On 26 August 2026 Alibaba's Qwen team published Qwen3.8-Flash-Next, an open-weight mixture-of-experts model with 125B total parameters that activates 6B per token, alongside a 51B n-gram embedding table and a 4B multi-token prediction module.1 Hugging Face, Qwen team model card 2026-08-26 125B total parameters with 6B activated per token, a 51B n-gram embedding layer and a 4B multi-token prediction module; 48 layers at hidden dimension 2560; hybrid of Gated DeltaNet and Qwen Sparse Attention with Gated Residual at bottleneck rank 320; Muon and AdamW optimizers; native context of 262,144 tokens extensible to one million; qwen-community-1.0 license; reported 62.5 on SWE-bench Pro, 73.9 on CoWorkBench, 84.5 on AndroidWorld. Open source The same day, Z.ai published GLM-5.3-Flash, a natively multimodal mixture-of-experts model with 320B total parameters and 18B active, under an MIT license.6 MarkTechPost 2026-08-26 Z.ai released GLM-5.3-Flash on 26 August 2026: 320B total parameters, 18B active per token, 1,048,576 token context, native image and video input, MIT licensed weights on Hugging Face; 84.3 on Terminal-Bench 2.1 against Claude Opus 4.8 at 85.0 and 63.4 on DeepSWE v1.1 against GLM-5.2 at 46.2; hybrid linear and sparse attention, IndexPool cutting KV cache 4.4 times, Manifold-Constrained Hyper-Connections; API pricing $0.15 input and $0.50 output per million tokens; self hosting roughly 306 GiB of FP8 weights on Hopper class GPUs or newer. Open source The stake is not the benchmark table. It is that Qwen reports training its model for roughly one ninth the cost of its own predecessor Qwen3.7-Plus, and prices the hosted version at $0.16 per million input tokens.3 MarkTechPost 2026-08-26 512 experts per mixture-of-experts layer with ten routed and one shared; roughly 20 million bigram and trigram entries at layer 2; training cost approximately one ninth that of Qwen3.7-Plus; up to 7.6 times prefill and 4.9 times decode speedup at one million tokens; FP8 weights 172.78 GiB and BF16 weights 335.28 GiB; tensor parallel 2 on GB300, TP8 recommended on H200; license qwen-community-1.0 rather than Apache 2.0. Open source 4 DataCamp 2026-08-26 Released 26 August 2026; text, image and video input; native context 262,144 tokens extensible to one million with YaRN; 62.5 against 53.4 for Claude Opus 4.6 Max on SWE-bench Pro, 73.9 against 68.2 on CoWorkBench, 55.7 against 36.6 on JobBench, trailing at 35.9 against 40.0 on Humanity's Last Exam; QwenCloud pricing $0.16 per million input tokens and $0.47 per million output tokens; training cost roughly a ninth of Qwen3.7-Plus. Open source We assess with moderate confidence that the competitive axis in open-weight models has moved from parameter count to activated parameters per token, and that the two releases landing on one day is a coordination of the calendar rather than of the labs.

