K2 Horizon 375B vs GLM 5.2 vs DeepSeek V4 Pro 2026
The Short Answer
Three open flagships, three different arguments. As of September 7, 2026:
| K2-Horizon-375B-A23B | GLM 5.2 | DeepSeek V4 Pro | |
|---|---|---|---|
| Lab | IFM / MBZUAI (Abu Dhabi) | Z.ai | DeepSeek |
| Released | Sep 3, 2026 | 2026 (5.3 weights Aug 28) | latest snapshot Aug 13, 2026 |
| Total params | 375B | — | 1.6T |
| Active/token | 23B | — | 49B |
| Architecture | Sparse MoE | Sparse MoE | Sparse MoE |
| Context | 512K | — | 1,000,000 |
| Max output | — | — | 384K |
| License | Apache 2.0 | open weights | MIT |
| Full training record | promised, not yet shipped | no | no |
The one-line verdict: GLM 5.2 wins the benchmark table, DeepSeek V4 Pro wins context and output length, and K2 Horizon wins on openness — a claim that is partly IOU as of this writing.
Benchmarks: GLM 5.2 Leads, But Not Everywhere
These rows come from IFM’s own model card for the 375B, which is the most useful kind of table: a lab publishing where its model loses.
| Benchmark | K2-Horizon-375B | GLM 5.2 | MiniMax-M3 |
|---|---|---|---|
| GDPVal-AA (Elo) | 1,441 | 1,498 | 1,380 |
| Toolathlon Verified | 65.3 | 59.9 | 53.7 |
| Terminal-Bench 2.1 | 70.2 (audited 66.9) | 77.9 | 65.2 |
| SWE-bench Pro (strict) | 42.6 | 46.7 | 43.8 |
| MCPMark | 67.7 | 72.4 | 48.8 |
| SWE-Atlas-QnA (strict) | 48.4 | 46.4 | 42.3 |
Two things to hold onto.
First, the 70.2 has a footnote. IFM audited its own Terminal-Bench 2.1 run for reward hacking using Artificial Analysis’s procedure with GPT-5.6 Sol as judge, found 24 flagged trials across 10 tasks, and published the corrected figure: 66.9. That widens the gap to GLM 5.2 by another three points. Quote 66.9.
Second, the comparison target is stale by design. IFM benchmarked against GLM 5.2, but GLM 5.3 weights landed August 28, 2026 — before K2 Horizon shipped. Every gap in that table is therefore a floor, and the real distance to the current best open-weight flagship is larger than printed.
DeepSeek V4 Pro is not in IFM’s table. Where it competes is on shape rather than score.
DeepSeek V4 Pro: Context Is the Product
1.6T total / 49B active · 1M context · 384K max output · MIT
DeepSeek’s flagship is roughly 4x the total parameter count of K2’s and activates more than twice as many parameters per token (49B vs 23B). That is a real serving-cost difference, not a spec-sheet flourish: active parameters set your per-token compute and latency floor.
What you buy for it:
- A 1,000,000-token default context, roughly double K2 Horizon’s 512K.
- A 384K maximum output — unusually large, and the reason V4 Pro shows up in long-form generation and whole-repo refactor pipelines.
- MIT license, the most permissive of the three.
There is also DeepSeek V4 Flash at 284B / 13B active with the same 1M context and MIT terms, plus DeepSeek-V4-Flash-Vision-Exp, an experimental multimodal checkpoint published to Hugging Face on August 31, 2026 under MIT. If you want DeepSeek’s context story at a fraction of the memory footprint, Flash is the version to test first.
⚠️ Note on version noise: a “DeepSeek V5” claim circulated in September 2026 with no changelog entry, model string or paper behind it. As of September 7, 2026 the current published line is V4 Pro / V4 Flash. Do not build a plan on V5.
Openness: The Only Category K2 Wins Outright
This is the actual reason to care about K2 Horizon.
| K2 Horizon | GLM 5.2 / 5.3 | DeepSeek V4 | |
|---|---|---|---|
| Final weights | ✅ | ✅ | ✅ |
| Permissive license | ✅ Apache 2.0 | open weights | ✅ MIT |
| Intermediate checkpoints | partial | ❌ | ❌ |
| Training data or recipe | partial | ❌ | ❌ |
| Training code | partial | ❌ | ❌ |
| Mixture composition + logs | partial | ❌ | ❌ |
“Partial” is doing real work in that column. Within the K2 fleet, only the 3.7B and 7B ship the complete artifact set today. The 375B and 36B cards state that intermediate checkpoints, data and training code will be released. The 32B is a Stage 1 checkpoint with the final still pending.
So if reproducibility is your requirement — regulated model provenance, academic replication, a data-lineage audit — K2 Horizon is the only one of the three even attempting it, but the flagship is not yet the model that delivers it. The 7B is.
Serving and Availability
| K2 Horizon | GLM 5.2 | DeepSeek V4 Pro | |
|---|---|---|---|
| Local runtimes | vLLM, SGLang, Ollama day-zero; GGUF builds | vLLM, SGLang | vLLM, SGLang |
| Hardware | NVIDIA, AMD, Cerebras | NVIDIA | NVIDIA |
| Hosted APIs | Compass, Cerebras, AWS, Nebius | multiple | DeepSeek + resellers |
| Published per-token price | none yet | vendor rate | vendor rate |
K2 Horizon’s day-zero Ollama and GGUF support is unusually complete for a launch, and Cerebras deployment is a differentiator if latency matters more than cost. But no K2-specific per-token price was published as of September 7, 2026 — do not read a hosting partner’s generic rate as K2’s price.
How to Choose
| If you need… | Pick |
|---|---|
| Best open-weight agent/coding scores | GLM 5.2 — and test 5.3 |
| 1M context or 384K outputs | DeepSeek V4 Pro (or V4 Flash for cheaper memory) |
| Reproducible training lineage | K2 Horizon — but the 7B, not the 375B |
| Apache 2.0 with a patent grant | K2 Horizon |
| MIT with the fewest strings | DeepSeek V4 |
| A small model that punches up | K2-Horizon-7B (70.6 SWE-bench Verified) |
The honest summary: K2 Horizon’s flagship is not the best open model, and IFM’s own table says so. Its contribution is the training record — and the small models built from it.
Last verified: September 7, 2026.