Pokee-Isaac 28B Launches: 10M Context on Single GPU, Tops BFCL v4
Startup Pokee AI has released the Pokee-Isaac 28B model. Featuring a new non-decoder-only architecture, it focuses on balancing massive context length with agentic capabilities at a minimal parameter size. It can be deployed on a single GPU (minimum RTX 4090) with an input cost of just $0.15/M token. Inference framework SGLang has announced day-one deployment support.
已确认
- 超长上下文: Supports an effective context of 1000 万 tokens, maintaining a 93.3% RULER score at full length.
- 智能体能力: Topping the leaderboard in the BFCL v4 agentic evaluation, Pokee-Isaac 28B scored 70.94, surpassing GPT-5.6-luna (70.61 分).
- 推理支持: SGLang officially announced day-one deployment support, achieving specific optimizations on a single B200 GPU.
为什么重要
- With a relatively small parameter size of 28B, Pokee-Isaac 28B outperforms industry-leading massive models in agentic capabilities, proving the potential of its new architecture in handling complex tasks.
- Compressing the requirement for a 1000 万 token context to run on a single consumer or enterprise GPU (like RTX 4090 or B200) drastically lowers the hardware barrier for ultra-long text processing, offering immense practical value for agent applications requiring massive context.
2026-08-05 ~ 2026-08-05 · 5 related posts
Primary sources
- [source] Pokee Launches 28B Agentic Model with 10M Token Context, Runs on a Single RTX 4090 — Scobleizer · 2026-08-05
- Pokee-Isaac 28B Launches: 10M-Token Context on a Single GPU — Kyrannio · 2026-08-05
- [source] SGLang Announces Day-0 Support for Pokee-Isaac 28B Deployment — Kyrannio · 2026-08-05
- [source] 28B Model Upsets Big Tech: Pokee-Isaac Takes #1 in BFCL v4 Agentic Capabilities — Kyrannio · 2026-08-05
1 near-duplicate retellings: Kyrannio