Verdict: Qwen3.8-27B is the best local LLM for coding for most developers with a single 24 GB GPU. It runs at 4-bit quantisation on that hardware, ships under Apache 2.0, and is the only one of the three that pairs agentic coding scores with weights you can actually download and serve yourself (model card). GLM-5.3 is the better pick only if you have a workstation cluster or multi-GPU server to host a large mixture-of-experts model. Gemma 4 26B A4B is the choice for sub-16 GB machines and latency-sensitive autocomplete, not for long agentic tasks.
TL;DR
- Qwen3.8-27B (Alibaba, released 14 August 2026, Apache 2.0): dense 27B, 262,144-token native context, fits a 24 GB GPU at 4-bit, roughly 55.6 GB in BF16 (Hugging Face).
- Vendor-reported scores put it ahead of Claude Opus 4.6 Max on 16 of 24 model-card benchmarks, and behind on GPQA Diamond and Humanity's Last Exam (model card).
- GLM-5.3 (Z.ai, 18 August 2026) scores Coding 75 on the Artificial Analysis index, but the GLM-5.2 generation is a 753B MoE, which puts self-hosting outside consumer reach (Artificial Analysis).
- Gemma 4 26B A4B (Google, Apache 2.0) activates about 3.8B parameters per token, giving the best capability-per-gigabyte at small tiers, but it failed a real agentic coding task in an independent field test, including a hard crash (EPAM).
- If you already run last generation's Qwen3.6-27B, the upgrade is incremental rather than urgent.
- Last verified: 26 August 2026.
Which is the best local LLM for coding on a single GPU?
Qwen3.8-27B, by a clear margin, provided you have 24 GB of VRAM. Alibaba's model card reports Terminal-Bench 2.1 at 73.0 (up from 63.4 for Qwen3.6-27B), SWE-bench Pro at 61.7 (from 53.5), LiveCodeBench v6 at 90.3 (from 83.9), OSWorld-Verified at 84.3 and IFBench at 79.5 (model card). Two caveats matter. All of those are self-reported, and the SWE-bench Pro run used the Claude Code harness per a model-card footnote, so the number describes a model-plus-scaffold pairing rather than the raw model. Treat them as a vendor's best-case claim, not as a neutral measurement.
The one independent signal so far is encouraging without being a coding result: Qwen3.8-27B placed first among open-weight models on Harvey's Legal Agent benchmark, in a study run by Harvey with Engram and reported in mid-August 2026. Long-horizon legal agent work rewards the same instruction-following and tool-use discipline that agentic coding does, which is why it is worth mentioning at all.
Architecturally it is a hybrid: 48 Gated DeltaNet linear-attention layers interleaved with 16 full Gated Attention layers, plus a vision encoder that brings the total to about 28B. The practical consequence is cheaper long-context inference than a pure dense attention stack of the same size, which is what makes a 262,144-token window usable rather than theoretical on one card. YaRN extension pushes that to 1M tokens, at the usual cost to precision.
For step-by-step hardware, quantisation and serving choices, see our Qwen3.8-27B local setup guide.
When is GLM-5.3 the better choice than Qwen3.8-27B?
When "local" means a rack, not a desktop. GLM-5.3 arrived from Z.ai on 18 August 2026 with an Artificial Analysis Intelligence score of 60 and a Coding score of 75, which is competitive at the frontier tier (Artificial Analysis). The problem is the shape of the model rather than its quality: the GLM-5.2 generation is a 753B-parameter mixture of experts, and nothing in that class fits a consumer GPU at any usable quantisation.
So the honest split is this. If you are provisioning a shared inference node for a team, with enough aggregate VRAM to hold a large MoE, GLM-5.3 is a reasonable candidate and may edge Qwen on some coding work. If you are one developer with one GPU, it is not an option, and the comparison ends there. Our earlier head-to-head on Qwen3-Coder versus GLM-5.2 covers that lineage in more detail.
Where does Gemma 4 26B A4B actually win?
