Best Open-Weight LLMs to Run Locally in September 2026: Qwen3.8-27B, DeepSeek V4.1-Flash, MiMo-V2.6 and gpt-oss

Which open-weight model to run on a 16 GB laptop, a 24 GB GPU or a multi-GPU server in September 2026: Qwen3.8-27B, DeepSeek V4.1-Flash, MiMo-V2.6 and gpt-oss, with memory needs, licenses, benchmarks and launch commands.

By AI Mastermind Lab · Published 25 Sept 2026

TL;DR: The best open-weight model you can run on one consumer GPU right now is Qwen3.8-27B (Apache 2.0, 262K context, native vision, roughly 14–16 GB at 4-bit). If you have a multi-GPU server, DeepSeek V4.1-Flash (552B mixture-of-experts, 8B active parameters for input and 16B for output) is the strongest open model DeepSeek has released and beats its own larger V4-Pro. Xiaomi's MiMo-V2.6 tops the Artificial Analysis open-weight index, and OpenAI's gpt-oss-20b remains the easiest 16 GB option for pure text. This guide covers what each needs, what it is good at, and how to start it.

Six weeks of open-weight releases have changed the local-LLM picture more than the previous six months did. Alibaba open-sourced Qwen3.8 in mid-August, DeepSeek shipped V4.1-Flash on 10 September, Xiaomi released MiMo-V2.6 on 22 September, and Z.ai, Moonshot and Poolside all published weights over the summer. This guide is organised by the hardware you have, because that decides more than any benchmark. Facts are taken from each model's card or announcement, linked inline; where only secondary reporting exists, that is stated.

Which model for which hardware

Hardware Recommended model Approximate memory License
16 GB laptop or single consumer GPU gpt-oss-20b, or Qwen3.8-27B at 4-bit ~16 GB Apache 2.0
24 GB GPU or 32 GB Apple Silicon Qwen3.8-27B at 4–8-bit 14–27 GB Apache 2.0
Multi-GPU server (8x 80 GB or more) DeepSeek V4.1-Flash Hundreds of GB total, but only 8–16B parameters active per token See model card
Anything, via a hosted API DeepSeek V4.1-Flash (deepseek-flash) — $0.30 / $1.20 per 1M tokens peak, half off-peak

Qwen3.8-27B: the single-GPU champion

Alibaba's Qwen3.8-27B model card describes a 27-billion-parameter causal language model with a vision encoder, 64 layers, and a hybrid attention design that mixes Gated DeltaNet with Gated Attention. The native context length is 262,144 tokens, extensible to 1,000,000. It is released under Apache 2.0, which is the license most enterprises can accept without a legal review, and it handles text, images and video input.

The published benchmarks are the reason it leads this list. On the model card: SWE-bench Pro 61.7, Terminal Bench 2.1 (Terminus) 73.0, LiveCodeBench v6 90.3, GPQA Diamond 89.2, IFBench 79.5, and, for computer use, OSWorld 84.3. Those coding and agentic scores are in the range that closed frontier models were posting in the spring, from a model that fits on one 24 GB card at 4-bit. VentureBeat's coverage makes the same point: frontier-class coding agents with no cloud API.

Running it:

# vLLM (the model card's recommended server)
vllm serve "Qwen/Qwen3.8-27B"

# SGLang
python3 -m sglang.launch_server --model-path "Qwen/Qwen3.8-27B" --host 0.0.0.0 --port 30000

For laptops, the card lists llama.cpp, Ollama and LM Studio support through quantised community builds, of which Hugging Face lists more than a thousand. Pick a Q4_K_M build for 16–24 GB machines and a Q8 build if you have 32 GB or more. The recommended sampling settings differ by mode: temperature 1.0 and top-p 0.95 when thinking is enabled, temperature 0.7 and top-p 0.80 for plain instruct use. Reasoning effort is adjustable (xhigh, medium, low), which is worth turning down for interactive chat and up for agent runs.

