What are the best open-weight LLMs released in August 2026?

By deployment class: GLM-5.3-Flash (MIT, 320B total / 18B active, 1M context, natively multimodal) for single-rack serving; GLM-5.3 (753B, Terminal-Bench 2.1 88.2) for agentic coding on an 8-GPU node; Tencent Hy4-preview (Apache 2.0, 770B / 49B) and DeepSeek-V4-Pro-0813 (MIT, 1.6-1.7T / 49B) for permissively licensed frontier scale; Qwen3.8-Flash-Next (125B / 6B active) for the lowest decode cost; and Meta Muse Glimmer 30B, IBM Granite 4.2-30B, or Qwen3.8-27B for a single 24-32 GB GPU.

Which new open-source models are MIT or Apache 2.0 licensed?

MIT: GLM-5.3-Flash, DeepSeek-V4-Pro-0813, and DeepSeek-V4-Flash-Vision-Exp. Apache 2.0: Tencent Hy4-preview, Meta Muse Glimmer 30B, IBM Granite 4.2 (3B/8B/30B), and Qwen3.8-27B. Custom licenses: GLM-5.3 (adds a security-review clause for Model-as-a-Service operators with more than US$10B group revenue), Qwen3.8-Flash-Next (Qwen Community License 1.0), Qwen3.8-Max weights, and Kimi K3.

What hardware do I need to run GLM-5.3-Flash or Qwen3.8-Flash-Next locally?

Both fit a 100-200 GB memory budget in quantized form. Unsloth's GLM-5.3-Flash GGUF is 109 GB at UD-Q2_K_XL and 200 GB at UD-Q4_K_XL (FP8 is about 306 GiB before KV cache). Qwen3.8-Flash-Next is 111 GB at UD-Q4_K_XL, 173 GiB in the official FP8 checkpoint, and its 51B N-gram table can be offloaded to system RAM. Practically that means a 2x H100/H200 server, a 4-GPU node, or a large unified-memory workstation cluster.

Is GLM-5.3 better than GLM-5.3-Flash?

On the published cards, GLM-5.3 leads on Terminal-Bench 2.1 (88.2 vs 84.3) and is the stronger pure coding model, but it is text-only, 753B parameters, needs a multi-GPU node even at 2-4-bit (roughly 245-400 GB), and ships under a custom license. GLM-5.3-Flash is a new hybrid-attention base, natively multimodal, 18B active, MIT licensed, and about one-tenth the API price. For most enterprise deployments Flash is the better default; GLM-5.3 is the upgrade for agentic coding workloads that already run GLM-5.2, since it shares that base.

Best Open-Weight Models, August 2026: Release Roundup

Between August 10 and August 31, 2026, ten open-weight language models landed on Hugging Face, and five of them are frontier-class mixture-of-experts models that arrived inside the final week. The best open-weight models of August 2026, by deployment class, are: GLM-5.3-Flash (Z.ai, MIT, 320B total / 18B active, 1M context, natively multimodal) for single-rack enterprise serving; GLM-5.3 (Z.ai, 753B, Terminal-Bench 2.1 of 88.2) for agentic coding on an 8-GPU node; Hy4-preview (Tencent, Apache 2.0, 770B / 49B) and DeepSeek-V4-Pro-0813 (MIT, 1.6-1.7T / 49B) for permissively licensed frontier scale; Qwen3.8-Flash-Next (Alibaba, 125B / 6B active) for the lowest decode cost; and Muse Glimmer 30B (Meta), Granite 4.2-30B (IBM), and Qwen3.8-27B (Alibaba) for a single 24-32 GB GPU. The differences that now matter are licence terms and memory footprint, not raw capability. This post is the release record; our enterprise open-weight comparison carries the portfolio recommendations.

Which open-weight models were released in August 2026?

All figures below come from the Hugging Face model cards or vendor documentation unless flagged. Kimi K3 is included as context because its weights landed on July 27, 2026, days before the window opened.

Date	Model	Org	Total / active params	Context	License	Hugging Face repo
Jul 27 (context)	Kimi K3	Moonshot AI	2.8T / 104B on card (press reports ~50B)	1M	Custom Kimi K3 License	`moonshotai/Kimi-K3`
Aug 10	Muse Glimmer 30B	Meta	~29.6B dense (incl. 1.8B vision encoder)	131K+	Apache 2.0	`meta-models/muse-glimmer` collection
Aug 12-13	Qwen3.8-Max weights	Alibaba	2.4T / 95B	262K native, 1M extended	Custom qwen3.8-max	`Qwen/Qwen3.8-2.4T-A95B`
Aug 13	DeepSeek-V4-Pro-0813	DeepSeek	1.6-1.7T / 49B	1M	MIT	`deepseek-ai/DeepSeek-V4-Pro-0813`
Mid-Aug	Qwen3.8-27B	Alibaba	27B dense, multimodal	262K native, 1M extended	Apache 2.0	`Qwen/Qwen3.8-27B`
Aug 25	Granite 4.2 (3B / 8B / 30B)	IBM	Dense	131K shipped (512K claimed)	Apache 2.0	`ibm-granite/granite-4.2-30b` (and -3b, -8b)
Aug 25-26	GLM-5.3-Flash	Z.ai	320B / 18B	1M (128K output)	MIT	`zai-org/GLM-5.3-Flash`
Aug 26	Qwen3.8-Flash-Next	Alibaba	125B / 6B (+51B N-gram table, 4B MTP head)	262K native, 1M with YaRN	Qwen Community 1.0	`Qwen/Qwen3.8-Flash-Next`
Aug 28	GLM-5.3	Z.ai	753B / ~40B (active figure is third-party derived)	1M	Custom GLM-5.3 License	`zai-org/GLM-5.3`
Aug 28	Hy4-preview	Tencent	770B / 49B	1M	Apache 2.0	`tencent/Hy4-preview`
Aug 31	DeepSeek-V4-Flash-Vision-Exp	DeepSeek	304.6B / 13B	1M	MIT	`deepseek-ai/DeepSeek-V4-Flash-Vision-Exp`

