Model Overview & Capabilities
K2 Horizon 375B A23B is an open-weights Mixture-of-Experts foundation model developed by the Institute of Foundation Models, ranked #15 globally with an Intelligence Index of 31.0%. Routing 23 billion active parameters per token out of 375 billion total parameters, it scores 31.8% on SWE-bench Verified across a 524,000-token context window. Priced at $0.80/1M input and $2.40/1M output ($1.20 blended), it is engineered for private enterprise clusters requiring transparent weights.
K2 Horizon 375B Architectural Specifications
Engineered by the Institute of Foundation Models (IFM), K2 Horizon represents one of the largest openly specified MoE networks evaluated in our test battery. The operational parameters below detail parameter sparsity, hosting requirements, and token costs:
| System Attribute | Verified Telemetry | Engineering Notes |
|---|---|---|
| Developing Lab | Institute of Foundation Models (IFM) | Academic-consortium backed foundation research |
| Global Leaderboard Rank | #15 of 22 Models | 31.0% Overall Intelligence Index |
| Parameter Architecture | 375B Total / 23B Active MoE | 16 expert modules with top-2 gating routing |
| Effective Context Window | 524,000 tokens (~400k words) | RoPE context extension with YaRN scaling factors |
| Max Output Length | 16,384 tokens | Supports full multi-file code diff generation |
| Licensing Framework | Open Weights / Permissive Research Spec | Allows commercial on-premise execution |
| Hosted Input Rate | $0.80 / 1,000,000 tokens | Prompt caching reduces repeat reads to $0.20/1M |
| Hosted Output Rate | $2.40 / 1,000,000 tokens | Cloud gateway endpoint commercial pricing |
| Blended Benchmark Price | $1.20 / 1,000,000 tokens | Standard 3:1 input:output ratio index standard |
| Model Identifiers | k2-horizon-375b-a23b |
OpenRouter: ifm/k2-horizon-375b-a23b |
| Release Checkpoint | August 2026 | Production GA release weights |
K2 Horizon 375B Empirical Benchmark Evaluation
In standardized test suites, K2 Horizon 375B provides an open alternative to corporate mid-tier proprietary systems. On SWE-bench Verified, K2 Horizon achieves 31.8% issue resolution across real GitHub repositories, operating within striking distance of Gemini 3.1 Pro (32.0%) while remaining fully deployable in private sovereign environments. On Aider Polyglot, it scores 32.0%, verifying high consistency across compiled languages like Rust, Go, and C++.
| Evaluation Metric | Score | Rank | Target Capability | Verified Source |
|---|---|---|---|---|
| SWE-bench Verified | 31.8% | #15 | Multi-file repository bug resolution | Artificial Analysis v4.3, Oct 2026 |
| Terminal-Bench 2.0 | 30.5% | #15 | Bash shell command execution & pipes | Terminal-Bench Benchmark, Oct 2026 |
| Aider Polyglot | 32.0% | #15 | Multi-language editing benchmark | Aider Leaderboard, Oct 2026 |
| LiveCodeBench | 28.5% | #15 | Uncontaminated algorithmic programming | LiveCodeBench Evaluator, Oct 2026 |
| GPQA Diamond | 46.0% | #15 | PhD-level scientific cross-domain tests | Artificial Analysis v4.3, Oct 2026 |
| SciCode | 32.5% | #15 | Scientific numerical modeling scripts | Artificial Analysis v4.3, Oct 2026 |

How K2 Horizon 375B Compares Against Nearest Rivals
In the mid-tier intelligence category, K2 Horizon directly challenges Google’s Gemini 3.1 Pro Preview (#16), Chinese MoE contender MiniMax-M3 (#17), and DeepSeek’s vision-capable DeepSeek V4 Flash Vision (#14). While Gemini 3.1 Pro offers a 2M token context, K2 Horizon offers open-weight deployability and matches Gemini’s SWE-bench score within 0.2 percentage points.
| Model | Rank | Intelligence Index | SWE-bench | Terminal-Bench | Blended $/1M | Context |
|---|---|---|---|---|---|---|
| K2 Horizon 375B A23B | #15 | 31.0% | 31.8% | 30.5% | $1.20 | 524,000 |
| DeepSeek V4 Flash Vision | #14 | 35.0% | 37.0% | 35.0% | $0.47 | 1,000,000 |
| Gemini 3.1 Pro Preview | #16 | 30.0% | 32.0% | 30.0% | $1.40 | 2,000,000 |
| MiniMax-M3 | #17 | 30.0% | 30.6% | 30.0% | $0.60 | 1,000,000 |

Price vs Performance Dynamics
At $1.20 per million blended tokens on public API endpoints, K2 Horizon 375B yields 25.8 Intelligence Index points per dollar. For self-hosted infrastructure, the unit economics become even more pronounced: once the server cluster is amortized, marginal inference cost drops to pure electricity and cooling expenses.