What sparsity now buys

The architecture of Qwen3.8-Flash-Next is the point of the release, not its scores. The model card describes 48 layers at hidden dimension 2560 arranged as a hybrid: three Gated DeltaNet blocks with mixture-of-experts, then one Qwen Sparse Attention block with mixture-of-experts, repeating.1 Hugging Face, Qwen team model card 2026-08-26 125B total parameters with 6B activated per token, a 51B n-gram embedding layer and a 4B multi-token prediction module; 48 layers at hidden dimension 2560; hybrid of Gated DeltaNet and Qwen Sparse Attention with Gated Residual at bottleneck rank 320; Muon and AdamW optimizers; native context of 262,144 tokens extensible to one million; qwen-community-1.0 license; reported 62.5 on SWE-bench Pro, 73.9 on CoWorkBench, 84.5 on AndroidWorld. Open source 2 Unite.AI 2026-08-26 Released 26 August 2026 as an open-weight preview of the Qwen4 architecture; BF16 weights total roughly 180B parameters; 48 layers repeating three Gated DeltaNet blocks with mixture-of-experts then one Qwen Sparse Attention block with mixture-of-experts; element-wise data dependent read gates; training recipe eliminates warmup; vendor reported 58.7 DeepSWE 1.1, 62.5 SWE-bench Pro, 91.7 GPQA Diamond, 91.9 LiveCodeBench v6. Open source Each mixture-of-experts layer holds 512 experts and routes ten plus one shared expert per token.3 MarkTechPost 2026-08-26 512 experts per mixture-of-experts layer with ten routed and one shared; roughly 20 million bigram and trigram entries at layer 2; training cost approximately one ninth that of Qwen3.7-Plus; up to 7.6 times prefill and 4.9 times decode speedup at one million tokens; FP8 weights 172.78 GiB and BF16 weights 335.28 GiB; tensor parallel 2 on GB300, TP8 recommended on H200; license qwen-community-1.0 rather than Apache 2.0. Open source A separate table of roughly 20 million bigram and trigram entries sits at layer 2 and carries 51B parameters that never all activate at once.3 MarkTechPost 2026-08-26 512 experts per mixture-of-experts layer with ten routed and one shared; roughly 20 million bigram and trigram entries at layer 2; training cost approximately one ninth that of Qwen3.7-Plus; up to 7.6 times prefill and 4.9 times decode speedup at one million tokens; FP8 weights 172.78 GiB and BF16 weights 335.28 GiB; tensor parallel 2 on GB300, TP8 recommended on H200; license qwen-community-1.0 rather than Apache 2.0. Open source Read gates are element-wise and data dependent, and the training recipe drops the conventional warmup phase and applies the Muon optimizer alongside AdamW.1 Hugging Face, Qwen team model card 2026-08-26 125B total parameters with 6B activated per token, a 51B n-gram embedding layer and a 4B multi-token prediction module; 48 layers at hidden dimension 2560; hybrid of Gated DeltaNet and Qwen Sparse Attention with Gated Residual at bottleneck rank 320; Muon and AdamW optimizers; native context of 262,144 tokens extensible to one million; qwen-community-1.0 license; reported 62.5 on SWE-bench Pro, 73.9 on CoWorkBench, 84.5 on AndroidWorld. Open source 2 Unite.AI 2026-08-26 Released 26 August 2026 as an open-weight preview of the Qwen4 architecture; BF16 weights total roughly 180B parameters; 48 layers repeating three Gated DeltaNet blocks with mixture-of-experts then one Qwen Sparse Attention block with mixture-of-experts; element-wise data dependent read gates; training recipe eliminates warmup; vendor reported 58.7 DeepSWE 1.1, 62.5 SWE-bench Pro, 91.7 GPQA Diamond, 91.9 LiveCodeBench v6. Open source

