Which single-GPU open-weight model is best in September 2026?

Qwen3.6-27B is the strongest general default on a 24 GB card — best all-around quality and agentic coding among models that fit. Xing4.0-29B-A4B, released September 22, 2026, is the better choice specifically for long-context agent work: 256K native context extensible to 512K, Apache 2.0, and ~15 GB at low-bit quantization per China Telecom's own optimization claim. gpt-oss-20b is the lightest at roughly 12 GB in Q4_K_M, which makes it the only one of the three comfortable on a 16 GB card with real context headroom.

How much VRAM does each model need?

At 4-bit quantization, before KV cache: gpt-oss-20b about 12 GB (21B total, ~3.6B active), Xing4.0-29B-A4B about 15-18 GB (29B total, 4B active; China Telecom claims ~15 GB with its memory optimizations, versus ~62.9 GB unquantized), and Qwen3.6-27B roughly 16-18 GB as a dense 27B. The rule of thumb is about 0.6 GB per billion total parameters at 4-bit. Mixture-of-experts models save compute, not memory — you hold every parameter and compute only the active ones.

Does mixture-of-experts help or hurt on a single GPU?

It helps speed and hurts memory efficiency per unit of quality. Xing4.0-29B-A4B activates 4B of 29B parameters per token, so it generates at roughly 4B-dense speed while reasoning closer to a much larger model — but you still need all 29B resident. A dense 27B like Qwen3.6-27B uses similar memory and runs slower per token. On a memory-constrained card MoE is usually the better trade, which is why nearly every locally deployable agent model shipped in 2026 is MoE.

Which has the longest context on consumer hardware?

Xing4.0-29B-A4B on paper, at 256K native and 512K extensible, against gpt-oss-20b's and Qwen3.6-27B's shorter windows. In practice the binding constraint on a 24 GB card is the KV cache, not the architecture. Expect tens of thousands of tokens of usable context at 4-bit on 24 GB regardless of which model you pick. The 256K figure is real on server hardware and aspirational on a desktop.

Can I run MiMo-V2.6-Flash on a single consumer GPU?

No, and this is the most common 2026 sizing mistake. MiMo-V2.6-Flash is roughly 309-310B total parameters with about 15B active. The active count makes it fast; the total count sets the memory requirement, which is datacenter-scale even at aggressive quantization. Active parameters determine speed, total parameters determine whether it fits. Xiaomi's MiMo line is an API and multi-GPU-server model, not a desktop model.

Quick Answer

Xing4.0-29B-A4B vs Qwen3.6-27B vs gpt-oss-20b (Sep 2026)

Published: September 26, 2026

The short answer

Qwen3.6-27B is the strongest default on one 24 GB card; Xing4.0-29B-A4B is the long-context agent specialist; gpt-oss-20b is the lightest and the only comfortable fit on 16 GB. China Telecom’s Xing4.0-29B-A4B, released September 22, 2026 under Apache 2.0, is the newest entrant and the only one of the three designed explicitly as an execution agent rather than a general assistant, with 256K native context extensible to 512K and FP8 and GGUF builds shipped alongside the base weights.

Side by side

	Xing4.0-29B-A4B	Qwen3.6-27B	gpt-oss-20b
Vendor	China Telecom AI (TeleAI)	Alibaba	OpenAI
Released	September 22, 2026	2026	2025, still current in the tier
Architecture	MoE	Dense	MoE
Total parameters	29B	27B	21B
Active per token	4B	27B	~3.6B
Native context	256K (→512K)	Long, below 256K	Below 256K
VRAM, 4-bit (pre-KV)	~15-18 GB	~16-18 GB	~12 GB
VRAM, unquantized	~62.9 GB	~54 GB	~42 GB
Fits 16 GB card	Tight	No	Yes
Fits 24 GB card	Yes	Yes	Yes, with large KV cache
License	Apache 2.0	Permissive (Qwen license)	Permissive (Apache 2.0)
Official GGUF	Yes	Community + official	Yes
Official FP8	Yes	Yes	Yes
Trained on	Huawei Ascend NPU, end to end	Nvidia	Nvidia
Designed as	Agentic execution	General + agentic coding	General, efficient
Independent AA index placement	None as of Sep 26, 2026	Yes	Yes

The VRAM arithmetic that decides this

Everything else is secondary to whether it fits. The working rule at 4-bit quantization is ~0.6 GB per billion total parameters, then add KV cache on top.