- Sovereign Hosting Cost Offsets: Bypassing external API rate limits and data retention policies provides significant cost certainty for regulated defense, medical, and banking software teams.
- Low Active Routing Overhead: Because only 23B parameters fire during token generation, GPU memory bandwidth requirements resemble those of a compact 30B model rather than a 375B behemoth.
- Index Efficiency Frontier: K2 Horizon clusters squarely within the high-leverage mid-tier bracket of our standardized price-performance scatter distribution.

What the “375B A23B” MoE Notation Means
The 375B A23B notation refers specifically to the model’s parameter routing density. Unlike monolithic dense architectures (such as gpt-oss-120b, where all 120 billion weights activate on every token), K2 Horizon utilizes a Mixture-of-Experts architecture containing 375 billion total parameters distributed across 16 specialized expert feed-forward networks.
During each token generation step, a top-2 gating router directs compute through only two experts totaling 23 billion active parameters. This design preserves the vast factual recall of a 375B model while maintaining the inference velocity and low operating cost of a much smaller system.
Strengths and Weaknesses
Empirical Strengths
- Open Weights Transparency: Freely inspectable parameter tensors for secure air-gapped enterprise deployments.
- 23B Active Compute Efficiency: High token generation throughput despite massive 375B total capacity.
- Competitive Code Pass Rates: 31.8% on SWE-bench Verified and 32.0% on Aider Polyglot.
- 524,000-Token Memory Window: Ingests substantial repository subtrees and complete API documentation manuals.
Architectural Trade-Offs
- Sub-Frontier Complexity: 31.8% pass rate trails frontier models like GPT-6 Astra (55.0%) and Claude Opus 5 (53.0%).
- High Cluster Footprint: Serving the full 375B parameters unquantized requires an 8x H100/A100 GPU node.
- Moderate Algorithmic Rating: 28.5% on LiveCodeBench indicates weaker competitive programming reasoning.
Best For and When to Avoid
Based on our verified empirical evaluations, K2 Horizon 375B excels in specific private development roles:
- Best For: Private VPC & On-Premise Code Intelligence: Organizations legally restricted from streaming proprietary code to commercial US cloud APIs.
- Best For: Repository Semantic Search & Documentation: Indexing multi-module architectures within its 524,000-token attention context.
- Best For: Cost-Predictable CI/CD Review Bots: Executing automated pull request linting at an economical $1.20/1M blended cost.
- Avoid For: Mission-Critical Architecture Refactoring: For zero-regression enterprise bug resolution, invest in Claude Opus 5.
- Avoid For: Single-Workstation Developer Laptops: Cannot fit in memory on consumer laptops; use Muse Glimmer High instead.
How to Access and Use K2 Horizon 375B
K2 Horizon 375B can be accessed via OpenRouter or served locally using modern high-throughput inference engines like vLLM. Here is a verified OpenAI-compatible Python snippet:
from openai import OpenAI
client = OpenAI(
base_url="https://openrouter.ai/api/v1",
api_key="YOUR_OPENROUTER_API_KEY"
)
response = client.chat.completions.create(
model="ifm/k2-horizon-375b-a23b",
temperature=0.2,
max_tokens=4096,
messages=[
{"role": "system", "content": "You are an enterprise systems engineer reviewing distributed message queues."},
{"role": "user", "content": "Analyze potential partition desynchronization risks in this Kafka consumer group rebalance handler."}
]
)
print(response.choices[0].message.content)
Version History and Related Models
K2 Horizon represents the pinnacle of the Institute of Foundation Models research family:
- K2 Horizon 180B (Early 2026): Precursor MoE checkpoint; the 375B release doubles total parameters and expands the context window from 128k to 524k tokens.
- MiniMax-M3: Competing Asian 1M-context model ($0.60/1M blended, 30.0% Index).
- DeepSeek V4 Pro 0813: Frontier open-architecture MoE competitor (38.0% SWE-bench, $0.96/1M blended).
Frequently Asked Questions
What does 375B A23B mean in K2 Horizon?
It designates a Mixture-of-Experts architecture with 375 billion total parameters where exactly 23 billion parameters are actively engaged per token.
What is K2 Horizon’s score on SWE-bench Verified?
K2 Horizon scores 31.8% on SWE-bench Verified, resolving roughly one out of every three real-world GitHub issues end-to-end.
How much does K2 Horizon 375B cost?
Hosted API pricing is $0.80 per million input tokens and $2.40 per million output tokens ($1.20/1M blended).
What is the context window of K2 Horizon?
It features an official 524,000-token context window with up to 16,384 tokens of continuous output generation.
Can K2 Horizon be deployed on private hardware?
Yes. Model weights and architecture specifications are accessible for enterprise self-hosted deployment on dedicated multi-GPU servers.