The consequences are measured in memory and throughput rather than in intelligence. Reported speedups reach up to 7.6 times on prefill and 4.9 times on decode at one million tokens of context.3 MarkTechPost 2026-08-26 512 experts per mixture-of-experts layer with ten routed and one shared; roughly 20 million bigram and trigram entries at layer 2; training cost approximately one ninth that of Qwen3.7-Plus; up to 7.6 times prefill and 4.9 times decode speedup at one million tokens; FP8 weights 172.78 GiB and BF16 weights 335.28 GiB; tensor parallel 2 on GB300, TP8 recommended on H200; license qwen-community-1.0 rather than Apache 2.0. Open source Native context is 262,144 tokens, extensible to one million.1 Hugging Face, Qwen team model card 2026-08-26 125B total parameters with 6B activated per token, a 51B n-gram embedding layer and a 4B multi-token prediction module; 48 layers at hidden dimension 2560; hybrid of Gated DeltaNet and Qwen Sparse Attention with Gated Residual at bottleneck rank 320; Muon and AdamW optimizers; native context of 262,144 tokens extensible to one million; qwen-community-1.0 license; reported 62.5 on SWE-bench Pro, 73.9 on CoWorkBench, 84.5 on AndroidWorld. Open source 4 DataCamp 2026-08-26 Released 26 August 2026; text, image and video input; native context 262,144 tokens extensible to one million with YaRN; 62.5 against 53.4 for Claude Opus 4.6 Max on SWE-bench Pro, 73.9 against 68.2 on CoWorkBench, 55.7 against 36.6 on JobBench, trailing at 35.9 against 40.0 on Humanity's Last Exam; QwenCloud pricing $0.16 per million input tokens and $0.47 per million output tokens; training cost roughly a ninth of Qwen3.7-Plus. Open source The n-gram table can be offloaded to ordinary system RAM rather than held in accelerator memory, which is the practical reason a 176B parameter footprint can be served on a machine sized for something much smaller.5 eesel AI 2026-08-26 Flash Next is the self hosted open weights while the production Qwen3.8-Flash API runs on QwenCloud with a one million token default context; the 51B n-gram lookup table can be offloaded to system RAM; hosted pricing $0.15 input and $0.47 output per million tokens with cached reads about 89 percent cheaper; Artificial Analysis ranks it fifth for intelligence in class at 56 on its Intelligence Index; the model is very verbose; described as an intentionally under-trained research preview with GLM 5.3 Flash recommended for production. Open source These figures are the lab's own, published without independent replication, and should be read as vendor claims.

GLM-5.3-Flash arrives at the same destination by a different route. Z.ai reports hybrid linear and sparse attention, an IndexPool compression scheme that cuts the key value cache by 4.4 times, and a structure it calls Manifold-Constrained Hyper-Connections.6 MarkTechPost 2026-08-26 Z.ai released GLM-5.3-Flash on 26 August 2026: 320B total parameters, 18B active per token, 1,048,576 token context, native image and video input, MIT licensed weights on Hugging Face; 84.3 on Terminal-Bench 2.1 against Claude Opus 4.8 at 85.0 and 63.4 on DeepSWE v1.1 against GLM-5.2 at 46.2; hybrid linear and sparse attention, IndexPool cutting KV cache 4.4 times, Manifold-Constrained Hyper-Connections; API pricing $0.15 input and $0.50 output per million tokens; self hosting roughly 306 GiB of FP8 weights on Hopper class GPUs or newer. Open source Its context window is 1,048,576 tokens natively and it takes image and video input directly.6 MarkTechPost 2026-08-26 Z.ai released GLM-5.3-Flash on 26 August 2026: 320B total parameters, 18B active per token, 1,048,576 token context, native image and video input, MIT licensed weights on Hugging Face; 84.3 on Terminal-Bench 2.1 against Claude Opus 4.8 at 85.0 and 63.4 on DeepSWE v1.1 against GLM-5.2 at 46.2; hybrid linear and sparse attention, IndexPool cutting KV cache 4.4 times, Manifold-Constrained Hyper-Connections; API pricing $0.15 input and $0.50 output per million tokens; self hosting roughly 306 GiB of FP8 weights on Hopper class GPUs or newer. Open source Two independent teams, in the same week, spent their engineering budget on the same problem: how little of a model you have to wake up to answer a token.

