Guide · 2026-08-02

Open-Weight LLM Releases — Q2 2026 Wrap-Up

Last updated 2026-08-02

The quarter that closed the gap: MiniMax M3, DeepSeek V4, GLM-5.2 and Gemma 4 turned Q2 2026 into the strongest open-weight window yet.

By Mohamed Meguedmi · 9 min read

The Q2 2026 story in one paragraph

Between April and June 2026 the open-weight field did something it had only threatened before: it caught the closed frontier on the workloads developers actually pay for. Coding, long-context retrieval and agentic tool use all saw open releases that matched or beat the previous generation's proprietary leaders. The quarter opened with a cluster of late-April drops (Xiaomi MiMo-V2.5-Pro among them), gathered pace through a permissive-license May, and closed with a June sprint — MiniMax M3 on the 1st, then a Moonshot/Z.ai one-two of Kimi K2.7 Code and GLM-5.2 around the 12th–13th, with Google's Gemma 4 family arriving to anchor the local tier. If Q1 was about MoE architectures maturing, Q2 was about them shipping at scale.

The full Q2 2026 release timeline

The table below groups the headline open-weight drops of the quarter. Parameter and context figures are as reported on the respective model cards; treat active-parameter counts (MoE) as the number actually resident per token, which is what governs your VRAM budget.

Model	Org	Date	License	Params (total / active)	Context	Headline strength
Xiaomi MiMo-V2.5-Pro	Xiaomi	Late Apr	Permissive	Dense	256K	Reasoning / math
MiniCPM-V 4.6 1.3B	OpenBMB	May 11	Apache 2.0	1.3B dense	32K	Edge vision-language
Command A+	Cohere	May 20	Apache 2.0	218B Sparse MoE	256K	Enterprise RAG
MiniMax M3	MiniMax	Jun 1	Open weight	MoE	1M	Frontier coding + multimodal
Kimi K2.7 Code	Moonshot AI	~Jun 12	Open weight	MoE	256K+	Agentic coding
GLM-5.2	Z.ai	~Jun 13	Open weight	MoE	256K+	General + coding
DeepSeek V4	DeepSeek	Q2	Open weight	Large MoE	256K+	Cost-efficient frontier
Gemma 4 (26B-A4B / E2B)	Google	Jun	Gemma license	26B/4B active · 2B	128K+	Local flagship + edge

For canonical specs and license text, the primary sources are the vendor model cards on Hugging Face — for example Google's Gemma org page, Moonshot AI and Cohere For AI. We normalize these into our own benchmarks hub and keep the machine-readable versions in the BestLLMfor public API (CC BY 4.0) and the open-source MCP server, so you can query the whole quarter programmatically.

The frontier tier: MiniMax M3, DeepSeek V4, GLM-5.2, Kimi K2.7

This is where Q2 rewrote expectations. MiniMax M3 led with the quarter's standout spec sheet — a genuine 1M-token context alongside native multimodality — and positioned itself as a frontier coding model rather than a research curiosity. In practice the 1M window is only useful if your serving stack can hold the KV cache, which is a server-GPU proposition, but for large-repo reasoning it is a real capability, not a headline number.

DeepSeek V4 continued the lab's signature play: near-frontier quality at a fraction of the inference cost, thanks to an aggressively sparse MoE design. It is the model to benchmark against when you are pricing a self-hosted deployment, because it sets the floor for cost-per-token that the closed vendors are now forced to answer. GLM-5.2 from Z.ai and Kimi K2.7 Code from Moonshot AI arrived within a day of each other and split the coding crown: GLM-5.2 is the stronger generalist, while K2.7 Code is tuned specifically for agentic, multi-file editing workflows. If your use case is an autonomous coding agent, K2.7 Code is the sharper tool; if you want one model to also handle chat, RAG and analysis, GLM-5.2 is the safer single bet.

Verdict for the frontier tier: on server-class hardware, GLM-5.2 is our default recommendation for mixed workloads and Kimi K2.7 Code for pure coding agents. Reach for MiniMax M3 only when you genuinely need the 1M context and can pay for it in VRAM.

