DeepSeek V4.1 Flash: 890 bytes/token KV cache, 552B MoE beats V4 Pro on agentic tasks
togethercompute · x · 2026-09-13
Together AI has launched DeepSeek V4.1 Flash, with full specs now published:
- Architecture: a multimodal MoE with a Causal Encoder-Decoder design (20 encoder + 20 decoder layers), 552B backbone plus 196B Engram lookup memory; only 8B parameters activate during prefill and 16B during decode, slashing cost for input-heavy agentic workloads.
- KV cache: Compressed Sparse Attention 2 plus FP4 KV caching shrink the global cache to 890 bytes per token — roughly a quarter of DeepSeek V4 Flash — across a 1M-token context.
- Capabilities: natively multimodal, trained from scratch on 45T tokens; reasoning effort continuously adjustable 1–100 per request; leads DeepSeek's published agentic suite, outperforming the much larger V4 Pro, and reportedly beats GPT-5.6 Sol on agentic benchmarks at one-third the cost per task.
- Availability: released under the MIT license, available on Together AI (serverless and dedicated) with a 99.9% SLA.
More from coding & agent
- astra is great at CAD itself, but telling it what to build remains the hard part — yacineMTB · 2026-09-13
- Yacine designs drone entirely from phone; custom NPU kernels max out chip without heatsink — yacineMTB · 2026-09-13
- GPT-Live-1 first impressions: most natural voice model yet, but instruction following is unreliable — kolchinski · 2026-09-13
- Open-source 'no-ai-slop' removes 20+ AI writing clichés, hits 8.7k stars — tom_doerr · 2026-09-13
- Devin SWE-2 first impressions: fast and cheap, but shorter endurance than Codex — CtrlAltDwayne · 2026-09-13
- Roboflow ships auto-label with GPT-6 Astra, making it the new default model — Roger_M_Taylor · 2026-09-13