The benchmark claims, and what they are worth

Both labs published comparisons against Anthropic models. Qwen reports 62.5 on SWE-bench Pro against 53.4 for Claude Opus 4.6 Max, 73.9 on CoWorkBench against 68.2, and 55.7 on JobBench against 36.6, while trailing on Humanity's Last Exam at 35.9 against 40.0.4 DataCamp 2026-08-26 Released 26 August 2026; text, image and video input; native context 262,144 tokens extensible to one million with YaRN; 62.5 against 53.4 for Claude Opus 4.6 Max on SWE-bench Pro, 73.9 against 68.2 on CoWorkBench, 55.7 against 36.6 on JobBench, trailing at 35.9 against 40.0 on Humanity's Last Exam; QwenCloud pricing $0.16 per million input tokens and $0.47 per million output tokens; training cost roughly a ninth of Qwen3.7-Plus. Open source Z.ai reports 84.3 on Terminal-Bench 2.1 against 85.0 for Claude Opus 4.8, and 63.4 on DeepSWE v1.1 against 46.2 for its own GLM-5.2.6 MarkTechPost 2026-08-26 Z.ai released GLM-5.3-Flash on 26 August 2026: 320B total parameters, 18B active per token, 1,048,576 token context, native image and video input, MIT licensed weights on Hugging Face; 84.3 on Terminal-Bench 2.1 against Claude Opus 4.8 at 85.0 and 63.4 on DeepSWE v1.1 against GLM-5.2 at 46.2; hybrid linear and sparse attention, IndexPool cutting KV cache 4.4 times, Manifold-Constrained Hyper-Connections; API pricing $0.15 input and $0.50 output per million tokens; self hosting roughly 306 GiB of FP8 weights on Hopper class GPUs or newer. Open source Every one of those numbers is self reported by the lab that shipped the model, on benchmarks the lab selected, against comparison scores it also selected.

One independent read exists and it cuts the other way. Artificial Analysis places Qwen3.8-Flash-Next fifth for intelligence in its class with a composite score of 56, and notes the model is verbose enough to consume materially more output tokens than comparable systems.5 eesel AI 2026-08-26 Flash Next is the self hosted open weights while the production Qwen3.8-Flash API runs on QwenCloud with a one million token default context; the 51B n-gram lookup table can be offloaded to system RAM; hosted pricing $0.15 input and $0.47 output per million tokens with cached reads about 89 percent cheaper; Artificial Analysis ranks it fifth for intelligence in class at 56 on its Intelligence Index; the model is very verbose; described as an intentionally under-trained research preview with GLM 5.3 Flash recommended for production. Open source Verbosity is not a footnote when output tokens are billed at $0.47 per million: a model that talks twice as much is not half the price of one that costs twice as much per token.4 DataCamp 2026-08-26 Released 26 August 2026; text, image and video input; native context 262,144 tokens extensible to one million with YaRN; 62.5 against 53.4 for Claude Opus 4.6 Max on SWE-bench Pro, 73.9 against 68.2 on CoWorkBench, 55.7 against 36.6 on JobBench, trailing at 35.9 against 40.0 on Humanity's Last Exam; QwenCloud pricing $0.16 per million input tokens and $0.47 per million output tokens; training cost roughly a ninth of Qwen3.7-Plus. Open source 5 eesel AI 2026-08-26 Flash Next is the self hosted open weights while the production Qwen3.8-Flash API runs on QwenCloud with a one million token default context; the 51B n-gram lookup table can be offloaded to system RAM; hosted pricing $0.15 input and $0.47 output per million tokens with cached reads about 89 percent cheaper; Artificial Analysis ranks it fifth for intelligence in class at 56 on its Intelligence Index; the model is very verbose; described as an intentionally under-trained research preview with GLM 5.3 Flash recommended for production. Open source Qwen itself frames the release as an architecture preview rather than a finished product, and at least one reviewer describes it as intentionally under-trained.2 Unite.AI 2026-08-26 Released 26 August 2026 as an open-weight preview of the Qwen4 architecture; BF16 weights total roughly 180B parameters; 48 layers repeating three Gated DeltaNet blocks with mixture-of-experts then one Qwen Sparse Attention block with mixture-of-experts; element-wise data dependent read gates; training recipe eliminates warmup; vendor reported 58.7 DeepSWE 1.1, 62.5 SWE-bench Pro, 91.7 GPQA Diamond, 91.9 LiveCodeBench v6. Open source 5 eesel AI 2026-08-26 Flash Next is the self hosted open weights while the production Qwen3.8-Flash API runs on QwenCloud with a one million token default context; the 51B n-gram lookup table can be offloaded to system RAM; hosted pricing $0.15 input and $0.47 output per million tokens with cached reads about 89 percent cheaper; Artificial Analysis ranks it fifth for intelligence in class at 56 on its Intelligence Index; the model is very verbose; described as an intentionally under-trained research preview with GLM 5.3 Flash recommended for production. Open source We assess with high confidence that the coding benchmark leads reported here will not survive independent evaluation intact.