GLM-5.3 was announced on August 14 inside Z.ai's GLM Coding Plan only; the weights followed on August 28 after a security review.

What happened in the last week of August?

Five frontier-class models shipped between August 25 and August 31, each changing a different part of the self-hosting calculus.

GLM-5.3-Flash (Aug 25-26) is the release of the month for on-premise teams. It is a new base model, not a GLM-5.2 derivative: the first GLM with hybrid sparse plus linear attention, the first natively multimodal GLM-5 (text, image, video, visual documents), 320B total with 18B active, 1M context with 128K output, and a plain MIT licence. The card reports Terminal-Bench 2.1 of 84.3 and DeepSWE v1.1 of 63.4 against 46.2 for GLM-5.2, at roughly one-tenth the API price of GLM-5.3. Our GLM-5.3-Flash model page has the serving detail.

Qwen3.8-Flash-Next (Aug 26) previews the Qwen4 architecture. Alibaba calls it "an early preview of the architecture that will underpin Qwen4": 125B main parameters with 6B active, Gated DeltaNet linear attention in three of every four layers, and a 51B-parameter N-gram embedding table that "can be offloaded to host memory and overlapped with model computation." Capacity moves into cheap system RAM instead of VRAM. It takes text, image, and video in, runs 262K natively (1M with YaRN), and reports SWE-bench Pro of 62.5. The licence is Qwen Community 1.0, not Apache 2.0.

GLM-5.3 (Aug 28) is the strongest open coding model on the card numbers. Same 753B base as GLM-5.2, with all gains from scaled post-training. Terminal-Bench 2.1 of 88.2, Terminal-Bench 3.0 of 28.3 (GLM-5.2: 4.6), DeepSWE v1.1 of 66.9. It is text-only and runs on the same vLLM and SGLang shape as GLM-5.2, so existing deployments upgrade in place. The licence is the story, covered below. See the GLM-5.3 model page.

Hy4-preview (Aug 28) is Tencent's Apache 2.0 answer at 770B. 49B active, 256 routed experts plus one shared with top-8 routing, 1M context, text-only. The card reports GPQA Diamond 92.3, SWE-bench Pro 65.7, and Terminal-Bench 2.1 85.4. The repo is 1.56 TB and the reference configs use tensor parallelism of 8.

DeepSeek-V4-Flash-Vision-Exp (Aug 31) puts DeepSeek vision under MIT. 304.6B total (the 284B V4-Flash plus a vision tower), 13B active, 1M context, up to 384 tokens per image. The API went live August 21 and the weights followed ten days later. DeepSeek ships a dedicated vLLM image and an SGLang config for a 4-GPU node.

How do the August 2026 releases compare on benchmarks?

These are vendor-reported figures from the respective model cards. Treat them as directional and reproduce them on your own harness before signing a procurement memo; that reproducibility is what a pinned on-premise deployment buys you.

Model	Terminal-Bench 2.1	DeepSWE v1.1	SWE-bench Pro	GPQA Diamond
Kimi K3 (context)	88.3	67.5	n/r	93.5
GLM-5.3	88.2	66.9	n/r	n/r
DeepSeek-V4-Pro-0813	87.9	62.7	n/r	n/r
Qwen3.8-Max (weights)	86.6	n/r	67.7	92.6
Hy4-preview	85.4	64.3	65.7	92.3
GLM-5.3-Flash	84.3	63.4	n/r	n/r
DeepSeek-V4-Flash-Vision-Exp	83.9	59.3	n/r	n/r
Qwen3.8-Flash-Next	n/r	58.7	62.5	91.7
Qwen3.8-27B	73.0	n/r	61.7	89.2
Muse Glimmer 30B	n/r	n/r	51.2	n/r
Granite 4.2-30B	n/r	n/r	57.0 (SWE-bench Verified)	n/r

n/r = not reported on the card. The spread at the top is now inside three points on Terminal-Bench 2.1 across five different labs, which is the clearest evidence yet that the open bench has converged on frontier agentic capability.