Sources and Verification Methodology
All telemetry, benchmark pass rates, and pricing parameters were confirmed on October 18, 2026:
- Artificial Analysis Intelligence Index v4.3 (October 2026)
- SWE-bench Verified Leaderboard (October 2026)
- Terminal-Bench 2.0 Evaluation Results (October 2026)
- OpenRouter Telemetry Data & Documentation (no longer listed on OpenRouter) (October 2026)
{
“@context”: “https://schema.org”,
“@graph”: [
{
“@type”: “TechArticle”,
“@id”: “https://vibecoderjournal.com/rankings/k2-horizon-375b-a23b/#article”,
“isPartOf”: {
“@type”: “WebPage”,
“@id”: “https://vibecoderjournal.com/rankings/k2-horizon-375b-a23b/”
},
“headline”: “K2 Horizon 375B Benchmarks & Architecture (Oct 2026)”,
“description”: “Ranked #15 with 375B total / 23B active MoE architecture. Complete empirical benchmark evaluation, $1.20/1M pricing, and specs verified Oct 2026.”,
“author”: {
“@type”: “Person”,
“name”: “Abdullah Zulfiqar”
},
“publisher”: {
“@type”: “Organization”,
“name”: “Vibecoder Journal”,
“url”: “https://vibecoderjournal.com”
},
“datePublished”: “2026-09-20T12:00:00+00:00”,
“dateModified”: “2026-10-18T16:00:00+00:00”,
“image”: “https://vibecoderjournal.com/wp-content/uploads/2026/10/k2-horizon-375b-a23b-benchmarks-oct-2026.webp”
},
{
“@type”: “Dataset”,
“@id”: “https://vibecoderjournal.com/rankings/k2-horizon-375b-a23b/#dataset”,
“name”: “K2 Horizon 375B Benchmark Scores October 2026”,
“description”: “Standardized benchmark evaluations for K2 Horizon 375B A23B including SWE-bench Verified (31.8%), Terminal-Bench 2.0 (30.5%), and Aider Polyglot (32.0%).”,
“creator”: {
“@type”: “Organization”,
“name”: “Vibecoder Journal Benchmark Lab”
},
“variableMeasured”: [
“SWE-bench Verified”,
“Terminal-Bench 2.0”,
“Aider Polyglot”,
“LiveCodeBench”,
“GPQA Diamond”,
“SciCode”
],
“dateModified”: “2026-10-18T16:00:00+00:00”,
“url”: “https://vibecoderjournal.com/rankings/k2-horizon-375b-a23b/”
},
{
“@type”: “BreadcrumbList”,
“@id”: “https://vibecoderjournal.com/rankings/k2-horizon-375b-a23b/#breadcrumbs”,
“itemListElement”: [
{
“@type”: “ListItem”,
“position”: 1,
“name”: “Home”,
“item”: “https://vibecoderjournal.com/”
},
{
“@type”: “ListItem”,
“position”: 2,
“name”: “Rankings”,
“item”: “https://vibecoderjournal.com/rankings/”
},
{
“@type”: “ListItem”,
“position”: 3,
“name”: “K2 Horizon 375B A23B”,
“item”: “https://vibecoderjournal.com/rankings/k2-horizon-375b-a23b/”
}
]
},
{
“@type”: “FAQPage”,
“@id”: “https://vibecoderjournal.com/rankings/k2-horizon-375b-a23b/#faq”,
“mainEntity”: [
{
“@type”: “Question”,
“name”: “What does 375B A23B mean in K2 Horizon?”,
“acceptedAnswer”: {
“@type”: “Answer”,
“text”: “It designates a Mixture-of-Experts architecture with 375 billion total parameters where exactly 23 billion parameters are actively engaged per token.”
}
},
{
“@type”: “Question”,
“name”: “What is K2 Horizon’s score on SWE-bench Verified?”,
“acceptedAnswer”: {
“@type”: “Answer”,
“text”: “K2 Horizon scores 31.8% on SWE-bench Verified, resolving roughly one out of every three real-world GitHub issues end-to-end.”
}
},
{
“@type”: “Question”,
“name”: “How much does K2 Horizon 375B cost?”,
“acceptedAnswer”: {
“@type”: “Answer”,
“text”: “Hosted API pricing is $0.80 per million input tokens and $2.40 per million output tokens ($1.20/1M blended).”
}
},
{
“@type”: “Question”,
“name”: “What is the context window of K2 Horizon?”,
“acceptedAnswer”: {
“@type”: “Answer”,
“text”: “It features an official 524,000-token context window with up to 16,384 tokens of continuous output generation.”
}
},
{
“@type”: “Question”,
“name”: “Can K2 Horizon be deployed on private hardware?”,
“acceptedAnswer”: {
“@type”: “Answer”,
“text”: “Yes. Model weights and architecture specifications are accessible for enterprise self-hosted deployment on dedicated multi-GPU servers.”
}
}
]
}
]
}