Second order effects: the ledger

Who gains. Inference providers and self hosting shops gain first, because the binding constraint on serving these models is now accelerator memory rather than compute, and the deployment notes make that arithmetic explicit: 172.78 GiB in FP8 for Qwen against 335.28 GiB in BF16, roughly 306 GiB for GLM-5.3-Flash in FP8.3 MarkTechPost 2026-08-26 512 experts per mixture-of-experts layer with ten routed and one shared; roughly 20 million bigram and trigram entries at layer 2; training cost approximately one ninth that of Qwen3.7-Plus; up to 7.6 times prefill and 4.9 times decode speedup at one million tokens; FP8 weights 172.78 GiB and BF16 weights 335.28 GiB; tensor parallel 2 on GB300, TP8 recommended on H200; license qwen-community-1.0 rather than Apache 2.0. Open source 6 MarkTechPost 2026-08-26 Z.ai released GLM-5.3-Flash on 26 August 2026: 320B total parameters, 18B active per token, 1,048,576 token context, native image and video input, MIT licensed weights on Hugging Face; 84.3 on Terminal-Bench 2.1 against Claude Opus 4.8 at 85.0 and 63.4 on DeepSWE v1.1 against GLM-5.2 at 46.2; hybrid linear and sparse attention, IndexPool cutting KV cache 4.4 times, Manifold-Constrained Hyper-Connections; API pricing $0.15 input and $0.50 output per million tokens; self hosting roughly 306 GiB of FP8 weights on Hopper class GPUs or newer. Open source A quantized Qwen fits on a two way GB300 configuration.3 MarkTechPost 2026-08-26 512 experts per mixture-of-experts layer with ten routed and one shared; roughly 20 million bigram and trigram entries at layer 2; training cost approximately one ninth that of Qwen3.7-Plus; up to 7.6 times prefill and 4.9 times decode speedup at one million tokens; FP8 weights 172.78 GiB and BF16 weights 335.28 GiB; tensor parallel 2 on GB300, TP8 recommended on H200; license qwen-community-1.0 rather than Apache 2.0. Open source Buyers of long context agentic work gain the most in cash terms, since prefill dominates that workload and prefill is where the reported 7.6 times speedup sits.3 MarkTechPost 2026-08-26 512 experts per mixture-of-experts layer with ten routed and one shared; roughly 20 million bigram and trigram entries at layer 2; training cost approximately one ninth that of Qwen3.7-Plus; up to 7.6 times prefill and 4.9 times decode speedup at one million tokens; FP8 weights 172.78 GiB and BF16 weights 335.28 GiB; tensor parallel 2 on GB300, TP8 recommended on H200; license qwen-community-1.0 rather than Apache 2.0. Open source

