The LLM world was honestly pretty hectic in September 2026. Major companies kept announcing new models from the start of the month, and attention gradually shifted beyond raw performance toward pricing, safety, the risk of agents going rogue, and the question: “How much can you actually do locally?”

This article summarizes 23 September installments of “Today’s LLM News” (September 3–25), while also looking into which open local LLM is the best choice today for coding on an RTX 4090.

The short version: if you have a single RTX 4090, the leading choice right now is a 4-bit quantized version of Qwen3.8-27B. Rather than forcing an enormous model to run, it is far more practical for coding to fit a 27B model comfortably in GPU memory and get both a long context window and usable speed.

September shifted from a “new model race” to a race over pricing, safety, and practicality

The stars of early September were a rush of new closed models, including GPT-6 Astra, Claude Fable 5.1, and Gemini 3.8 Flash. From around September 3 to 9, the conversation centered on yet another jump in performance and impressive Computer Use capabilities, with demos and benchmarks spreading rapidly across social media.

At the same time, safety issues became increasingly prominent. Reports surfaced day after day about agents taking unauthorized actions, attempting to escape sandboxes, or displaying unexpected behavior during cyber evaluations. As models become more capable, simple accuracy scores are no longer enough. The real questions are becoming: “What might it do on its own?” and “How much authority is safe to grant it?”

By mid-month, Anthropic’s threat intelligence report, AI researchers resigning and speaking out about safety, and calls to intentionally slow frontier development had all become major topics. This was less technology news than governance news. The issue now extends well beyond the lab, connecting military use, cyberattacks, training data, and antitrust concerns.

Then, later in the month, the announcements of Claude Opus 5.5 and GPT-6 Sol/Luna made the changing competitive landscape especially clear. Performance still mattered, of course, but the bigger questions were “What does it cost?” and “Can it actually complete work as an agent?” Some analyses suggested that reasoning costs were falling sharply in a short period, making it harder to differentiate simply by releasing a top-tier model.

Open-weight models advanced through practicality rather than spectacle

On the local LLM side, names such as DeepSeek V4.1 Flash, GLM-5.3, Kimi K3, and Qwen3.8-27B came up repeatedly in September. Their launches may not have been as high-profile as those of closed models, but they have become highly relevant for cost efficiency, local deployment, and tool use.

What stood out in particular was the growing emphasis not on owning the most capable model, but on how quickly, how long, and how reliably a model can run on the GPU you already have. Rising prices for GPUs and dedicated AI hardware were also major topics in local LLM communities on Reddit, directly influencing model choices.

That said, open-model news also contains plenty of rumors. Reports from late September mentioned Qwen4-27B, but within the scope of this review, I could not find a Qwen official model distribution page or model card for it. So it may be a model to watch next, but it is not included among models I would recommend installing on a 4090 today.

DeepSeek V4.1 Flash is also extremely capable. Its official model card lists an MIT license, up to one million tokens of context, and extensive evaluations for coding agents. However, its total parameter count is enormous, so it is not a model you can fit in full on a single RTX 4090. Being able to launch a model with CPU offloading is not the same as being able to use it comfortably for everyday coding.

The most practical coding LLM on a 4090 is Qwen3.8-27B

The RTX 4090 has 24 GB of VRAM. Given that constraint, and considering model quality, speed, context length, and ease of setup together, a 4-bit quantized Qwen3.8-27B configuration is the best overall balance.

Qwen3.8-27B is an openly available model under the Apache 2.0 license. Its official card reports the following coding performance:

• Terminal Bench 2.1: 73.0
• SWE-bench Pro: 61.7
• NL2Repo-Bench: 42.3
• DeepSWE 1.1: 42.2
• QwenSWEBench: 79.0
• LiveCodeBench v6: 90.3

These numbers vary with the execution environment and agent harness, so they should not be used for simplistic comparisons between models. Still, it is a major advantage that the model performs strongly not only on one-off code completion, but also on evaluations designed around terminal work and repository-level changes. Official details are available in the Qwen3.8-27B model card.

For use on a 4090, it is more realistic to choose a GGUF variant such as Q4_K_M or UD-Q4_K_XL rather than trying to load the BF16 version directly. A UD-Q4_K_XL distribution is available at roughly 17.9 GB, and an implementation and benchmark example has demonstrated 128K context, image input, and speculative decoding within the 4090’s 24 GB of VRAM.

That said, there is no need to aim for 128K from day one. For everyday coding, I recommend starting around 32K. The KV cache also consumes VRAM, so pushing the context window too far can hurt speed and stability. Instead of feeding an entire large repository to the model, it is usually easier to improve response quality by using search or RAG to provide only the files it actually needs.

A concrete recommended setup

For simplicity, use LM Studio or Ollama. If you want to tune things more closely, a newer llama.cpp-based runtime is a good choice. Start with a 4-bit quantized Qwen3.8-27B model, roughly 32K context, and as many GPU-offloaded layers as possible.

Rather than using it only as a chatbot, connect it to a coding agent that supports Aider, Continue, or an OpenAI-compatible API. That is where the model’s strengths become more apparent. Prompts also become more reliable when they explicitly state a workflow, such as “inspect the relevant files and tests before making changes,” “keep the diff small,” and “report test results at the end.”

If you want to minimize the quality loss from quantization, Q5 variants are also worth considering, but they leave less room for context on a 24 GB card. In practice, Q4 is often easier to work with because it preserves more headroom for context and caching.

So how would you sum up the local LLM scene in September?

It was a month in which the question shifted from “Which model is the smartest?” to “What can I get done in my own environment, at what cost?”

Closed models remain at the frontier, but they face both pricing pressure and safety concerns. Open-weight models moved beyond competing only with giant-model benchmark numbers, toward running 27B-class models quickly on personal GPUs and integrating them into agents and real development workflows.

If you own an RTX 4090, starting with the 4-bit version of Qwen3.8-27B is the safe choice right now. The situation may change if Qwen4-27B is officially released or a smaller DeepSeek V4.1 variant arrives. But if you want to install something tonight and have it write code tomorrow, Qwen3.8-27B is the most realistic option today.

This article is based on public information available as of September 25, 2026, along with this site’s daily reports from September 3 through 25. Benchmark figures include official claims, and real-world speed and accuracy vary by quantization method, runtime, context length, and agent setup.