Skip to content

DeepSeek V4 Flash 0731 r33 393K Primary promotion

Date: 2026-08-11

Evidence: local functional, capacity, and bounded quality

Decision: human-approved current for llm.primary; subsequent live ownership remains private operator state

Outcome

The digest-pinned r33 DeepSeek 0731 recipe became the routed text Primary at a 393,216-token window after a managed profile application proved exact model identity, readiness, and exclusive ownership of both RTX PRO 6000 cards. The profile keeps the quality-first choices: FP8 DS-MLA KV, DSpark K5, 4,096 batched tokens, 16 admitted sequences, 0.975 GPU utilization, and no CPU or filesystem KV offload. The engine reported 725,543 GPU KV tokens, or about 1.845 full configured windows.

OpenClaw's existing anvil/llm.primary model entry was updated from a 262,144 to a 393,216 context window and its generic 8,192-token output limit was raised to the quality-oriented 32,768-token ceiling. Config validation passed, the gateway completed safe restarts, and isolated post-restart basic-agent and real tool-loop requests both used provider anvil, model llm.primary, without fallback. The tool action was recorded as started and succeeded.

Exact promoted configuration

  • Model: deepseek-ai/DeepSeek-V4-Flash-0731, revision 9e165c30e2704aec5d9d593cce3eebd58bbef1cb
  • Served name: deepseek-v4-flash-0731-r33-b12x-dspark5-maxseq16-batch4096-tp2-393k
  • Image: sha256:fdde59fed7f9fc12f9fd5ef1b3b3ea8d5097bf10ebad54b348497102c3a83f82
  • Runtime: r33 B12X W4A8 mixed NVFP4-MoE/FP8, FP8 DS-MLA KV, DSpark fixed probabilistic K5
  • Hardware: 2x RTX PRO 6000 Blackwell Max-Q, exclusive TP=2 over PCIe without NVLink
  • Envelope: 393,216 context, max_num_seqs=16, max_num_batched_tokens=4096, utilization 0.975
  • Capacity tiers: GPU only; no KV offload and no reserve gate

The portable managed configuration is the 393K recipe.

Capacity and functional evidence

The calibrated direct context ladder passed at 238,507, 339,310, and 359,900 actual API prompt tokens. The largest request measured 65.2-second TTFT, 5,599 effective prefill tok/s, and returned 158 output tokens. A post-probe functional gate passed.

The routed preflight passed smoke, JSON, 3/3 typed tools, streaming tools, tool-result continuation, and the Responses API. Its legacy nominal-320K needle failed closed with HTTP 413 because the generated prompt was 1,800,115 bytes and the router's conservative bytes/4 estimate was 450,028 tokens—above the 393,216-token route limit. This is a benchmark-harness calibration mismatch, not a successful routed long-context result. The direct 359,900-token request proves model capacity above 300K; a calibrated routed context job is still required before claiming the same through the router.

The preflight's reasoning-evidence-required check also failed on trivial prompts that completed without a reasoning channel. The earlier repeated high-reasoning quality suite remains the quality evidence; this result is not silently converted to a pass.

OpenClaw verification

OpenClaw version 2026.7.1-2 retained agents.defaults.model.primary = anvil/llm.primary. The provider model now declares 393,216 context tokens and 32,768 maximum output tokens. After safe gateway restarts:

  • Run 2a0968e4-27b6-4652-a32e-dd466f3b6887 returned the exact readiness marker, reported a 393,216-token context budget, and used no fallback.
  • Run df38095f-1029-4d62-9ac3-ca7368e6d6d4 completed an actual workspace tool call, returned the expected marker, and used no fallback.
  • Session agent:main:config-393k-32k-smoke-20260811 returned the exact OPENCLAW_393K_32K_READY marker at high thinking after the output-cap correction; gateway config audit and RPC health passed.

These are functional client-path checks, not an OpenClaw 300K context test.

Hermes verification

Hermes Agent 0.20.0 already selected custom model llm.primary but retained a 262,144-token provider metadata entry, no explicit output cap, and persistent reasoning effort none. Its active default-model configuration now pins model.context_length=393216, model.max_tokens=32768, and agent.reasoning_effort=high. Hermes' own configuration command performed the write; the launchd messaging gateway restarted with zero active sessions.

An isolated one-shot request returned the exact HERMES_393K_32K_HIGH_READY marker through provider custom, model llm.primary, in one API call. This proves the corrected Hermes client path, not a Hermes request above 300K.

The operator then restarted Hermes for an upgrade. Version 0.20.0 retained all three settings, but the upgrade left the current launchd definition unloaded. hermes gateway start reloaded the matching service definition and restored launchd supervision with auto-start/auto-restart. A post-upgrade one-shot returned the exact HERMES_POST_UPGRADE_393K_32K_HIGH_READY marker in one API call.

Promotion and benchmark failures retained

The first lower-level promotion attempt exposed two product defects: the transaction supplied mutually exclusive reasoning controls to preflight, and its automatic recovery could not transition directly between exclusive serve owners. The stopped candidate was removed through managed lifecycle commands, then the declared dual-gpu-exclusive profile applied successfully and readmitted the route only after the expected and observed model identities matched. The defects are recorded in the promotion recovery ticket.

The requested remote-worker SWE smoke did not submit a benchmark job. Controller 0.33.1 accepted the typed dispatch but returned profile_unavailable because the installed wheel lacks the repository-level benchmark-profile JSON files. No SWE score exists for this promotion. The packaging gap is recorded in the benchmark profile ticket.

Scope and restoration boundary

This promotion proves exact managed routing, direct capacity through 359,900 actual prompt tokens, corrected OpenClaw and Hermes client budgets, and post-restart client paths. It does not prove routed, OpenClaw, or Hermes context above 300K, broad concurrency at 393K, or a SWE score. During the approved qualification, exclusive TP=2 kept other GPU inference offline. The post-session owner and operating mode remain private operator state rather than a public deployment claim.

The sanitized machine-readable record is promotion-summary.json.