Who loses. The middle tier of Western hosted models loses pricing cover. GLM-5.3-Flash lists at $0.15 and $0.50 per million tokens and Qwen at $0.15 to $0.16 and $0.47, and both labs are anchoring their marketing to Opus class comparisons.4 DataCamp 2026-08-26 Released 26 August 2026; text, image and video input; native context 262,144 tokens extensible to one million with YaRN; 62.5 against 53.4 for Claude Opus 4.6 Max on SWE-bench Pro, 73.9 against 68.2 on CoWorkBench, 55.7 against 36.6 on JobBench, trailing at 35.9 against 40.0 on Humanity's Last Exam; QwenCloud pricing $0.16 per million input tokens and $0.47 per million output tokens; training cost roughly a ninth of Qwen3.7-Plus. Open source 5 eesel AI 2026-08-26 Flash Next is the self hosted open weights while the production Qwen3.8-Flash API runs on QwenCloud with a one million token default context; the 51B n-gram lookup table can be offloaded to system RAM; hosted pricing $0.15 input and $0.47 output per million tokens with cached reads about 89 percent cheaper; Artificial Analysis ranks it fifth for intelligence in class at 56 on its Intelligence Index; the model is very verbose; described as an intentionally under-trained research preview with GLM 5.3 Flash recommended for production. Open source 6 MarkTechPost 2026-08-26 Z.ai released GLM-5.3-Flash on 26 August 2026: 320B total parameters, 18B active per token, 1,048,576 token context, native image and video input, MIT licensed weights on Hugging Face; 84.3 on Terminal-Bench 2.1 against Claude Opus 4.8 at 85.0 and 63.4 on DeepSWE v1.1 against GLM-5.2 at 46.2; hybrid linear and sparse attention, IndexPool cutting KV cache 4.4 times, Manifold-Constrained Hyper-Connections; API pricing $0.15 input and $0.50 output per million tokens; self hosting roughly 306 GiB of FP8 weights on Hopper class GPUs or newer. Open source Anyone selling a proprietary model whose case rests on coding benchmark position rather than on reliability, support or data handling now has to argue against a free weights file. The other loser is Qwen's own licensing position: GLM-5.3-Flash ships under MIT while Qwen3.8-Flash-Next ships under qwen-community-1.0, not Apache 2.0, which is a real difference for a commercial adopter comparing the two on the same afternoon.3 MarkTechPost 2026-08-26 512 experts per mixture-of-experts layer with ten routed and one shared; roughly 20 million bigram and trigram entries at layer 2; training cost approximately one ninth that of Qwen3.7-Plus; up to 7.6 times prefill and 4.9 times decode speedup at one million tokens; FP8 weights 172.78 GiB and BF16 weights 335.28 GiB; tensor parallel 2 on GB300, TP8 recommended on H200; license qwen-community-1.0 rather than Apache 2.0. Open source 6 MarkTechPost 2026-08-26 Z.ai released GLM-5.3-Flash on 26 August 2026: 320B total parameters, 18B active per token, 1,048,576 token context, native image and video input, MIT licensed weights on Hugging Face; 84.3 on Terminal-Bench 2.1 against Claude Opus 4.8 at 85.0 and 63.4 on DeepSWE v1.1 against GLM-5.2 at 46.2; hybrid linear and sparse attention, IndexPool cutting KV cache 4.4 times, Manifold-Constrained Hyper-Connections; API pricing $0.15 input and $0.50 output per million tokens; self hosting roughly 306 GiB of FP8 weights on Hopper class GPUs or newer. Open source

