The Known Good Updated 25 Jul 2026
DeepSeek: R1 Distill Llama 70B vs Microsoft: Phi 4 ·DeepSeek vs Microsoft

DeepSeek: R1 Distill Llama 70B vs Microsoft: Phi 4 — head-to-head comparison

DeepSeek: R1 Distill Llama 70B profile Microsoft: Phi 4 profile All comparisons
DeepSeek: R1 Distill Llama 70B — DeepSeek
Quality
#41 of 61
52based on 3 of 6 evaluations
Known Good Index
Speed
#264 of 316
24.3
Output tokens per second
Price
#176 of 340
$0.80
Blended per 1M tokens
In $0.80 · Out $0.80
Cache hit price
unranked
—
per 1M tokens
Context
#338 of 345
8k
tokens (~13 pages)
Microsoft: Phi 4 — Microsoft
Quality
#50 of 61
33based on 3 of 6 evaluations
Known Good Index
Speed
#142 of 316
56.1
Output tokens per second
Price
#40 of 340
$0.09
Blended per 1M tokens
In $0.07 · Out $0.14
Cache hit price
unranked
—
per 1M tokens
Context
#334 of 345
16k
tokens (~25 pages)
Summary

On the Known Good Index, DeepSeek: R1 Distill Llama 70B scores 52 and Microsoft: Phi 4 scores 33 — a 20-point lead for DeepSeek: R1 Distill Llama 70B. They share 10 published evaluations, with DeepSeek: R1 Distill Llama 70B ahead on 8 and Microsoft: Phi 4 ahead on 2.

Blended pricing is $0.80 per 1M tokens for DeepSeek: R1 Distill Llama 70B and $0.09 for Microsoft: Phi 4 — Microsoft: Phi 4 is about 9.1× cheaper.

Throughput is 24 tokens per second for DeepSeek: R1 Distill Llama 70B and 56 for Microsoft: Phi 4, with time to first token at 0.8s and 0.2s. DeepSeek: R1 Distill Llama 70B accepts 8k tokens of context against 16k.

DeepSeek: R1 Distill Llama 70B is open weights under Open weights (unrestricted); Microsoft: Phi 4 is open weights under Open weights (unrestricted). 2 tracked hosts serve one or both.

Generated from tracked data on 25 Jul 2026 · how these comparisons are computed →

At a glance
MetricDeepSeek: R1 Distill Llama 70BMicrosoft: Phi 4
Known Good Index 52 based on 3 of 6 evaluations 33 based on 3 of 6 evaluations
Blended $/1M $0.80 $0.09
Output tok/s 24 56
Context window 8k 16k
Providers 1 1
Jump to section

Quality

Known Good Index and the capability indices, side by side

IndicesComponents
Index scores

Each index is min-max normalised across tracked models and scaled 0–100 from published evaluations. We do not run these evaluations.

MetricDeepSeek: R1 Distill Llama 70BMicrosoft: Phi 4
Known Good Index 52 33
Mathematics Index 71 39
Science Index 52 53
Reasoning Index 59 47
Openness Index 60 60
The Known Good

Quality Breakdown

Every evaluation both models report a published score for

Evaluations both models report

Bar = DeepSeek: R1 Distill Llama 70B · marker = Microsoft: Phi 4 · scores ingested from Epoch AI (CC-BY 4.0), LiveBench (Apache-2.0)

GPQA Diamond56 / 56
MATH / AIME51 / 14
LiveBench54 / 42
LiveBench Coding52 / 31
LiveBench Data Analysis56 / 45
LiveBench Instruction Following70 / 58
LiveBench Language24 / 26
LiveBench Mathematics58 / 42
LiveBench Reasoning68 / 48
MATH Level 590 / 65

Right-hand figures read DeepSeek: R1 Distill Llama 70B / Microsoft: Phi 4. DeepSeek: R1 Distill Llama 70B leads on 8 of 10, Microsoft: Phi 4 on 2.

The Known Good

Arena Elo

Human preference rating from blind pairwise votes

Arena ratings

Elo from blind pairwise votes, ingested from Arena. Higher is better.

MetricDeepSeek: R1 Distill Llama 70BMicrosoft: Phi 4
Text Arena — 1,217
Text Arena (style controlled) — 1,256
The Known Good

Price & Cost

Published API pricing by token type

Per token typeCache discount
Published pricing

USD per 1M tokens. Blended is a 3:1 input:output mix. Live from OpenRouter, cross-checked against models.dev. Lower is better.

MetricDeepSeek: R1 Distill Llama 70BMicrosoft: Phi 4
Input $/1M $0.80 $0.07
Output $/1M $0.80 $0.14
Blended $/1M $0.80 $0.09
Cache read $/1M — —
The Known Good
We publish blended price rather than cost-per-task. Cost-per-task requires token counts from an evaluation suite we do not run. Blended price is computed from published rates and is fully reproducible — see the pricing methodology.

Context Window

Maximum input accepted and maximum output emitted

Context and output limits

Provider-declared limits, from metadata via models.dev and OpenRouter. Higher is better.

MetricDeepSeek: R1 Distill Llama 70BMicrosoft: Phi 4
Context window 8k 16k
Max output tokens 8k 16k
The Known Good

Speed & Latency

Median throughput and time to first token across hosting providers

Output speedTime to first token
Throughput and latency

Medians across all tracked providers, derived from OpenRouter endpoint telemetry.

MetricDeepSeek: R1 Distill Llama 70BMicrosoft: Phi 4
Output tokens/s 24 56
Time to first token 0.8s 0.2s
The Known Good

Providers

Which hosts serve each model, and at what price and speed

2 hosts
Full provider benchmarks →
Provider DeepSeek: R1 Distill Llama 70B blendedDeepSeek: R1 Distill Llama 70B tok/s Microsoft: Phi 4 blendedMicrosoft: Phi 4 tok/s
Novita $0.80 19 — —
DeepInfra — — $0.09 70

Per-host pricing and telemetry from OpenRouter. A dash means that host does not serve that model.

The Known Good
Sources for this page
Pricing & providers — OpenRouter Metadata — models.dev Benchmarks — Epoch AI (CC-BY 4.0)Benchmarks — LiveBench (Apache-2.0)Elo — Arena Full attribution →