Skip to content

Heavy intelligence challengers: Mistral Small 4 and Nemotron 3 Super

Point-in-time record, 2026-07-12. Two current, official Hugging Face checkpoints were served one at a time on Primary Node's RTX PRO 6000 and run through the same Heavy correctness, context, independent-sanity, and five-session workloads. nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4 is the selected resident experiment because it passed the complete built-in gate and the thinking-enabled tie-break. This is not a router-profile promotion.

Raw artifacts are under 2026-07-12-heavy-intelligence-challengers-evidence/.

Candidate prior and source freshness

Candidate Official source observed 2026-07-12 Source age Why it entered the local gate
mistralai/Mistral-Small-4-119B-2603-NVFP4 Hugging Face model card current, within 60 days Apache-2.0, 119B total / 6.5B active, 256K advertised context, unified instruct/reasoning/coding behavior, official NVFP4
nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4 Hugging Face model card release was about 123 days old, stale under the 120-day research rule; model page remained current 120B total / 12B active hybrid model, configurable thinking, up to 1M advertised context, official NVFP4

The external claims above were candidate-selection priors only. They did not affect the local pass/fail result. The imported llmrequirements.com database, dated 2026-07-11, was also used as an advisory fit prior; it was not treated as hardware-matched benchmark evidence.

Reproducible single-card recipes

Both checkpoints used pinned cache revisions, vLLM nightly 0.23.1rc1.dev531+ga65f93fb2, FP8 KV cache, a 131,072-token served limit, and a five-sequence admission cap. Prefix caching was disabled so independent-prompt concurrency results could not be inflated by reuse.

The managed recipes default --revision to the tested commits. The vllm/vllm-openai:nightly image tag remains mutable, so the observed engine version is recorded evidence rather than a future engine pin. That drift risk is another reason the selection remains experimental.

Candidate Pinned revision Cached checkpoint Engine details KV capacity at 131K
Mistral Small 4 d57a94c74a961e1f9b489b8b3e792923ca29149b 66.0 GiB TRITON_MLA, Mistral reasoning/tool parsers, text-only 2,141,968 tokens; 16.34 full-window equivalents
Nemotron 3 Super 4f0cf9daaeb7a4d5e23f80a00e7ed15f0e03caf6 74.8 GiB official super_v3 reasoning parser, qwen3_coder tool parser, mamba_cache_dtype=float16, text-only 3,999,467 tokens; 30.51 full-window equivalents

The vendored super_v3 parser was taken from the official NVIDIA checkpoint repository at the same pinned revision.

These capacity figures describe the engine's shared KV pool. They do not mean that every admitted session can grow to the model-card maximum simultaneously. Neither candidate's advertised maximum was validated: this round deliberately held the served and correctness-tested context at 131K.

Preflight and current Heavy eval

Nemotron passed the exact final recipe's preflight: short generation, structured JSON, a 131K needle, and 20/20 tool calls. It then passed the built-in context, tool, session-recall, and both deterministic intelligence checks with zero failures. Its 131K request measured 16.76 s TTFT and 17.50 s end to end.

Mistral also passed preflight without a thinking override, including 20/20 tool calls and the 131K needle. The first attempt with the harness's Qwen-style chat_template_kwargs failed closed with HTTP 400 because Mistral tokenizers do not accept that field. The model card instead specifies the OpenAI reasoning_effort request field; the current harness does not expose it. On the final no-prefix-cache recipe, Mistral failed both built-in intelligence checks: the edit was not a valid minimal unified diff and the timeout recommendation did not satisfy the deterministic fix check. A prior run passed one of two, so the default sampling behavior was also not stable enough for selection. Its 131K request measured 51.90 s TTFT and 52.55 s end to end.

Independent Hugging Face sanity slices

The no-thinking comparison used the pinned ten-row allenai/ai2_arc ARC-Challenge fixture already documented in the Qwen bakeoff. Mistral scored 9/10; Nemotron scored 7/10. Nemotron then scored 5/5 on the separate thinking-enabled slice with a 1,024-token completion cap, versus the earlier ThinkingCap Qwen result of 4/5.

Candidate / mode Correct visible finals Median request latency Wall time
Mistral Small 4, default non-reasoning behavior 9/10 - 1.18 s
Nemotron 3 Super, thinking disabled 7/10 - 2.62 s
Nemotron 3 Super, thinking enabled 5/5 4.35 s 21.26 s
ThinkingCap Qwen3.6-27B FP8, thinking enabled (prior comparison) 4/5 6.69 s 33.49 s

These are small, one-shot exact-marker sanity tests. They are useful for finding obvious regressions and reasoning starvation, not for claiming general model quality. The artifacts still do not retain finish reason or a distinct hidden- reasoning token budget, so the result is not promotion evidence and does not repair the known cross-model eval protocol gap.

Five-session result

Each concurrency run sent five independent 8,192-token prompts. All requests completed at both admission settings. The prompts elicited short outputs, so aggregate output tok/s is an operational batch measure rather than a controlled long-generation decode rate.

Candidate Concurrency 1: TTFT / E2E p50 Concurrency 5: TTFT / E2E p50 Aggregate output tok/s, 1 / 5 Completed
Mistral Small 4 0.30 / 0.58 s 1.85 / 2.46 s 57.82 / 67.04 5/5, 5/5
Nemotron 3 Super 0.62 / 1.02 s 2.52 / 3.63 s 33.19 / 45.90 5/5, 5/5

Five sessions fit because vLLM schedules them against a shared KV block pool; the five-sequence cap is admission control, not five pre-reserved maximum-length contexts. RadixAttention or prefix caching could improve workloads with shared prefixes, but these prompts were independent and prefix caching was deliberately disabled for a clean comparison.

Decision

Nemotron 3 Super replaces ThinkingCap Qwen3.6-27B as the best currently validated Heavy experiment. Mistral is the lower-latency candidate for short independent prompts, but its current Heavy intelligence failures and unsupported harness reasoning control outweigh that advantage. Nemotron passed the full built-in gate, produced the best bounded thinking result measured in this round, and remained healthy at five concurrent sessions.

The selected direct endpoint is nemotron3-super-120b-a12b-nvfp4 on http://127.0.0.1:39033/v1. The production Heavy router profile, quality calibration, and promotion state remain unchanged pending a separate human gate.