Which August 2026 model fits your hardware?

Memory figures are checkpoint or GGUF sizes from the repos; add KV cache on top for real serving.

Memory class	Typical hardware	Models	Footprint evidence
24-32 GB VRAM	One RTX 5090 or workstation GPU	Muse Glimmer 30B, Granite 4.2-30B, Qwen3.8-27B	Glimmer ships K-quant variants at 17 GB and 32 GB (55+ GB full precision) with a DFlash speculative decoder rated 3.1x on an RTX 5090. Granite 4.2-30B and Qwen3.8-27B are dense models that reach this class at 4-bit.
100-200 GB	2x H100/H200, a 4-GPU node, or a unified-memory cluster	Qwen3.8-Flash-Next, GLM-5.3-Flash, DeepSeek-V4-Flash-Vision-Exp	Flash-Next: 111 GB at UD-Q4_K_XL, 173 GiB official FP8. GLM-5.3-Flash: 109 GB at UD-Q2_K_XL, 200 GB at UD-Q4_K_XL, ~306 GiB FP8. DeepSeek Vision repo ~168 GB.
8-GPU node	8x H200 / GB300-class or multi-node	GLM-5.3, Hy4-preview, DeepSeek-V4-Pro-0813, Qwen3.8-Max, Kimi K3	GLM-5.3: ~245 GB at 2-bit to ~810 GB at 8-bit (third-party estimate). Hy4: 1.56 TB, TP=8. V4-Pro-0813: ~893 GB. Kimi K3: 1.56 TB, MXFP4 supported.

The middle tier is where most regulated organizations should look. In September 2026 a two-GPU server runs GLM-5.3-Flash at 4-bit with 1M context and a multimodal encoder, or Qwen3.8-Flash-Next with 6B-active decode speed. Our on-premise deployment cost guide has the break-even math for each tier.

Why is the licence now the differentiator?

With capability compressed into a three-point band, procurement teams are choosing on legal terms, and August split the field cleanly.

Plain MIT: GLM-5.3-Flash, DeepSeek-V4-Pro-0813, DeepSeek-V4-Flash-Vision-Exp. No restrictions on commercial use, modification, redistribution of fine-tunes, or air-gapped deployment.

Apache 2.0: Hy4-preview, Muse Glimmer 30B, Granite 4.2, Qwen3.8-27B. Same practical freedom, plus an explicit patent grant. Meta's return to a permissive licence after its closed pivot, and IBM continuing under Apache 2.0, matter for organizations whose policy requires Western-origin weights.

Custom terms: GLM-5.3 ships under a "GLM-5.3 License" that grants the MIT-style rights (use, copy, modify, distribute, sublicense, sell, fine-tune) but adds one clause: Model-as-a-Service operators whose group revenue exceeds US$10 billion over any consecutive 12 months must pass a Z.ai security review before commercial use. Embedding the model in an end-user product does not trigger the review. For nearly every enterprise and sovereign deployment this is a non-event, but it is no longer MIT and your counsel will read it. Qwen3.8-Flash-Next uses the Qwen Community License 1.0 and Qwen3.8-Max its own custom licence; we have not reviewed the full text of either, and one analyst report of geographic restrictions in the Qwen3.8-Max licence is unverified. Kimi K3 remains under its custom Kimi K3 License.

For a bank or insurer preparing for OSFI Guideline E-23, the licence line is not academic. A model-risk file needs the licence, the checksum, and the evaluation harness in one place, and the fewer bespoke clauses in that file, the faster it clears review.

What did the proprietary labs ship in the same window?

The closed frontier did not stand still. Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, sharing one underlying model, with 1M-token input, 128K output, and cache-read pricing cut 75%; Mythos is gated to vetted cyber and life-science organizations. Google followed on September 2 with Gemini 3.8 Flash at 1M context and an introductory US$0.75 / US$3.75 per million that doubles on January 1, 2027. Both are strong, and both are rented: pricing, availability, and behaviour change on the vendor's schedule. The August open releases are the alternative you can pin.

What should an enterprise do with this in September 2026?

Standardize the single-rack tier on GLM-5.3-Flash. MIT, multimodal, 1M context, 18B active, 200 GB at 4-bit. It replaces GLM-5.2 for most deployments at less than half the memory.
Upgrade coding agents that run GLM-5.2 to GLM-5.3 in place. Same base and serving shape; have counsel file the licence change.
Put Hy4-preview and DeepSeek-V4-Pro-0813 on the evaluation list for 8-GPU nodes. Both are permissively licensed at frontier scale.
Use Muse Glimmer, Granite 4.2-30B, or Qwen3.8-27B for edge and workstation deployments, and to satisfy Western-origin procurement rules where they apply.
Trial Qwen3.8-Flash-Next where decode cost dominates, after reading the Qwen Community License.

Every model in this post is listed with hardware tiers and serving notes in our models directory. The comparison and portfolio recommendation, refreshed to this bench, is in Best Open-Weight LLM for Enterprise in 2026.

Best Open-Weight Models of August 2026: The Release Roundup