The counter-case

The strongest argument against our reading is that neither of these is a frontier model and both labs know it. Qwen named the release a preview of an architecture that does not yet have a flagship built on it, and the honest recommendation from at least one reviewer is to run GLM-5.3-Flash in production and treat Flash Next as a research artifact.5 eesel AI 2026-08-26 Flash Next is the self hosted open weights while the production Qwen3.8-Flash API runs on QwenCloud with a one million token default context; the 51B n-gram lookup table can be offloaded to system RAM; hosted pricing $0.15 input and $0.47 output per million tokens with cached reads about 89 percent cheaper; Artificial Analysis ranks it fifth for intelligence in class at 56 on its Intelligence Index; the model is very verbose; described as an intentionally under-trained research preview with GLM 5.3 Flash recommended for production. Open source A one ninth training cost against Qwen3.7-Plus is a claim about a specific predecessor, not about the frontier, and an under-trained model is cheap to train by construction.3 MarkTechPost 2026-08-26 512 experts per mixture-of-experts layer with ten routed and one shared; roughly 20 million bigram and trigram entries at layer 2; training cost approximately one ninth that of Qwen3.7-Plus; up to 7.6 times prefill and 4.9 times decode speedup at one million tokens; FP8 weights 172.78 GiB and BF16 weights 335.28 GiB; tensor parallel 2 on GB300, TP8 recommended on H200; license qwen-community-1.0 rather than Apache 2.0. Open source 5 eesel AI 2026-08-26 Flash Next is the self hosted open weights while the production Qwen3.8-Flash API runs on QwenCloud with a one million token default context; the 51B n-gram lookup table can be offloaded to system RAM; hosted pricing $0.15 input and $0.47 output per million tokens with cached reads about 89 percent cheaper; Artificial Analysis ranks it fifth for intelligence in class at 56 on its Intelligence Index; the model is very verbose; described as an intentionally under-trained research preview with GLM 5.3 Flash recommended for production. Open source Sparsity has been the direction of travel in open-weight releases for two years, so the same-day timing may be nothing but two teams aiming at the end of a month. For our assessment to be wrong, the efficiency gains would have to fail to carry to a fully trained flagship, which is exactly what a preview cannot demonstrate.

