Deepgram Self-Hosted September 2026 Release (260915)
By Steven Van ·
Self-hosted release 260915 requires NVIDIA driver 580+ (open kernel module) and adds Blackwell GPU support, Flux forced end-of-turn, and FIPS 140-3 images.
Deepgram's September 2026 self-hosted release (260915) raises the minimum NVIDIA driver requirement to 580 or later, and it must be the open kernel module flavor, for example nvidia-driver-580-open; the proprietary build isn't supported. The Engine image is now built against CUDA 13, whose official NVIDIA support begins on the 580 driver branch. On Blackwell GPUs the effect is total: NVIDIA never added Blackwell support to the proprietary 580 branch, so on that build a Blackwell card disappears from the system, nvidia-smi reports no devices, and Engine won't start.
The release also includes:
- General availability of NVIDIA Blackwell GPU support for self-hosted deployments
- Flux Forced End-of-Turn, letting a client end a /v2/listen turn on demand rather than waiting for the model to detect end of speech, via listen_v2_force_end_turn and a ForceEndTurn message
- Mid-stream numeral configuration for Flux, so numeral formatting can be changed during an active /v2/listen stream with a Configure message instead of only at connection time
- Improved reading of German numbers, currency, times, and addresses
- FIPS 140-3 images for all four self-hosted services (api, engine, license-proxy, billing), tagged with a -fips suffix and running OpenSSL's FIPS 140-3 validated cryptographic provider