The local hero: Google Gemma 4

For the readers running models on a single desktop GPU or a laptop, Gemma 4 is the release of the quarter. The 26B-A4B variant is a mixture-of-experts design that activates roughly 4B parameters per token while reportedly scoring around 89% on AIME 2026 — reasoning quality that would have needed a closed API a year ago, now fitting comfortably on a 24 GB card at 4-bit. At the other end, Gemma 4 E2B targets ~2 GB of RAM and will run on a Raspberry Pi 5, making it the most credible "real model on tiny hardware" option we have tested this year.

The table below shows realistic memory footprints for the local-tier models at common quantizations. Figures are weights-only estimates; add roughly 15–30% headroom for the KV cache at long context.

Model & quant	Approx. weights size	Min GPU VRAM	Runs on	Best for
Gemma 4 E2B (Q4_K_M)	~2 GB	Integrated / CPU	Raspberry Pi 5, laptops	Edge, offline assistants
Gemma 4 26B-A4B (Q4_K_M)	~16 GB	24 GB	RTX 4090 / 5090 class	Local flagship, reasoning
MiniCPM-V 4.6 1.3B (Q4)	~1 GB	Integrated	Phones, edge boards	On-device vision-language
GLM-5.2 (Q4_K_M)	Large (MoE)	48–80 GB+	A6000 / H100 class	Server coding + chat

Ollama support arrived quickly for the mainstream drops — the Ollama library is the fastest path to a first run — and you can size a deployment against your electricity and hardware costs with our cost calculator.

Enterprise and permissive: Command A+ and MiniCPM-V

Q2's quieter but strategically important story was licensing. Cohere Command A+ (May 20) shipped a 218B Sparse MoE under Apache 2.0, which matters more than any single benchmark: for a commercial team, an Apache-2.0 frontier-class model removes the legal ambiguity that shadows several "open weight" releases whose custom licenses restrict competitive use. Command A+ is tuned for enterprise retrieval and long-document reasoning, and its permissive license makes it the natural choice for products that ship the model to customers.

OpenBMB MiniCPM-V 4.6 1.3B (May 11, also Apache 2.0) rounded out the edge-vision niche. It is not a frontier model and is not trying to be — it is a small, fast vision-language model you can embed on-device without a licensing conversation. Pair it with Gemma 4 E2B and you have a complete offline stack for constrained hardware.

What to actually run

Cutting through the release noise, here is how we'd allocate the quarter's models by hardware budget. The recommendation column is our editorial call, not a spec restatement.

Your hardware	Primary pick	Why	Alternative
Raspberry Pi 5 / laptop iGPU	Gemma 4 E2B	Real reasoning in ~2 GB, fully offline	MiniCPM-V 4.6 (vision)
Single 24 GB GPU	Gemma 4 26B-A4B Q4_K_M	~89% AIME with a 4B active footprint	DeepSeek V4 (quantized)
48–80 GB GPU / server	GLM-5.2	Best open generalist + coding balance	Kimi K2.7 Code (agents)
Multi-GPU / long context	MiniMax M3	Only open model with a true 1M window	DeepSeek V4 (cost)
Commercial product shipping the model	Cohere Command A+	Apache 2.0 removes license risk	MiniCPM-V (edge)

Browse the full set of quantizations and community GGUF builds in our model catalog, and compare head-to-head scores on the benchmarks page before committing hardware.

Conclusion — the verdict

Q2 2026 is the quarter open weights stopped being the budget option and became the default. The right move depends entirely on your hardware: most readers should run Gemma 4, coding teams with server GPUs should run GLM-5.2 or Kimi K2.7 Code, and anyone shipping a commercial product should lean on Apache-2.0 releases like Cohere Command A+. Match the model to the machine and Q2's lineup covers everything from a Raspberry Pi to a multi-GPU server — see our best-of picks for updated rankings as Q3 lands.