What to watch

  • Qwen4 ships on this architecture, or it does not. If a full Qwen4 family lands within roughly six months carrying Gated DeltaNet, Qwen Sparse Attention and the n-gram embedding intact, the preview did its job. A Qwen4 that quietly drops the n-gram table would say the 51B side car did not pay for itself.1 Hugging Face, Qwen team model card 2026-08-26 125B total parameters with 6B activated per token, a 51B n-gram embedding layer and a 4B multi-token prediction module; 48 layers at hidden dimension 2560; hybrid of Gated DeltaNet and Qwen Sparse Attention with Gated Residual at bottleneck rank 320; Muon and AdamW optimizers; native context of 262,144 tokens extensible to one million; qwen-community-1.0 license; reported 62.5 on SWE-bench Pro, 73.9 on CoWorkBench, 84.5 on AndroidWorld. Open source 2 Unite.AI 2026-08-26 Released 26 August 2026 as an open-weight preview of the Qwen4 architecture; BF16 weights total roughly 180B parameters; 48 layers repeating three Gated DeltaNet blocks with mixture-of-experts then one Qwen Sparse Attention block with mixture-of-experts; element-wise data dependent read gates; training recipe eliminates warmup; vendor reported 58.7 DeepSWE 1.1, 62.5 SWE-bench Pro, 91.7 GPQA Diamond, 91.9 LiveCodeBench v6. Open source
  • Independent evaluation of the SWE-bench Pro lead. Watch whether a third party reproduces 62.5 on SWE-bench Pro within one quarter. Artificial Analysis already places the model fifth in class on a composite index, so a large gap between the two readings is the signal that the coding claim is benchmark specific.4 DataCamp 2026-08-26 Released 26 August 2026; text, image and video input; native context 262,144 tokens extensible to one million with YaRN; 62.5 against 53.4 for Claude Opus 4.6 Max on SWE-bench Pro, 73.9 against 68.2 on CoWorkBench, 55.7 against 36.6 on JobBench, trailing at 35.9 against 40.0 on Humanity's Last Exam; QwenCloud pricing $0.16 per million input tokens and $0.47 per million output tokens; training cost roughly a ninth of Qwen3.7-Plus. Open source 5 eesel AI 2026-08-26 Flash Next is the self hosted open weights while the production Qwen3.8-Flash API runs on QwenCloud with a one million token default context; the 51B n-gram lookup table can be offloaded to system RAM; hosted pricing $0.15 input and $0.47 output per million tokens with cached reads about 89 percent cheaper; Artificial Analysis ranks it fifth for intelligence in class at 56 on its Intelligence Index; the model is very verbose; described as an intentionally under-trained research preview with GLM 5.3 Flash recommended for production. Open source
  • Effective cost per completed task, not per token. If verbosity holds, the posted $0.47 per million output tokens will not translate into the price advantage the headline implies. Track cost per resolved issue rather than per token over the next two quarters.4 DataCamp 2026-08-26 Released 26 August 2026; text, image and video input; native context 262,144 tokens extensible to one million with YaRN; 62.5 against 53.4 for Claude Opus 4.6 Max on SWE-bench Pro, 73.9 against 68.2 on CoWorkBench, 55.7 against 36.6 on JobBench, trailing at 35.9 against 40.0 on Humanity's Last Exam; QwenCloud pricing $0.16 per million input tokens and $0.47 per million output tokens; training cost roughly a ninth of Qwen3.7-Plus. Open source 5 eesel AI 2026-08-26 Flash Next is the self hosted open weights while the production Qwen3.8-Flash API runs on QwenCloud with a one million token default context; the 51B n-gram lookup table can be offloaded to system RAM; hosted pricing $0.15 input and $0.47 output per million tokens with cached reads about 89 percent cheaper; Artificial Analysis ranks it fifth for intelligence in class at 56 on its Intelligence Index; the model is very verbose; described as an intentionally under-trained research preview with GLM 5.3 Flash recommended for production. Open source
  • The GLM launch promotion expiry. Z.ai lists GLM-5.3-Flash at $0.15 and $0.50 per million tokens.6 MarkTechPost 2026-08-26 Z.ai released GLM-5.3-Flash on 26 August 2026: 320B total parameters, 18B active per token, 1,048,576 token context, native image and video input, MIT licensed weights on Hugging Face; 84.3 on Terminal-Bench 2.1 against Claude Opus 4.8 at 85.0 and 63.4 on DeepSWE v1.1 against GLM-5.2 at 46.2; hybrid linear and sparse attention, IndexPool cutting KV cache 4.4 times, Manifold-Constrained Hyper-Connections; API pricing $0.15 input and $0.50 output per million tokens; self hosting roughly 306 GiB of FP8 weights on Hopper class GPUs or newer. Open source Whether that list price holds through the fourth quarter, or drifts toward the promotional floor, tells you whether the price is a strategy or a launch tactic.
  • Activated parameter counts in the next release cycle. If the next round of open-weight flagships from either lab lands below 20B active per token while total parameters rise, sparsity is the settled design. A return to denser activation would mean the memory savings cost more accuracy than the reported numbers admit.1 Hugging Face, Qwen team model card 2026-08-26 125B total parameters with 6B activated per token, a 51B n-gram embedding layer and a 4B multi-token prediction module; 48 layers at hidden dimension 2560; hybrid of Gated DeltaNet and Qwen Sparse Attention with Gated Residual at bottleneck rank 320; Muon and AdamW optimizers; native context of 262,144 tokens extensible to one million; qwen-community-1.0 license; reported 62.5 on SWE-bench Pro, 73.9 on CoWorkBench, 84.5 on AndroidWorld. Open source 6 MarkTechPost 2026-08-26 Z.ai released GLM-5.3-Flash on 26 August 2026: 320B total parameters, 18B active per token, 1,048,576 token context, native image and video input, MIT licensed weights on Hugging Face; 84.3 on Terminal-Bench 2.1 against Claude Opus 4.8 at 85.0 and 63.4 on DeepSWE v1.1 against GLM-5.2 at 46.2; hybrid linear and sparse attention, IndexPool cutting KV cache 4.4 times, Manifold-Constrained Hyper-Connections; API pricing $0.15 input and $0.50 output per million tokens; self hosting roughly 306 GiB of FP8 weights on Hopper class GPUs or newer. Open source

The interesting question these two releases raise is not which one wins a coding benchmark. It is what happens to the economics of a proprietary API when the cheapest way to run a competent long context agent is a weights file you download once and serve on memory you already own.