WhatisthebestlocalLLMforcodingin2026?

Facts checked against model cards and the Ollama library on October 26, 2026. This field changes monthly, so check the cards before you download.

Which is the best local LLM for coding right now?

The best local LLM for coding in late 2026 comes from a short list: seven open-weight models cover almost every consumer setup, and four of them are Qwen releases. The table lists each with its size, license and the best coding score its own card reports. The scores come from different benchmarks and harnesses, so compare within a column, not across rows.

Download size is not the same as memory use. The context cache and runtime overhead come on top, so leave a few gigabytes of headroom. For the arithmetic, see local LLM hardware requirements; for what Q4 means, see GGUF quantization explained.

In the “active” column, a mixture-of-experts (MoE) model only runs a fraction of its weights per token, but all of them still have to sit in memory. That is why a 35B-A3B model needs roughly the memory of a 35B model.

What fits in 8GB, 12GB, 16GB, 24GB and 48GB+ of VRAM?

Developers on 16GB cards already show where the limit is. One Hacker News user running Qwen3.8 27B at a 2-bit quant on an RTX 5070 Ti with 16GB reported 128k context at 50-60 tokens/sec (HN, Aug 14, 2026), which is only possible by dropping below the default Q4 download. The rule that follows: pick the largest dense or MoE model whose download leaves a few gigabytes free. One recommendation per tier:

Two honest gaps. First, the 12GB row is a non-answer, because the model families in this table have a hole between 9B and 24B. Second, the two 3B-active MoE models at the 36GB+ tier don’t beat the dense 27B Qwen on the SWE-bench Pro numbers their cards report: 49.5 for Qwen3.6 35B-A3B and 47.6 for Laguna XS 2.1, versus 61.7 for Qwen3.8 27B. If you have 24GB, you may not need more memory.

A third gap is the 16GB row. Several community reports run Qwen3.8 27B on 16GB cards at Q2 or Q3 with long contexts (see “What do developers report in practice?” below), which this table does not list because those quants are not the default downloads. If you choose that route, expect lower quality than the Q4 builds the vendor scores describe.

What about Macs with unified memory?

On a Mac, the GPU and CPU share one memory pool, so the whole download must fit alongside macOS and your editor. Ollama publishes MLX builds, Apple’s native format, for the Qwen line: qwen3.8:27b-mlx is 18GB, the same size as the default build. Poolside states that Laguna XS 2.1 is “compact enough to run on a Mac with 36 GB of RAM.”

A reasonable mapping from the table above: 16GB Macs use the 9B tier, 24GB Macs are borderline for the 27B, 32-36GB Macs are comfortable with Qwen3.8 27B or Laguna XS 2.1, and 64GB Macs can load Qwen3-Coder-Next. Treat that as a starting point and confirm on your machine, because other applications compete for the same pool.

What do developers report in practice?

Community reports on Hacker News say Qwen3.8 27B is usable on 16GB cards, but only at 2- and 3-bit quants, and that a smaller MoE model often matches it for agent work. These are single-user reports on unspecified prompts, not controlled tests, and none has been independently verified. They are useful because they state hardware and quant, which the model cards don’t.

Three patterns repeat. First, a 16GB card runs the 27B only below Q4. One commenter argues that Q4 variants leave too little room for a useful context on 16GB (HN). Second, quality at low bit depth is a judgment call: one user calls a 2-bit quant much better than the previous generation, another finds the drop from Q4 noticeable even at Q2 versus Q3. Third, the 3B-active MoE is a real alternative. One user running agentic work on systems under 24GB VRAM prefers Qwen3.6 35B-A3B, with experts offloaded to system RAM, and reports that the dense 27B gave comparable results while taking about 1.2 to 2 times longer per task, because it runs more test attempts and compactions (HN). That fits the benchmark gap noted above.

Reliability is the other recurring complaint. A user who tried Qwen3.6-35B-A3B and Qwen3.8-Flash-Next in an agent harness said they could complete tasks but failed often enough to be unusable, and asked how others route work between local and hosted models (HN). A more positive account describes leaving Qwen3.8 27B running an agent loop at the end of the day and keeping the useful pieces (HN). The practical reading: treat a local agent as an overnight draft generator you review, not as an unattended committer.

Which benchmarks matter for coding, and what do they miss?

Prefer SWE-bench Pro and Terminal-Bench over SWE-bench Verified, and check which harness produced the number. Among the models whose cards report SWE-bench Pro, the dense 27B leads:

The scores you will see most:

• SWE-bench Verified: real GitHub issues, 500 tasks. OpenAI reported that it stopped using the benchmark and, per AI Weekly’s summary, that an audit of 138 of the hardest tasks found 59.4% had material issues in test design or problem descriptions. Qwen3.8’s card doesn’t list it and leads with SWE-bench Pro instead.
• SWE-bench Pro: a harder, multi-repository successor from Scale AI. It is not settled either: the same article notes a May 2026 audit claimed its graders mis-graded about a third of trials, which Scale AI had not confirmed.
• LiveCodeBench: competitive-programming style problems, which test algorithms more than repository work. A high score doesn’t mean a model can navigate your codebase.
• Aider polyglot: a useful edit-focused test, but Aider’s leaderboard was last updated on November 20, 2025. It lists DeepSeek-V3.2-Exp at 74.2% as its best open-weight entry and none of the models in this article.

The bigger problem is the harness. Qwen’s card says its numbers use specific settings, such as the Claude Code harness at temperature 1.0. A score is a property of the model plus the scaffold around it, and your editor plugin is a different scaffold. Run your own repository through any shortlisted model before committing.

Agentic coding vs autocomplete: do you need different models?

Yes, in practice. Autocomplete needs low latency on short contexts, so small and fast models win. Agentic work, where a model reads files, runs commands and edits across a repository, needs long context and reliable tool calls.

The model cards line up with this split. Devstral Small 2 ships alongside Mistral’s Vibe CLI. Qwen3-Coder-Next and Laguna XS 2.1 are pitched at agents, with Poolside describing Laguna as built for agentic coding on local hardware. For autocomplete, the 9B tier is the realistic choice because the download leaves memory free for an editor. The cards report no vendor tokens-per-second figures, so this article gives none. For Qwen-specific setup, see the Qwen Coder local guide.

How do local models compare with hosted frontier models?

Local models have closed much of the gap on published agentic benchmarks, but not all of it. Qwen’s own table for Qwen3.8 27B lists Claude Opus 4.6 (Max) at 78.2 on Terminal-Bench 2.1 against 73.0 for the local model, and 53.4 on SWE-bench Pro against 61.7. Those are the vendor’s numbers, run by the vendor. The trade-offs that decide the choice are privacy, cost at volume and refusal behavior rather than a single score. See local LLM vs Claude for coding for the full comparison. If your concern is over-refusal on security tooling, how to run an uncensored LLM locally covers that separately; the license terms of the base model still apply either way.

FAQ