Flash Next 47 Run Evidence

Flash Next 47 tok/s: published run evidence

This appendix preserves the generated text and exact measurements from the September 7, 2026 public-build verification. See Notable runs for the current headline results.

Measurement record

source_commit=b46f7f7a436f990932d3da3ec53380e2b9effc89
public_branch=moe-cache
build_source_clean=true
model=Qwen3.8-Flash-Next-UD-Q3_K_XL
model_revision=38bb39ee97821de2c9009abb7e93950eec396e66
expert_cache_slots=80
context_capacity=12288
batch=4096
ubatch=512
parallel=1
kv_key=q8_0
kv_value=q8_0
speculative_decoding=none
predictive_expert_prefetch=absent
grouped_frequency_eviction=default_enabled
profiling=none
temperature=0
seed=12345
prompt_tokens=158
completion_tokens=1024
total_tokens=1182
finish_reason=length
server_prompt_ms=1657.688
server_prompt_tok_s=95.3134727403468
server_decode_ms=21767.858
server_decode_tok_s=46.99589642674075
client_ttft_ms=1682.9886990017258
client_decode_seconds=21.76632235397119
client_decode_tok_s=46.99921205629656
client_request_wall_seconds=23.476998583995737
loaded_device_vram_mib=13908
peak_device_vram_mib=14384
minimum_system_mem_available_mib=2290.46
server_peak_rss=not_recorded
grouped_registered=48
grouped_covered=48
grouped_ready=49104
grouped_completed=49104
grouped_fallback=0
grouped_rollback=0
grouped_prepare_error=0
grouped_finish_error=0
logical_grouped_h2d_banks=292737
logical_grouped_h2d_bytes=223657574400
server_error_lines=0
client_error=none
health_ok=true
teardown_process_clear=true
teardown_port_clear=true
teardown_gpu_clear=true

The transfer counter describes logical expert-bank bytes, not a measurement of PCIe wire traffic or transfer duration. The RAM minimum is a system-level resource observation, not the server's private allocation.

The following hashes cover the concatenated UTF-8 stream strings, without the Markdown fences used to display them here:

content_sha256=b6f509bedcdab739b071b42c666ca844b1acfacaec45172c2c3881c3c9cd13d1
reasoning_sha256=18af1c0b8727d8c8790aefc38fcca3ee578ef12498db02b962d00d2f60a2e62e

Reasoning output

The server streamed this text through the model's reasoning field.

We need answer user's request. Need act as release operator. Need first reason through deployment hazards, dependencies, rollback criteria. Then explicitly finish reasoning and produce final executable rollout plan dominated by shell commands and config snippets. Deploy payments API to blue canary pool, hold traffic at 10%, verify latency and error budgets, publish signed go-or-rollback decision. Keep reasoning brief enough to leave most response budget for final plan. Include concrete scripts with error handling, config examples, verification commands rather than only prose.

Need likely produce final answer with brief reasoning then final plan. Need maybe mention assumptions: Kubernetes, Istio/Envoy? Blue canary pool. Need executable shell commands and config snippets. Could use kubectl, helm, istioctl, curl, openssl signing. Need signed go-or-rollback decision. Could create decision JSON, sign with openssl dgst using private key, verify. Need maybe environment variables. Need error handling set -euo pipefail. Need deploy to blue canary pool, hold traffic 10%. Need verify latency and error budgets. Could use Prometheus queries, kubectl rollout status, istio VirtualService weights. Need rollback criteria. Need maybe use blue/green/canary: blue canary pool? Deploy payments APINow give the requested answer and finish all requested sections.

Answer output

The server streamed this text through the answer field. The token ceiling truncates the response; it does not finish all requested sections. These are model-generated examples, not a reviewed production deployment script. Do not execute them against a live cluster.

```markdown
## Release Operator Reasoning

Primary hazards:
- Payments API is stateful and transaction-sensitive; bad deploys can cause double charges, failed captures, or webhook loss.
- Blue canary pool must not receive production traffic until health, schema compatibility, and downstream dependencies are verified.
- Latency and error-budget regressions can be caused by database migrations, cache misses, cold JVM/Node pools, or third-party PSP latency.
- Rollback must be fast and reversible without requiring a new build.

Dependencies:
- Kubernetes cluster with `payments` namespace.
- Blue canary deployment: `payments-api-blue-canary`.
- Istio or equivalent traffic management for weighted routing.
- Prometheus/Grafana for latency and error-rate verification.
- PostgreSQL, Redis, PSP gateway, webhook consumer, and feature flags.
- Signing key for release decision artifact.

Rollback criteria:
- HTTP 5xx rate > 0.5% over 5 minutes.
- p95 latency > 350 ms over 5 minutes.
- Payment success rate < 99.5% over 5 minutes.
- Pod restarts, CrashLoopBackOff, readiness failures, or DB migration errors.
- Any duplicate-charge or webhook-loss signal.

Decision:
- If all canary checks pass, sign `GO`.
- If any check fails, sign `ROLLBACK` and shift traffic back to stable immediately.

---

## Final Executable Rollout Plan

### 0. Environment and Safety Setup

```bash
#!/usr/bin/env bash
set -euo pipefail

# -----------------------------------------------------------------------------
# Release operator rollout script for payments API blue canary
# -----------------------------------------------------------------------------

export RELEASE_ID="${RELEASE_ID:-$(date -u +%Y%m%dT%H%M%SZ)}"
export APP_NAME="payments-api"
export NAMESPACE="payments"
export BLUE_CANARY_DEPLOYMENT="payments-api-blue-canary"
export STABLE_DEPLOYMENT="payments-api-stable"
export VIRTUAL_SERVICE="payments-api"
export CANARY_PERCENT="10"
export STABLE_PERCENT="90"
export CANARY_WINDOW_SECONDS="300"
export PROM_URL="${PROM_URL:-http://prometheus.monitoring.svc.cluster.local:9090}"
export DECISION_DIR="${DECISION_DIR:-./release-decisions}"
export SIGNING_KEY="${SIGNING_KEY:-./release-signing-key.pem}"
export SIGNING_CERT="${SIGNING_CERT:-./release-signing-cert.pem}"
export KUBECONFIG="${KUBECONFIG:-$HOME/.kube/config}"

mkdir -p "$DECISION_DIR"

log() {
  printf '[%s] %s\n' "$(date -u +%Y-%m-%dT%H:%M:%SZ)" "$*"
}

fail() {
  log "ERROR: $*"
  exit 1
}

require_cmd() {
  local cmd="$1"
  command -v "$cmd" >/dev/null 2>&1 || fail "Required command not found: $cmd"
}

require_env() {
  local var="$1"
  [[ -n "${!var:-}" ]] || fail "Required environment variable not set: $var"
}

require_cmd kubectl
require_cmd istioctl
require_cmd curl
require_cmd jq
require_cmd openssl
require_cmd awk

require_env RELEASE_ID
require_env SIGNING_KEY
require_env PROM_URL

log "Starting rollout for release ${RELEASE_ID}"
```

---

###

Reproduce and tune this run - 16 GB setup guide