Card	gpt-oss-20b	Xing4.0-29B-A4B	Qwen3.6-27B
12 GB (RTX 3060, 4070)	Marginal	No	No
16 GB (4060 Ti 16GB, 5060 Ti)	Comfortable	Tight, short context	No
24 GB (3090, 4090)	Comfortable, big KV	Comfortable	Comfortable
32 GB (5090)	Trivial	Comfortable, longer KV	Comfortable
32-64 GB Apple unified	Yes via GGUF	Yes via GGUF	Yes via GGUF

Two traps here.

MoE does not save memory. Xing4.0-29B-A4B activates 4B parameters per token, which is why it feels fast. All 29B still sit in VRAM. People read “4B active” and size a 12 GB card for it; it does not fit.

KV cache eats the headroom. The 256K context on Xing4.0-29B-A4B is an architectural property. On a 24 GB card at 4-bit with ~16 GB of weights, the remaining ~8 GB is your working context, which is tens of thousands of tokens, not 256,000. Size context to the card and raise it until you OOM, rather than requesting the headline number.

Where each one wins

Qwen3.6-27B is the strongest single default for a 24 GB GPU in 2026. It is the best all-around model in the tier and the best at agentic coding among models that actually fit, with mature tooling, Thinking and Non-Thinking modes, and the broadest quantization ecosystem. If you are going to run exactly one local model and you have 24 GB, run this. Being dense is the cost: it generates noticeably slower than the MoE options at the same memory footprint.

Xing4.0-29B-A4B wins long-context agent work and licensing purity. Apache 2.0 with no field-of-use restrictions, 256K native context, FP8 and GGUF published on day one, and a design brief that is explicitly execution rather than conversation — multi-step planning, tool calling, long documents. China Telecom publishes SWE-bench Verified and Multilingual via the SWE-agent harness at 210K context, Terminal-Bench 2.1 via terminus-2, and Claw-Eval at 256K, with full sampling settings disclosed, which is more methodological honesty than most releases. Third-party write-ups report coding results ahead of Gemma 4 and Qwen3.6 on several benchmarks — treat that as unverified until an independent evaluator publishes, because as of September 26, 2026 it has no Artificial Analysis placement. Full detail in what is Xing4.0-29B-A4B.

gpt-oss-20b wins on hardware reach. At roughly 12 GB in Q4_K_M it is the only one of the three that leaves real headroom on a 16 GB card, and it is trivially comfortable on 24 GB with a large KV cache. That matters more than benchmark deltas if your hardware is a laptop, a 4060 Ti, or a shared machine. It is also the safest choice for a fleet where you cannot guarantee every machine has 24 GB.

Benchmarks: why you should not pick on them

All three publish agentic coding scores. None of them are comparable to each other, because agent benchmark results are a function of the harness as much as the model. China Telecom’s SWE-bench numbers use SWE-agent at 210K context, temperature 1.0, top_p 0.95, repetition penalty 1.05. Change the harness and the number moves by double digits. See agent harness vs model: why scores differ.

The defensible evaluation for this tier is to run your own three tasks — one repository edit, one long-document extraction, one multi-step tool-calling workflow — against all three at the quantization you will actually deploy. That takes an afternoon and beats any published table, because quantization level alone can swing quality more than the model choice does.

The Ascend footnote

Xing4.0-29B-A4B was trained end to end on Huawei Ascend NPUs. For a Western builder this changes nothing operationally — the weights run on Nvidia, Apple Silicon or CPU like any other GGUF. It matters as a signal: the Chinese domestic training stack has produced a competitive agentic model without Nvidia silicon, which removes the last argument that Ascend is inference-only. Combined with Huawei’s Atlas 960E SuperPod, the domestic pipeline now covers both ends.

Decision rule

24 GB card, one model, general use: Qwen3.6-27B.
24 GB card, building an agent that reads long documents or works across a large repo: Xing4.0-29B-A4B.
16 GB card, or a fleet with mixed hardware: gpt-oss-20b.
Apache 2.0 required with no ambiguity: Xing4.0-29B-A4B or gpt-oss-20b.
You were considering MiMo-V2.6-Flash for a desktop: you cannot. It is ~309B total parameters; the ~15B active count is speed, not size. Use the API or a multi-GPU server.
All three: deploy at the quantization you will run in production before you benchmark. A Q4 model and a Q8 model of the same weights are different products.

The broader field, including the models that did not make this three-way, is ranked in best open-weight models for a single GPU.

Last verified: September 26, 2026. VRAM figures are 4-bit estimates before KV cache and vary by quantization method.