dropkit.contents

Ledger on GitHub. Hub here. Branches elsewhere.

← Drops · 2026-09-04

Jalapeño 1.9×, Hy4 770B, and InferenceBench's 15× price spread

Jalapeño 1.9×, Hy4 770B, and InferenceBench’s 15× price spread

OpenAI published benchmark results for its Jalapeño inference chip, showing 1.5×–1.9× higher inference performance than Nvidia’s GB300 and up to 3.6× better work-per-watt across three open models.

The test suite was SemiAnalysis’ public InferenceX benchmark, run against GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5.

No training performance data appeared in the published results.

Tencent released Hy4 preview, a 770B-parameter open model with 49B active parameters and a 1M-token context window.

Hy4 uses sparse mixture-of-experts routing: 256 routed experts plus one shared expert. Only 21 of the model’s 78 layers compute a sparse routing index, and each query attends to 2048 tokens.

The model ships under Apache 2.0.

Tencent’s internal rubric gave Hy4 preview a score of 2.99, versus 2.92 for GLM-5.3 and 2.94 for Kimi K3.

InferenceBench published research tracking the same GPT-OSS 120B weights deployed across providers. Median time-to-first-token ranged from 263 ms to 2,036 ms. Output pricing ranged from $0.17 to $0.70 per million tokens.

A second InferenceBench summary, covering a broader set of providers serving GPT-OSS 120B, reported median TTFT from 111 ms to 545 ms and output prices from $0.45 to $17.50 per million tokens.

The gap between identical weights on different infrastructure was fifteen-fold in one sample and thirty-fold in the other.

One analysis piece described Chinese open models as dominating the 2026 summer scale frontier. The monthly ceiling for Chinese releases ran from 754B to 2.78T parameters. U.S. open models stayed under 130B in five of the seven months surveyed.

DeepSeek V4-Pro-0813 was reported as an MIT-licensed open-weight model with approximately 1.6T parameters, 49B active parameters, and a 1M-token context window.

Moonshot’s Kimi K3 was described as a 2.8T-parameter open-weight model and called the largest open-weight release.

Z.ai’s GLM-5.3-Flash was described as a 320B multimodal mixture-of-experts model with 18B active parameters, 1M-token context, MIT license, and training on Chinese AI chips.

Meta’s Muse Glimmer was described as a 30B Apache 2.0 agentic model runnable on a single 24GB consumer GPU.


Model or chipLabelScale / performanceURL
OpenAI Jalapeñoinference chip1.5×–1.9× vs. GB300; up to 3.6× work-per-watt; 3 modelstechradar.com
OpenAI Jalapeño benchmark suiteInferenceXGPT-OSS 120B; DeepSeek R1 670B; Kimi K2.5aimastery.page
Tencent Hy4 previewopen MoE770B params; 49B active; 1M contexttencent.com
Hy4 routingsparse MoE256 routed experts + 1 shared; 21/78 layers; 2048 tokensx.com/TencentHunyuan
Hy4 preview rubricinternal score2.99 (Hy4) vs. 2.92 (GLM-5.3), 2.94 (Kimi K3)datanorth.ai
InferenceBench: GPT-OSS 120Bprovider spread (sample 1)TTFT 263–2,036 ms; output $0.17–$0.70/M tokensinferencebench.dev
InferenceBench: GPT-OSS 120Bprovider spread (sample 2)TTFT 111–545 ms; output $0.45–$17.50/M tokensinferencebench.dev
Chinese open-model frontier2026 summermonthly ceiling 754B–2.78T; U.S. under 130B in 5/7 monthscgtn.com
DeepSeek V4-Pro-0813open MoE~1.6T params; 49B active; 1M context; MITthursdai.news
Moonshot Kimi K3open weights2.8T params; largest open releaseforbes.com
Z.ai GLM-5.3-Flashmultimodal MoE320B; 18B active; 1M context; MIT; Chinese chipsrequesty.ai
Meta Muse Glimmeragentic model30B; Apache 2.0; runs on 24GB consumer GPUrequesty.ai

Act: Benchmark the same open-weight checkpoint on two inference providers and measure TTFT plus output pricing before signing annual contracts.

Watch: Hy4 preview’s 256-expert routing and the 21-layer sparse schedule; routing topology matters when context reaches 1M tokens and active parameters stay below 50B.

Ignore: Headlines that compare parameter counts without stating active parameters, context length, and license; the table already shows total and active separated by four to fifty-seven fold.