Ledger on GitHub. Hub here. Branches elsewhere.
Jalapeño First Silicon: 700W Broadcom ASIC Posts 1.9× Efficiency Gain Over GB300
Jalapeño First Silicon: 700W Broadcom ASIC Posts 1.9× Efficiency Gain Over GB300
OpenAI published first results for Jalapeño, a Broadcom co-designed inference ASIC built for dense transformer inference at scale.
The chip delivered 1.5–1.9× more AI work per watt and 1.7–3.6× lower end-to-end latency than Nvidia GB200/GB300 in the published comparison.
Package power is 700W. The die carries 216GB of HBM4 with 15.4 TB/s aggregate bandwidth.
OpenAI benchmarked the ASIC across three open-weights models: GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T. The benchmark set supports the inference-specific design claim rather than serving as a general-purpose GPU replacement story.
The results mark the first public performance data from a hyperscaler-commissioned inference ASIC that shows measurable latency and efficiency separation from the current Blackwell generation.
Jalapeño’s design focused on transformer decode rather than training or embedding workloads. The HBM4 choice and package power envelope suggest the ASIC targets rack density and thermal limits that constrain GPU deployments in large-scale inference clusters.
Open-Weights MoE Releases: Context and Active Parameter Comparisons
Three open-weights mixture-of-experts models surfaced in August with different active parameter and context strategies.
| Model | Total Parameters | Active Parameters | Native Context | Benchmark | Score | URL | |---|---|---|---|---|---| | Qwen3.8-2.4T-A95B | 2.4T | 95B | 262,144 (extends to ~1,010,000) | Terminal Bench 2.1 | 86.6 | https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B | | Qwen3.8-2.4T-A95B | 2.4T | 95B | 262,144 | SWE-bench Pro | 67.7 | https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B | | DeepSeek V4-Pro 0813 | 1.7T | 49B | 262K | Terminal-Bench 2.1 | 87.9 | https://morningmail.ai/en/newsletter/ai-developer-tools-what-shipped-kncCRQDENxYjgBnSFQK5/2026-08-16 | | DeepSeek V4-Pro 0813 | 1.7T | 49B | 262K | DeepSWE | 62.7 | https://morningmail.ai/en/newsletter/ai-developer-tools-what-shipped-kncCRQDENxYjgBnSFQK5/2026-08-16 | | GLM-5.3 | 753B | 40B | 1M | — | — | https://fruition.net/frontier/ |
Qwen3.8-2.4T-A95B is the largest total-parameter model in the set, with 2.4T total and 95B active. The documented context extends from 262,144 native tokens to roughly 1,010,000 tokens through extension techniques.
Benchmark results on Terminal Bench 2.1 and SWE-bench Pro are solid but not dominant. The model sits in the middle of current open-weights performance rather than at the top.
DeepSeek V4-Pro 0813 uses 1.7T total parameters and 49B active. Native context is 262K. Terminal-Bench 2.1 score is 87.9, slightly above Qwen3.8’s 86.6. DeepSWE score is 62.7.
DeepSeek introduced peak/off-peak API pricing tiers alongside V4-Pro. The pricing change reflects load-balancing incentives for inference providers facing uneven utilization patterns.
GLM-5.3 is a 753B total, 40B active open-weights family with 1M native context. The model documentation emphasizes agentic coding and cyber-defense tasks.
GLM-5.3 reportedly surfaced a serious vulnerability in Cursor during cybersecurity testing. The claim suggests the model’s capabilities extend to probing developer-tool attack surfaces, though the severity and remediation timeline are not in sources.
Inference Stack Notes: Z.ai Token Volume and Meta Muse Glimmer
Z.ai’s inference stack reportedly served 62T AI tokens on sanctioned Chinese chips. The volume claim indicates sustained production inference at scale, though architecture details, chip generation, and throughput per node are not in sources.
Meta released Muse Glimmer, described as a 30B Apache 2.0 agentic model designed to run locally on consumer GPUs. The license and parameter count position the model for on-device or single-GPU deployment rather than cluster inference.
Three Practitioner Lines
Act: Benchmark Jalapeño ASIC efficiency claims against your current GB200/GB300 inference cost structure if you operate decode-heavy workloads above 100B active parameters. The latency and power numbers matter most in multi-rack deployments where thermal and bandwidth limits constrain GPU density.
Watch: Track context-extension techniques in Qwen3.8 and GLM-5.3. The gap between native (262K) and extended (~1M) context windows is large enough to change task feasibility for document analysis and multi-turn agent loops, but throughput and accuracy under extension are not yet documented in public benchmarks.
Ignore: Hype around “agentic” labels without task-specific benchmarks. Muse Glimmer’s 30B parameter count and Apache 2.0 license are concrete; claims about agentic behavior require reproducible task definitions and published results before they inform deployment decisions.