At the low end of the VRAM curve, and in autocomplete rather than agentic work. Gemma 4 26B A4B is a sparse MoE that activates roughly 3.8B parameters per token under Apache 2.0, which gives it strong capability per gigabyte and low per-token latency. If your machine has 8-16 GB of VRAM, or you want inline completions that return fast enough not to interrupt typing, it is the sensible default.
It is weaker as an agent. In an independent field test published on 21 July 2026, the QAT build of Gemma 4 26B A4B running locally failed every assigned task and hard-crashed, while Qwen3.6-35B-A3B was the only local model that completed a realistic settings-screen task on a 15,000-line Flutter application (EPAM). That is a single test on one codebase, so do not over-read it, but it matches the general pattern: heavily sparse small models handle short, well-scoped generation better than multi-step tool use. If Gemma is your tier, our Gemma 4 QAT guide covers the quantisation-aware builds.
What does each model cost you in VRAM?
| Model | Params | Local viability | Licence |
|---|---|---|---|
| Qwen3.8-27B | Dense 27B (28B with vision) | 24 GB GPU at 4-bit; about 55.6 GB in BF16 | Apache 2.0 |
| GLM-5.3 | Large MoE (GLM-5.2 was 753B) | Multi-GPU server or cluster | Open weights |
| Gemma 4 26B A4B | 26B total, about 3.8B active | Sub-16 GB tiers, fast autocomplete | Apache 2.0 |
| Kimi K3 | 2.8T MoE | 594 GB download, 4x A100 class | Open weights |
Figures for Qwen3.8-27B are from its model card; Kimi K3's weights were published on 26 July 2026 and are included to mark the boundary of what "local" can mean. It is a server model, not a desktop one. We wrote separately about using it inside an agent stack in our Kimi K3 agent guide.
Should you upgrade from Qwen3.6-27B?
Not urgently. Qwen3.6-27B, released 22 April 2026, scored 77.2% on SWE-bench Verified, ahead of the 397B flagship of its own generation at 76.2% (willitrun.ai), and its Q4_K_M quantisation lands around 16.8 GB, which fits RTX 4080-class hardware. If that setup is working, the gains in Qwen3.8-27B are concentrated in agentic and terminal tasks rather than single-file generation. The stronger reasons to move are the longer native context and the vision input.
If you are starting from nothing, begin with our local AI setup guide, then pick the tier that matches your card.
How should you choose in practice?
Match the model to your hardware first, then to your workflow. One 24 GB GPU and agentic coding: Qwen3.8-27B at 4-bit. Under 16 GB, or autocomplete-heavy work: Gemma 4 26B A4B. A shared multi-GPU node and a team to justify it: evaluate GLM-5.3 against Qwen on your own repositories.
That last point is the part most comparisons skip. Every published number here is either vendor-reported or drawn from a single independent test, and none of them were measured on your codebase, your language, or your tooling. Run a fixed set of five to ten real tasks from your own backlog through each candidate and count completions. That local measurement will beat any leaderboard for deciding what to install.
FAQ
Q: What is the best local LLM for coding in 2026?
A: Qwen3.8-27B for a single 24 GB GPU, and Gemma 4 26B A4B for smaller rigs or fast autocomplete.
Q: Can a 24 GB GPU run Qwen3.8-27B?
A: Yes, at 4-bit quantisation. Full BF16 weights need around 55.6 GB, per the model card.
Q: Is Qwen3.8-27B open source?
A: It is open-weight under Apache 2.0, which permits commercial use, though the training data is not released.
Q: Can you run Kimi K3 locally?
A: Only on multi-GPU server hardware. The weights are a 594 GB download and were benchmarked on four A100 GPUs (explainx.ai).
Q: Are Qwen3.8-27B's benchmark numbers independent?
A: Mostly no. The coding and agentic scores are self-reported by Alibaba; the one independent result is a first-place open-weight finish on Harvey's Legal Agent benchmark.
Q: Why does Gemma 4 score well on efficiency but poorly as a coding agent?
A: It activates only about 3.8B parameters per token, which helps latency and memory but leaves less capacity for the multi-step tool use that agentic coding requires.
Top comments (0)