Serve chat without forced thinking

Test think prefill with pinned SFT epoch 3 revision

Fix Unicode characters in streamed responses

Compact chat context at 1024 tokens

Deploy PIQA-selected full-SFT epoch 2 root model

Download only refinement inference files

Deploy MoE architecture metadata fix

Serve the 160B refinement model as raw completion

Deploy TR-HASH MoE 200M 160B-token Thinking LoRA

Serve intermediate TR-HASH-200M-70B-SFT checkpoint

Serve AETHORIA TR-HASH-200M-130B

Bump TR-Hash-i64 to 8dd8256

Bump TR-Hash-i64 pin to the multi-hash routing loader fix

Bump tr_hash_i64 pin to 9f9766b (MPSGraphRunner, INL - 2025 cleanup)

Bump tr_hash_i64 pin to 8b9d77a (checkpoint path resolution fix)

Bump tr_hash_i64 pin to f79196f (MPS support, 200M registry entry)

Bump tr_hash_i64 pin to 8040970 (decode_step CUDA graph fix, hash-routing audit)

Sync with the renamed TR-Hash-i64 server (was vllm-i64): repo URL, module import, env var names

Reload optional-thinking chat template

Rebuild with bounded thinking finalization

Rebuild with vllm-i64 a44a472

Deploy reasoning-history context isolation

Isolate chat prefix cache by conversation

Isolate prefix cache by conversation

Do not charge telemetry against chat rate limit

Fix TR-Hash Space model identifier

Rename public demo to TR-HASH-0.5B

Apply embedded Hugging Face chat template

Mount TR-Hash 500M model volume

Serve TR-Hash MoE 500M

Rebuild with vllm-i64 71e005a

Add live top-2 expert telemetry e787df1

Rebuild with rolling context manager e772ddf

Pin vllm-i64 dynamic model metadata

Deploy TR-MOE-306 with vllm-i64