If you are new to local models, our Ollama on Mac guide and the Ollama vs LM Studio comparison cover setup; the vLLM quick-start covers serving on a GPU box.

DeepSeek V4.1-Flash: the strongest open model, if you have the hardware

DeepSeek's V4.1-Flash announcement (10 September 2026) describes a 552-billion-parameter mixture-of-experts model that activates 8B parameters when processing input and 16B when generating output, with native visual understanding. DeepSeek says tests by multiple parties put it ahead of its flagship V4-Pro on performance, cost, speed and total runtime, and it has begun routing V4-Pro API requests to V4.1-Flash. The weights are published on Hugging Face; check the repository for the license text before commercial use.

The catch is memory. Even though only 8–16B parameters are active per token, all 552B must be resident, so this is a multi-GPU deployment. For everyone else, the hosted API is the practical route: the pricing page lists deepseek-flash at $0.30 per million input tokens (cache miss) and $1.20 per million output at peak, with a 1M context, 384K maximum output and a cache-hit input price of $0.006. Off-peak, outside 01:00–04:00 and 06:00–10:00 UTC on weekdays, everything is half price. That makes it the cheapest way to get near-frontier quality without owning hardware, and a good "large model in the loop" for a local Qwen agent.

MiMo-V2.6: Xiaomi's surprise leader

Xiaomi open-sourced MiMo-V2.6 on 22 September 2026 in Pro, Flash and a distilled 9B Qwen-based variant, according to TechNode. TechNode reports an Artificial Analysis intelligence index score of 46, ahead of Kimi K3 (44) and GLM-5.3 (45), and a stated training cost of $2.62 million for the Pro model. The license terms and parameter counts were not stated in that report; read the model card on Hugging Face before relying on either. The Distill-Qwen-9B variant is the one to try on a laptop if you want MiMo's reasoning style in a small footprint.

gpt-oss-20b: the reliable 16 GB option

OpenAI's gpt-oss models, released in August 2025, are still the simplest way to get a capable Apache 2.0 text model on a 16 GB machine: gpt-oss-20b was designed to run in 16 GB of memory, and the larger gpt-oss-120b targets a single 80 GB GPU. They are text-only and a year old, so Qwen3.8-27B beats them on coding and adds vision, but gpt-oss-20b remains a sensible default for pure-text assistants on constrained hardware and is available in Ollama and LM Studio with no configuration.

Also worth knowing

  • Qwen3.8-Max is the 2.4-trillion-parameter, 95B-active sibling of the 27B model, released under a custom license rather than Apache 2.0, per VentureBeat. It is a datacentre model; most teams will use it through an API.
  • Kimi K3 from Moonshot (2.8T parameters, 104B active, weights released 27 July, custom license) and GLM-5.3 from Z.ai (14 August, weights following) are the other large open releases of the summer, reported by Interconnects and MarkTechPost respectively. Both are server-class.
  • Poolside Laguna XS 2.1 (33B total, 3B active, OpenMDW-1.1 license, July) is a coding-focused small model worth testing if Qwen3.8-27B is too slow on your hardware; see Poolside's announcement.
  • Small language models are covered separately in our SLM guide, and Best Ollama models for coding has the earlier-2026 picks.

A practical local setup for September 2026

For a developer laptop with 32 GB of unified memory or a 24 GB GPU, the setup that gives the most capability per watt today:

  1. Qwen3.8-27B at Q4_K_M in Ollama or LM Studio as the daily driver: chat, code, and reading screenshots or PDFs through the vision encoder.
  2. Reasoning effort low for chat, xhigh for agent tasks so you are not paying thinking-token latency for simple questions.
  3. DeepSeek V4.1-Flash via API, off-peak, as the escalation model for tasks the 27B model gets wrong, at roughly $0.15 / $0.60 per million tokens.
  4. A local RAG stack (Ollama + ChromaDB + LangChain) so your documents never leave the machine.

Re-run your own evaluation set when you switch models. Model cards report the vendor's numbers under the vendor's settings; the quantised build on your laptop will be a little worse, and the difference varies by task.

Sources