Guide · 2026-08-03

Open-Weight LLM Releases — Q3 2026 Wrap-Up

Last updated 2026-08-03

The summer 2026 open-weight wave, ranked: which models to actually download, and which to skip.

By Mohamed Meguedmi · 8 min read

The Q3 2026 release timeline at a glance

The back half of the summer was dense. Below is the shortlist of open-weight drops that matter for people running models locally, not the churn of minor endpoint bumps. Dates and highlights are drawn from official model cards and release trackers; treat headline benchmark figures as indicative until you validate them against your own workload on our benchmarks hub.

Model	Org	License	Released	Headline
MiniMax M3	MiniMax	Apache-2.0	Jun 1, 2026	1M context, native multimodal, frontier coding
Kimi K2.7 Code	Moonshot AI	Modified MIT	Jun 12, 2026	Agentic coding specialist, long-horizon tool use
GLM-5.2	Z.ai	MIT	Jun 13, 2026	Balanced generalist, strong reasoning-per-watt
Nemotron-3 (compressed)	NVIDIA	NVIDIA Open	Jul 6, 2026	120B-A12B pruned via "Iterative Puzzle"
DeepSeek V4	DeepSeek	DeepSeek (permissive)	Jul 2026	Frontier MoE, upgraded reasoning
DeepSeek-V4-Flash-0731	DeepSeek	DeepSeek (permissive)	Jul 31, 2026	Fast distill, single-GPU friendly

Coding-focused releases: MiniMax M3, Kimi K2.7 Code, GLM-5.2

The June cluster made one thing clear: open-weight labs now treat agentic coding as the flagship benchmark, not chat. MiniMax M3 (model card) is the most ambitious — a 1M-token context window and native multimodality mean it can ingest a full repository plus screenshots in a single pass. That capability is real, but the context length is a memory tax: you pay for KV cache linearly, and few local setups will feed it a genuine million tokens without spilling.

Kimi K2.7 Code from Moonshot AI is the pragmatic counterweight. It trades raw context for long-horizon tool-use reliability — it finishes multi-step edit/run/test loops without losing the thread more often than its rivals in our internal harness. GLM-5.2 from Z.ai is the generalist that punches above its size on reasoning-per-watt, and its clean MIT license makes it the easiest to embed in a commercial product.

Model	Coding (SWE-style, %)	Reasoning (MMLU-Pro, %)	Notes
MiniMax M3	~71	~82	Best when context volume is the bottleneck
Kimi K2.7 Code	~73	~78	Most reliable agent loops
GLM-5.2	~66	~80	Best license + efficiency balance
DeepSeek-V4-Flash	~64	~79	Best on a single 24 GB GPU

Figures are indicative composites, not vendor-reported single scores. Reproduce with your own prompts before committing.

DeepSeek V4 and the Flash distills

DeepSeek V4 is the release that reshuffles the top of the open-weight leaderboard. The full MoE is a data-center model — you are not running it comfortably on consumer hardware — but the lineage is what matters. On Jul 31 the team shipped DeepSeek-V4-Flash-0731, a distilled variant that keeps most of V4's reasoning while collapsing into a footprint a single high-end GPU can host. This is the pattern DeepSeek has repeated since V2: ship the frontier model, then ship the distill a few weeks later. For local users, the Flash line is almost always the one to download. Weights and configs land on the DeepSeek HuggingFace org and mirror to ollama.com within days.

If you only pull one model from this wrap-up onto a 24 GB card, make it DeepSeek-V4-Flash at Q4_K_M. It is the best accuracy-per-gigabyte on the list.

Efficiency plays: NVIDIA Nemotron compression

The most technically interesting release wasn't a new frontier model — it was NVIDIA's compressed Nemotron-3 (Jul 6). It takes Nemotron-3-Super-120B-A12B and applies "Iterative Puzzle," a post-training framework that jointly prunes MoE experts, the active-parameter budget, and Mamba state to raise inference throughput while holding accuracy. For anyone deploying at scale, the story is tokens-per-second-per-dollar, not leaderboard position. Run the numbers for your own volume in our cost calculator before assuming a bigger model is worse value.

Hardware & quantization — what you can actually run

Benchmarks are noise if the model doesn't fit in VRAM. Here is the practical view for the Q3 releases at Q4_K_M, the quantization most local users default to. Add roughly 15-25% headroom for context and you have a realistic floor.

Model (Q4_K_M)	Total / active params	Approx. VRAM	Usable context	Best fit
DeepSeek-V4-Flash-0731	~30B / ~3B	~19 GB	128K	Single 24 GB GPU
GLM-5.2	~32B dense	~20 GB	128K	Single 24 GB GPU
Kimi K2.7 Code	~48B / ~8B	~30 GB	256K	2× 24 GB or 48 GB card
MiniMax M3	~70B+ MoE	~44 GB	1M (KV-limited)	48 GB+ / multi-GPU
Nemotron-3 (compressed)	~120B / ~12B	~70 GB	128K	Server / A100-class

Every model in the table is browsable with quantization and license filters in our model catalog, and the same numbers are exposed programmatically through the BestLLMfor public API (CC BY 4.0) and our open-source MCP server, so you can wire release data straight into your own tooling.

Licensing watch — Meta goes closed

The structural story of Q3 isn't a benchmark — it's a license. While DeepSeek, MiniMax, Moonshot and Z.ai shipped permissive weights (Apache-2.0 and MIT dominate this cohort), Meta moved its flagship line closed with Muse Spark. If your product assumed a perpetual open Llama flagship, that assumption expired this quarter. The practical takeaway: the center of gravity for genuinely open weights has shifted decisively toward Chinese labs, and any long-term local-LLM strategy should plan around that.

Verdict

There is no single winner — there is a right pick per constraint. Match the model to your hardware and use case rather than chasing the top of a leaderboard.

Your situation	Pick	Why
Single 24 GB GPU, general use	DeepSeek-V4-Flash-0731	Best accuracy-per-gigabyte on the list
Commercial product, clean license	GLM-5.2	MIT + strong reasoning-per-watt
Agentic coding on bigger hardware	Kimi K2.7 Code	Most reliable long-horizon tool loops
Repo-scale / multimodal context	MiniMax M3	Genuine 1M context, native multimodal
High-volume server deployment	Nemotron-3 compressed	Best throughput-per-dollar at scale

For per-task rankings that update as new weights drop, keep our best-for guides bookmarked.