 model releases have genuinely clean licenses? Qwen3.8-27B and Muse Glimmer 30B both ship under plain Apache 2.0 with no user caps or attribution clauses. Kimi K3 is MIT-derived but adds revenue/MAU clauses (irrelevant at home), and GLM-5.3’s license was unconfirmed at publication because the weights hadn’t shipped yet.

Sources

  • RTX 3090 — the used-market 24GB workhorse for the Qwen3.8 / Muse Glimmer tier
  • RTX 4060 Ti 16GB — cheapest new card that comfortably runs Gemma 4 26B-A4B QAT

Was this article helpful?