K2 Horizon vs DeepSeek V4 Pro, GLM-5.3, Nemotron 3 Ultra

Short answer: K2 Horizon and Nemotron 3 Ultra cost nothing, DeepSeek V4 Pro and GLM-5.3 are paid

LLMTR is a gateway that gives access to language models from different providers through one OpenAI-compatible API. K2 Horizon 375B A23B (ifm/k2-horizon-375b-a23b), Nemotron 3 Ultra 550B A55B (nvidia/nemotron-3-ultra-550b-a55b), DeepSeek V4 Pro (deepseek/deepseek-v4-pro) and GLM-5.3 (zai/glm-5.3) are called here on the /v1/chat/completions endpoint at https://llmtr.com/v1 with the same API key from your llmtr.com account.

According to IFM's model card, K2 Horizon is an Apache 2.0 model with 375 billion total and 23 billion active parameters; it reasons before every answer and is free with a daily allowance of 200,000 tokens until 5 November 2026, 23:59 (TRT). From 6 November 2026, 00:00, it returns 410. The option at no cost with no end date is the Nemotron 3 Ultra row with its daily quota. On the paid side, DeepSeek V4 Pro runs at $0.66 input and $1.98 output per 1M tokens, and GLM-5.3 at $1.26 and $3.96.

• IFM, K2-Horizon-375B-A23B model card · 2026-10-19
• DeepSeek, models and pricing · 2026-10-19

K2 Horizon, DeepSeek V4 Pro, GLM-5.3 or Nemotron 3 Ultra?

Four questions decide it: whether the work runs past 5 November, how sensitive the data you send is, how you want to control reasoning, and how much output you need in one response. The data notice of K2 Horizon and Nemotron 3 Ultra says prompts may be logged by the provider and used to improve the model; do not send sensitive, personal or customer data to them.

The table shows the four rows with the values in the LLMTR catalog and docs; prices are per 1M tokens, in the order input / cached input / output.

What does the same agent step cost on each model?

Take an agent step with a 50,000-token prompt and 8,000 output tokens including reasoning, without a cache hit. On DeepSeek V4 Pro this step costs $0.04884 off-peak and $0.09768 on weekday peak hours. On GLM-5.3 it costs $0.09468. On K2 Horizon it takes nothing from your balance until 5 November 2026; the step uses 58,000 tokens of the daily 200,000-token allowance, so the allowance covers 3 steps of this size a day. On Nemotron 3 Ultra there is no charge and requests count toward the daily quota. If 40,000 tokens of the prompt are read from the cache, the same step drops to $0.02332 off-peak on DeepSeek V4 Pro and to $0.05364 on GLM-5.3; K2 Horizon and Nemotron 3 Ultra list no cache discount.

The Python example below sends the same planning prompt to the four models and prints finish_reason, the token counts and the cost. reasoning_effort goes only to the two models that accept it, K2 Horizon and GLM-5.3, and both are set to high. Try it with a few steps from your real workflow: the token count changes from model to model.

OpenAI Python SDK example that sends the same planning prompt to K2 Horizon, Nemotron 3 Ultra, DeepSeek V4 Pro and GLM-5.3 through https://llmtr.com/v1/chat/completions and prints token usage and cost.

import os

from openai import OpenAI

client = OpenAI(base_url="https://llmtr.com/v1", api_key=os.environ["LLMTR_API_KEY"])

# Per 1M tokens: (input, output) on 19 October 2026, without cache.
# K2 Horizon: free through 5 November 2026 23:59 (TRT), then retired.
# DeepSeek V4 Pro: off-peak price; every item doubles on weekday peak hours.
PRICES = {
    "ifm/k2-horizon-375b-a23b": (0.0, 0.0),
    "nvidia/nemotron-3-ultra-550b-a55b": (0.0, 0.0),
    "deepseek/deepseek-v4-pro": (0.66, 1.98),
    "zai/glm-5.3": (1.26, 3.96),
}
EXTRA = {
    "ifm/k2-horizon-375b-a23b": {"reasoning_effort": "high"},
    "zai/glm-5.3": {"reasoning_effort": "high"},
}
messages = [{"role": "user", "content": "Plan the migration of a 40-table PostgreSQL schema to a new service in numbered steps, with a rollback for each step."}]

for model, (inp, out) in PRICES.items():
    response = client.chat.completions.create(
        model=model, messages=messages, max_tokens=16384, **EXTRA.get(model, {})
    )
    usage = response.usage
    cost = (usage.prompt_tokens * inp + usage.completion_tokens * out) / 1_000_000
    print(model, response.choices[0].finish_reason, usage.prompt_tokens, usage.completion_tokens, f"${cost:.6f}")

