K2 Horizon 375B A23B
By Institute of Foundation Models
Released: 2026-09-03
K2 Horizon 375B A23B is Institute of Foundation Models's open-weight AI model, released 03/09/2026. IFM's Hugging Face card for the flagship K2 Horizon (375B total, 23B active, Apache 2.0, 512K context) prints a chart of 18 benchmarks against four open-weight models (Nemotron 3 Ultra, Inkling xhigh, MiniMax-M3, GLM 5.2 max) and three closed ones (GPT 5.6 Luna, GPT 5.6 Terra, Claude Sonnet 5). It tops only SWE-Atlas-QnA and trails the best closed model on most rows. The card states no release date; 3 September 2026 comes from MBZUAI's launch announcement as cited in our article. Recommended settings are reasoning_effort high, temperature 1.0, top_p 0.95.
Visit K2 Horizon 375B A23BAI-Powered
Leverages advanced AI technology to deliver cutting-edge capabilities and results.
Fast & Efficient
Optimized performance ensures quick results without compromising on quality.
Purpose-Built
Specifically designed for llms tasks and workflows.
How does K2 Horizon 375B-A23B score on independent benchmarks?
K2 Horizon 375B-A23B ranks #26 of 28 models on the Artificial Analysis Intelligence Index (v5.0.1) with a score of 17.8, as measured on 22 November 2026.
Where does K2 Horizon 375B-A23B rank against other models?
How does K2 Horizon 375B-A23B do on individual tests?
| Evaluation (run by Artificial Analysis) | K2 Horizon 375B-A23B | Best of any model |
|---|---|---|
| Terminal-Bench 4.0 | 1.5% | 67.2% |
| FrontierCode 2.0 | 11.8% | 48.8% |
| Humanity's Last Exam | 32% | 74.3% |
| GDPval-AA (Elo) | 1363 | 1866 |
How has K2 Horizon 375B-A23B's score changed over time?
We have tracked K2 Horizon 375B-A23B on Intelligence Index v5.0.1 since 22/11/2026. A point is added each day the score, coding index or price changes, so this chart fills in over time.
What does the K2 Horizon 375B-A23B system card report?
IFM's Hugging Face card for the flagship K2 Horizon (375B total, 23B active, Apache 2.0, 512K context) prints a chart of 18 benchmarks against four open-weight models (Nemotron 3 Ultra, Inkling xhigh, MiniMax-M3, GLM 5.2 max) and three closed ones (GPT 5.6 Luna, GPT 5.6 Terra, Claude Sonnet 5). It tops only SWE-Atlas-QnA and trails the best closed model on most rows. The card states no release date; 3 September 2026 comes from MBZUAI's launch announcement as cited in our article. Recommended settings are reasoning_effort high, temperature 1.0, top_p 0.95.
These are IFM (Institute of Foundation Models, MBZUAI)'s own published results (self-reported, with the vendor's harness and settings), so they are not directly comparable with other labs' numbers. Source: IFM/K2-Horizon-375B-A23B on Hugging Face. Last checked 22/11/2026.
Agents & tool use
| Benchmark | K2 Horizon 375B-A23B | Nemotron 3 Ultra | Inkling (xhigh) | MiniMax-M3 |
|---|---|---|---|---|
| GDPval-AA | 1441 Elo | 1162 | 1234 | 1380 |
| tau3-Banking | 34% | 14.2% | 29.1% | 15.3% |
| Toolathlon Verified | 65.3% | 34.3% | 45.5% | 53.7% |
| Automation Bench Public | 25.3% | 8% | 12.8% | 20.5% |
| Apex-Agents (pass@1)Subset of text-only tasks | 24.8% | 9% | 19% | 23.8% |
| MCPMark | 67.7% | 45.7% | 51.2% | 48.8% |
| BrowseCompDiscard-all@95k context setting; other models may use different harnesses | 72.8% | 44.4% | 77.1% | 83.5% |
| WildClawBenchSubset of text-only tasks | 50.9% | 34.2% | 52.3% | 56.4% |
Coding
| Benchmark | K2 Horizon 375B-A23B | Nemotron 3 Ultra | Inkling (xhigh) | MiniMax-M3 |
|---|---|---|---|---|
| Terminal-Bench 2.1Figure printed on the card chart; IFM later reported 66.9% after a reward-hacking audit (reported in our article, not on the card) | 70.2% | 53.9% | 55.1% | 65.2% |
| SciCode | 42.7% | 39.9% | 46.1% | 45.4% |
| SWE-Atlas-QnA (strict)Strict mode, no internet access | 48.4% | n/a | 25.5% | 42.3% |
| SWE Bench Pro (strict)Strict mode, no internet access | 42.6% | 38.7% | 43.1% | 43.8% |
Knowledge & science
| Benchmark | K2 Horizon 375B-A23B | Nemotron 3 Ultra | Inkling (xhigh) | MiniMax-M3 |
|---|---|---|---|---|
| Humanity's Last Exam (no tools) | 32% | 28.4% | 31.9% | 39% |
| GPQA Diamond | 87.3% | 86.7% | 87.2% | 92.9% |
| CritPt | 8.6% | 3.1% | 5.4% | 3.7% |
| AA-Omniscience Accuracy | 23% | 23% | 42% | 17% |
Long context
| Benchmark | K2 Horizon 375B-A23B | Nemotron 3 Ultra | Inkling (xhigh) | MiniMax-M3 |
|---|---|---|---|---|
| AA-LCR | 76% | 71% | 73.3% | 80.3% |
Safety & honesty
| Benchmark | K2 Horizon 375B-A23B | Nemotron 3 Ultra | Inkling (xhigh) | MiniMax-M3 |
|---|---|---|---|---|
| AA-Omniscience Non-Hallucination | 74.7% | 70% | 32% | 82% |
What did safety testing find for K2 Horizon 375B-A23B?
Safety level: Not published on the model card
Independent figures: Artificial Analysis, refreshed daily on our AI model benchmarks page. Full analysis: our K2 Horizon 375B-A23B deep dive.
Institute of Foundation Models Model Timeline
512K tokens (524,288) context
Specifications
Related Tools
Claude Opus 5.5
Claude Opus 5.5, launched 22 September 2026, is Anthropic's current Opus-line flagship, replacing Opus 5. It performs at roughly the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5, at $4 input / $20 output per million tokens (both 20% down, cache reads cut 60% to $0.20). It leads on Terminal-Bench 4.0, GDPval-AA v2.1 and Humanity's Last Exam, and is the first Opus model to ship with Fable-5.1-class cybersecurity, biology and distillation safeguards.
Claude Sonnet 5.5
Claude Sonnet 5.5, released 28 September 2026, is the second model in Anthropic's Claude 5.5 family and the faster, lower-cost complement to Claude Opus 5.5. It keeps Sonnet 5's $2 input / $10 output per million tokens ($0.20 cache reads) while generating output over 30% faster, and scores 70.6% on Terminal-Bench 4.0, above Opus 5.5's 66.4%. It has a 1M-token context window, 128K max output, adaptive thinking, and is the first Sonnet with cyber and anti-distillation classifiers.
GPT-6.1 Sol
GPT-6.1 Sol, launched at OpenAI DevDay on 29 September 2026, is OpenAI's mid-tier workhorse model. OpenAI says it delivers nearly the same intelligence as GPT-6 Astra at one-fifth of Astra's token price: $2 input / $10 output per million tokens with cached input cut to $0.10. It scores 75.2% on DeepSWE v1.1 (high effort) and 71.4% on OSWorld 2.0, has a 1.05M-token context window and 128K max output, and is available in the API plus ChatGPT Work and Codex.
