Unified Model Catalog

Stable Aliases. Dynamic Multi-Provider Routing.

Instead of pinning your code to fragile vendor endpoints, target Attenval model aliases. We route transparently to verified providers with automatic health failover.

attenval-auto

Recommended

Target: DeepSeek V3, Workers AI, Llama 3.3

Analyzes prompt structure and routes to the most cost-effective tier with optimal response depth.

Input Tokens 10 credits / 1k tokens

Output Tokens 30 credits / 1k tokens

Typical Latency ~350ms

Client Request

Attenval Gateway

Auto Router

attenval-fast

Low Latency

Target: Workers AI Edge, Llama 3.1 8B

Engineered for autocomplete, streaming chat bubbles, classification, and summarization tasks.

Input Tokens 4 credits / 1k tokens

Output Tokens 12 credits / 1k tokens

Typical Latency ~180ms

Client Request

Attenval Gateway

Fast Engine

attenval-code

High Accuracy

Target: Qwen 2.5 Coder, DeepSeek Coder

Trained on modern programming languages, AST refactoring, and test suite generation.

Input Tokens 12 credits / 1k tokens

Output Tokens 40 credits / 1k tokens

Typical Latency ~420ms

Client Request

Attenval Gateway

Code Specialist

attenval-smart

Deep Thinking

Target: DeepSeek R1, QwQ 32B

Solves complex algorithmic logic, rigorous math, system design, and competitive coding problems.

Input Tokens 25 credits / 1k tokens

Output Tokens 75 credits / 1k tokens

Typical Latency ~650ms

Client Request

Attenval Gateway

Reasoning Core

Catalog Specifications

Model Alias Deep Dive

Exact rates in integer credits per 1,000 tokens. Zero floating-point micro-billing.

attenval-auto

Attenval Auto Router

Recommended Default

Analyzes prompt length, instruction density, and reasoning requirements to automatically pick the most efficient provider tier without manual switching.

Input / 1k tokens 10 credits

Output / 1k tokens 30 credits

Context Limit 128k tokens

Median Latency 350 ms

Underlying Engine: DeepSeek V3, Workers AI, Llama 3.3 70B

Recommended For: General chat, document Q&A, synthesis, mixed development queries.

attenval-fast

Attenval Fast Engine

Sub-200ms TTFT

Optimized for raw time-to-first-token. Runs on Cloudflare global edge GPU nodes nearest to the client.

Input / 1k tokens 4 credits

Output / 1k tokens 12 credits

Context Limit 64k tokens

Median Latency 180 ms

Underlying Engine: Cloudflare Workers AI Edge, Llama 3.1 8B Instruct

Recommended For: Autocomplete, inline IDE suggestions, classification, semantic routing, real-time chatbots.

attenval-code

Attenval Code Specialist

Code Synthesis

Trained on modern languages, frameworks, and syntax trees. Excels at generating compile-clean code with zero hallucinations.

Input / 1k tokens 12 credits

Output / 1k tokens 40 credits

Context Limit 128k tokens

Median Latency 420 ms

Underlying Engine: Qwen 2.5 Coder 32B, DeepSeek Coder V2.5

Recommended For: Full-stack development, AST refactoring, unit test generation, CI debugging, API schema parsing.

attenval-smart

Attenval Reasoning Core

Deep Thinking

Employs test-time compute and chain-of-thought verification before generating the final answer for complex logical challenges.

Input / 1k tokens 25 credits

Output / 1k tokens 75 credits

Context Limit 128k tokens

Median Latency 650 ms

Underlying Engine: DeepSeek R1, QwQ 32B Preview

Recommended For: Rigorous math, multi-step agentic planning, formal verification, competition programming.

Performance & Accuracy

Benchmark Comparisons

Evaluated across standard coding, reasoning, and latency benchmarks.

Evaluation Benchmark	`attenval-auto`	`attenval-fast`	`attenval-code`	`attenval-smart`
HumanEval (Python Code)	84.2%	68.5%	90.4%	92.1%
GSM8K (Math Reasoning)	89.1%	74.2%	86.5%	95.8%
MMLU (General Knowledge)	82.6%	71.0%	79.2%	88.5%
Time to First Token (TTFT)	320 ms	160 ms	380 ms	580 ms
Context Window	128k	64k	128k	128k

* Benchmarks measured across zero-shot standard evaluations. TTFT represents median Cloudflare edge to client streaming start.