Mistral Medium 3.5¶
Current status and review date¶
Decision snapshot
- Product role: unqualified replacement candidate.
- Selected or best-qualified configuration: none.
- Measured hardware: two RTX PRO 6000 Blackwell Max-Q GPUs in TP2.
- Evidence: bounded protocol and native reliability gates; the dated finding holds the sole metric table.
- Decision:
no-promotion; no route or performance ranking follows. - Important limitation: configured context is not full-context or concurrent-capacity evidence.
- Review dates: 2026-10-15.
Review narrative¶
r2 passed protocol but failed the campaign resource policy. r3 cleared the generic native floor but missed the frozen campaign reliability requirement. These are final gate outcomes and neither qualifies a replacement; exact GLM was retained. See the replacement evaluation.
Immutable identity¶
mistralai/Mistral-Medium-3.5-128B@22b2b868a15677cfa6061277ed2f653d1349a9ab- r3 served identity:
mistral35-fp8kv-tp2-c4-262k-r3 - image:
vllm/vllm-openai@sha256:cb9bc64f0f038c2c2ebf13b31248fb87aca40c356c3d20f363be087cfcff1fcd
Tested hardware and topology¶
Two RTX PRO 6000 Blackwell Max-Q GPUs in direct TP2 qualification. This is measured hardware for the r2/r3 gates, not a claim about a live deployment.
Engine, quantization, KV, context, and concurrency recipe¶
vLLM with FP8 KV; r3 declares 262,144 configured tokens, C4 admission, and up to eight images. It is not measured full-context, C4, multimodal, or performance evidence.
Evidence by measurement class¶
Resource-policy gate¶
r2 was rejected by resource policy despite protocol completion.
Native reliability gate¶
r3 missed the frozen campaign reliability gate. The dated finding contains the authoritative measured values.
Decision and promotion state¶
no-promotion. No configuration, load, health check, or benchmark result
changes a serve, route, or client catalog.
Failures and gotchas¶
A generic native-floor pass does not override the campaign reliability gate. Configured context and C4 do not prove capacity.
Dated run history¶
- 2026-10-15 — replacement evaluation: r2 resource-policy and r3 reliability outcomes.