Reasoning control, output limits and peak hours

K2 Horizon reasons before every answer and this cannot be turned off; reasoning_effort selects medium or high, while low, none and minimal return 400. IFM's model card recommends high. Reasoning tokens count toward max_tokens; if the response comes back with finish_reason: length, raise the limit. K2 Horizon accepts all four forms of tool_choice and supports json_schema in response_format.

On GLM-5.3 reasoning also runs on every answer; its depth is set with low, high and max, and low applies when the field is omitted. DeepSeek V4 Pro bills every item at double the price on weekdays between 01:00-04:00 and 06:00-10:00 UTC, which is 04:00-07:00 and 09:00-13:00 Türkiye time; weekends and Chinese public holidays are billed at the normal price all day. On the Nemotron 3 Ultra row at no cost, reasoning is off by default and there is no cache discount; once the quota is reached, requests are limited until the next day.

• IFM, K2-Horizon-375B-A23B model card · 2026-10-19

Which one when?

The choices below follow from the differences in duration, price, data notice and parameters between the four models; compare output quality with your own tasks.

• Trying an agent at no cost until 5 November 2026 that reasons on every answer and needs required or named tool_choice and JSON Schema: K2 Horizon.
• An option at no cost with no end date, a 1,000,000-token context and long-document work where reasoning starts off: Nemotron 3 Ultra.
• Paid, up to 393,216 output tokens in one response, and batch work you can schedule outside peak hours: DeepSeek V4 Pro.
• Paid code and agent work where you want to set reasoning depth with low, high and max: GLM-5.3.
• Sensitive, personal or customer data: do not use K2 Horizon or Nemotron 3 Ultra; their data notices say such content may be used to improve the model.

Trying all four with the same key

Open an account on llmtr.com and create a key under Dashboard > API Keys. K2 Horizon and Nemotron 3 Ultra run without charging your balance; for DeepSeek V4 Pro and GLM-5.3, top up your balance under Dashboard > Billing. The same key works on all four models, and you do not need a separate account with each of the four makers.

Move a flow you built on K2 Horizon to another model before 6 November 2026: from that date requests return 410 and are not redirected to another model. When you move it, change the reasoning_effort value too; each model accepts different levels, and a level it does not list returns 400.

Comparing four reasoning models on the same task

Send the same task to K2 Horizon, Nemotron 3 Ultra, DeepSeek V4 Pro and GLM-5.3 with one LLMTR key, then compare the result and the cost.

1. Key. Create an API key on llmtr.com; top up your balance for the two paid models.
2. Same task. Send your own tasks with the four model ids to the https://llmtr.com/v1/chat/completions endpoint.
3. Choice. Compare the results, token counts, cost and the data notice, then choose the model id.

Frequently asked questions

What happens to K2 Horizon after 6 November 2026?

Requests are answered with 410 model_retired and are not redirected to another model. To keep going at no cost choose Nemotron 3 Ultra; for a paid model choose DeepSeek V4 Pro or GLM-5.3, and change the model id.

When does DeepSeek V4 Pro cost double?

On weekdays between 01:00-04:00 and 06:00-10:00 UTC, which is 04:00-07:00 and 09:00-13:00 Türkiye time. In those hours input, cache and output double; weekends and Chinese public holidays are billed at the normal price.

Can reasoning be turned off on K2 Horizon?

No. reasoning_effort selects medium or high; low, none and minimal return 400. Reasoning tokens count toward max_tokens and toward the daily allowance.

Related posts

• The K2 Horizon API: a 375B-parameter open-weight model
• reasoning_effort and max_tokens on K2 Horizon
• Nemotron 3 Ultra 550B A55B 262K: NVIDIA's largest open-weight MoE model
• GLM-5.3: reasoning can't be turned off, choosing between low, high, and max
• Jais 3 or Claude Haiku 5.5 for Arabic?

• K2 Horizon 375B A23B model
• Nemotron 3 Ultra 550B A55B model
• DeepSeek V4 Pro model
• GLM-5.3 model
• IFM Jais documentation
• Z.AI thinking mode
• DeepSeek peak-hour pricing