Ledger on GitHub. Hub here. Branches elsewhere.
Qwen3.8-Max ships at 2.4 trillion sparse parameters; DeepSeek V4-Flash adds vision with no surcharge; OpenAI Jalapeño chip targets inference speed
Qwen3.8-Max ships at 2.4 trillion sparse parameters; DeepSeek V4-Flash adds vision with no surcharge; OpenAI Jalapeño chip targets inference speed
Qwen3.8-Max reached open weights on August 12–13, 2026. The architecture is a 2.4 trillion parameter sparse mixture-of-experts model that activates 95 billion parameters per forward pass. The system designates active compute at routing time rather than loading the entire parameter set for each token.
The 27-billion dense variant in the same family shipped with 262,144 tokens of native context. That window extends to one million tokens when YaRN positional interpolation is applied. The method recomputes position embeddings to support sequence lengths beyond the pretrain horizon without a full retrain cycle.
Open weights became available through the usual repository channels. Documentation and benchmark scripts went live alongside the checkpoint files. No restricted license appears in the release notes examined.
DeepSeek made an experimental vision branch of V4-Flash available through the API. DeepSeek V4-Flash-Vision-Exp adds image understanding to the existing text and agent capabilities that shipped in the earlier V4-Flash release. The vision branch handles image inputs in the same request stream as text tokens.
Pricing for V4-Flash-Vision-Exp was described as matching the text-only tier with no additional charge for vision processing. That structure removes the surcharge that multimodal APIs have historically carried for image encoding. The experimental label indicates that parameter counts, context limits, and reliability margins have not yet been frozen.
OpenAI’s Jalapeño inference chip appeared in reporting as a custom silicon project focused on fast-inference workloads rather than training. Benchmark figures cited in the articles placed Jalapeño performance at 1.5–1.9 times NVIDIA GB200 speeds and 1.7–3.6 times GB300 speeds under the test conditions described. The comparison reflects throughput in a narrow set of inference tasks, not general FLOPS or memory bandwidth for all model types.
Deployment volume was reported as very small at the end of 2026, with larger-scale rollout planned for 2027. The chip does not replace training infrastructure; the design targets serving latency and token-per-second throughput once weights are frozen. The articles characterized Jalapeño as an effort to reduce dependency on GPU supply for inference scaling.
Table
| Item | Detail | Numbers | URL |
|---|---|---|---|
| Qwen3.8-Max | 2.4T sparse MoE, 95B active parameters | 2.4T; 95B | https://www.digitalapplied.com/blog/qwen3-8-max-full-release-benchmarks-open-weights |
| Qwen3.8-27B | 262K native context, extends to 1M via YaRN | 262K; 1M | https://www.vorplabs.com/models/releases/qwen-3-8-max |
| Qwen3.8 open weights | Released August 2026 | Aug 12; Aug 13 | https://www.aireiter.com/blog/qwen-3-8-max-open-weights |
| DeepSeek V4-Flash-Vision-Exp | Experimental multimodal vision API | experimental | https://api-docs.deepseek.com/updates/ |
| DeepSeek V4-Flash-Vision-Exp | Image understanding added to V4-Flash text/agent capability | multimodal | https://api-docs.deepseek.com/updates/ |
| DeepSeek V4-Flash-Vision-Exp pricing | Text-only price with no vision surcharge | no surcharge | https://docs.apiyi.com/en/live/2026-08/deepseek-v4-flash-vision-exp |
| OpenAI Jalapeño chip | Fast-inference design, 1.5–1.9× GB200 / 1.7–3.6× GB300 in benchmarks | 1.5–1.9×; 1.7–3.6× | https://www.forbes.com/sites/luisromero/2026/08/27/openais-jalapeo-chip-isnt-hot-and-thats-a-good-thing/ |
| OpenAI Jalapeño deployment | Small volumes end-2026, larger rollout 2027 | end-2026; 2027 | https://www.techcrunch.com/2026/08/25/openais-jalapeno-chip-is-built-for-fast-inference-at-scale-benchmarks-show/ |
Three practitioner lines
Act: Test Qwen3.8-Max at 95B active parameters if your workload can route through a sparse MoE and your memory budget does not accommodate a 2.4T dense model. Evaluate DeepSeek V4-Flash-Vision-Exp for multimodal tasks where removing the vision surcharge changes cost structure enough to justify switching APIs.
Watch: Monitor Jalapeño chip deployment announcements in 2027 for evidence that custom inference silicon can match or exceed GB200/GB300 throughput at lower cost or power, since that outcome would shift procurement assumptions for inference clusters serving frozen weights.
Ignore: Marketing claims about “breakthrough” context windows or “revolutionary” vision pricing unless the documentation includes token-per-second numbers, error rates at the stated context length, and cost-per-token tables that allow direct comparison with the existing API you run today.