Apriel-H1

Apriel-H1

Why efficiency-optimized reasoning matters now

Torsten Scholak — Lead Research Scientist

SLAM Lab — ServiceNow

November 2025

Efficiency = Capability

Apriel matches frontier reasoning at 15B.
But full attention pays the quadratic tax → long ctx throughput is now the bottleneck.

Speed creates capability:

  • Agents keep full tickets/logs in memory → fewer compactions
  • More tools per turn at same latency
  • Deeper reasoning chains with more steps → better accuracy
  • Larger RAG contexts stay in-context
  • Higher req/s on existing fleet → lower unit cost, better UX

That's why we're building efficient hybrids.

How Hybrids Work

Full Attention Efficient (Linear/Sparse) Hybrid
Complexity O(n²) O(n) or sub-quadratic Mixed
KV cache Large, grows with n² Small or none Reduced ~50-75%
Global fidelity Perfect Limited Preserved in key layers
Throughput gain 1× 2-10× (but quality risk) 2-10× at minimal Δ

Pattern: Keep ~20-30% full attention for global reasoning,
replace rest with Mamba/linear/sparse mechanisms.

Hybrids Are Shipping at Scale

  • Apr: NVIDIA Nemotron-H-47B
    9:1 Mamba-2:FA hybrid, ≈3× faster vs dense 70B at long ctx

  • May: Falcon-H1-34B
    Parallel Mamba-2 + FA hybrid, 4× prefill, 8× decode at long ctx

  • Jun: MiniMax-M1
    7:1 Lightning:FA hybrid, ≈3–4× faster decode @100k tokens

  • Aug: Nemotron-Nano-9B-v2
    7:1 Mamba-2:FA hybrid, up to 6× throughput vs Qwen3-8B

  • Sep: Qwen3-Next-80B-A3B
    3:1 Gated-DeltaNet:FA hybrid, >10× throughput vs Qwen3-32B @>32k

  • Sep: DeepSeek V3.2-Exp
    MLA+DSA sparse, 1:64 attended:total tokens @128k, 3× faster at long ctx

  • Oct: Kimi-Linear-48B-A3B
    3:1 KLA:FA hybrid, 75% KV↓, up to 6× decode @1M

Apriel-H1

What You Get

Apriel-H

Today we release Apriel-H1:

  • Hybrid reasoner distilled from Apriel-15B
  • ~2× throughput in vLLM with minimal quality deltas
  • Runs today in vLLM

Proof It Works

Metric Apriel 15B Apriel-H1 30 Δ
Throughput (vLLM) 1× ~2× +2×
MATH500 90 92 +2
GSM8k 97 95 −2
AIME'24 70 65 −5
GPQA-D 59 55 −4
MBPP 86 85 −1
MT-Bench 8.30 8.58 +0.28

Evaluation Results

Apriel-H1 Evaluation Results

How Apriel-H1 Works

Architecture — H1-30

  • Start: Apriel-15B teacher (50 FA layers)
  • Replace least-critical FA layers with Mamba (no KV cache, linear time)
  • Keep 20 FA layers to preserve global patterns

Distillation — 3 steps

  1. Score layer importance (LOO perf drop + MMR distill loss)
  2. Swap low-importance FA → Mamba (MIL-style init from attention)
  3. Stage & gate: H1-25 → H1-27 → H1-30 (… H1-34/37/40) with reverse-KL; ship at best quality/throughput trade
Teacher (50L): [FA][FA][FA][FA][FA][FA][FA][FA][FA][FA] ...
H1-30:         [FA][FA][FA][M ][FA][M ][M ][M ][M ][FA] ...
                ^           ^           ^           ^
              "keep"     "convert"   "convert"    "keep"

Eval Score vs Throughput

Apriel-H1 Family Performance

What's Next

Apriel-H2 roadmap:

  • Advanced mixers: Gated DeltaNet, Kimi Linear Attention
  • Higher efficient-to-full ratios: Search-guided layer placement
  • Stacked optimizations: Sliding window + quantization + sparse attention
  • Target: 5-10× throughput while maintaining reasoning quality

The path forward:

  • From-scratch hybrid training gives the best ceiling
  • Distillation offers practical retrofitting for existing models
  • Both approaches matter for different constraints

Thank You

Apriel-H1: Efficient Reasoning Through Hybrid Architectures

SLAM Lab — ServiceNow

Contact: Torsten Scholak ()

Team: Oleksiy Ostapenko, Luke Kumar, Raymond Li, Denis Kocetkov, Joel Lamy-Poirier