SiliconFlow Free Models Routing Guide: Nex-N2-Pro, DeepSeek-V4-Flash, and Zero-Cost Fallback Strategies

SiliconFlow runs a growing catalog of models at zero cost per token — no credit card, no billing commitment. As of June 2026, that catalog includes Nex-N2-Pro, a 397B MoE model that trades blows with GPT-5.5 and Claude Opus 4.7 on coding benchmarks, and it costs nothing to call through the API. Combined with budget options like DeepSeek-V4-Flash at $0.13/M input tokens, you can build routing setups where the happy path costs pennies and the fallback costs zero.

Update, 2026-10-10: Nex-N2-Pro has since been retired by its provider and is no longer served on TheRouter; requests for nex-agi/nex-n2-pro return a deprecation error. The closest live option is qwen/qwen3.5-397b-a17b, the Qwen3.5 base model N2 Pro was post-trained from (same 397B-A17B class, different model). The rest of this article is kept as published.

This guide covers the specific free models, their practical limits, and when zero-cost routing makes sense versus when you need to pay. For general SiliconFlow setup — API keys, SDK integration, the full model catalog — see our SiliconFlow API Complete Guide.

30-Second Answer: Which SiliconFlow Models Are Free?

As of June 2026, SiliconFlow offers these models at $0.00 per million tokens (both input and output):

Pricing caveat: SiliconFlow's free tier may change. Nex-N2-Pro launched in early June 2026 and was offered free during its initial availability window. Verify current pricing at siliconflow.com/pricing before building production dependencies on zero-cost access.

The standout is Nex-N2-Pro. It is a MoE model built on Qwen3.5-397B-A17B with 17B active parameters per token. Reported benchmarks include 80.8 on SWE-Bench Verified and 75.3 on Terminal-Bench 2.1 — competitive with closed frontier models on coding tasks. At $0.00/M tokens, it is the most capable free model available through any OpenAI-compatible endpoint we have tested.

How SiliconFlow Free Models Work

OpenAI-compatible means a provider exposes a chat-completions endpoint whose request and response shape matches the OpenAI API contract closely enough that an unmodified OpenAI SDK call works against it after swapping three values: API key, base URL, and model name. The minimum surface in practice is POST /v1/chat/completions with messages, model, and an OpenAI-shaped streaming response.

SiliconFlow's free tier works like this:

1. Sign up at cloud.siliconflow.cn and complete identity verification.
2. Free models become available immediately — no credit card or deposit needed.
3. Rate limits are fixed. Free models have static RPM/TPM caps that do not increase with spending. Paid models offer tiered rate limits that scale with your monthly consumption (L0 through L5).
4. Pro/ prefix = paid version. Some models exist in both free and paid variants. The free version uses the original name (e.g., Qwen/Qwen2.5-7B-Instruct), while the paid version adds a Pro/ prefix (e.g., Pro/Qwen/Qwen2.5-7B-Instruct) with higher rate limits.

Free Model Rate Limits

Free model rate limits are fixed — they do not scale with spending tier. From the SiliconFlow docs:

• Language models (chat): RPM 1,000–10,000, TPM 50,000–5,000,000 (varies by model size)
• Embedding models: RPM 2,000–10,000, TPM 500,000–10,000,000
• Image generation: IPM 2+, IPD 400+

These limits are per-account, not per-key. Exceeding them returns HTTP 429 with a Too Many Requests response. The limits are generous enough for development, testing, and low-traffic production — but not for sustained high-throughput workloads.

Source: SiliconFlow Rate Limits docs, retrieved 2026-06-19.

When Free Models Make Sense

Free models work well for specific use cases:

Development and staging. Point your dev environment at a free model to iterate without burning budget. Nex-N2-Pro's 262K context window and strong coding performance make it a genuinely useful dev model, not just a placeholder.

Low-priority fallback. When your primary model (paid, with guaranteed throughput) returns 429 or a transient error, route overflow to a free model. The user gets a response instead of an error, and you pay nothing for the fallback path. See the architecture section below.

Cost ceiling enforcement. If your budget allows N calls per day at $X per call, route any overflow beyond that budget to a free model. You cap spend while keeping the system responsive.

Prototyping and evaluation. Test prompt templates, eval harnesses, and pipeline architectures against a free model before committing to paid inference.

Embedding and reranking. BAAI's BGE models on SiliconFlow are free and solid for text embedding workflows where you want to avoid per-token charges.

When Free Models Don't Make Sense

Be honest about the limits:

SLA requirements. Free models come with no uptime guarantees. If your application needs 99.9% availability, you need a paid tier with support.

Consistent latency. Free models may queue behind paid traffic during peak hours. We have not independently measured this, but it is a reasonable expectation for any free tier.

Production-critical paths. If a failed response costs you real money (e-commerce checkout, financial workflows, customer-facing chat), the primary path should always be a paid model with guaranteed capacity. Free models belong in the fallback or overflow slot.

Nex-N2-Pro durability. The $0.00 pricing launched alongside the model in June 2026. We have no confirmation this is permanent. Build your architecture so the free model slot can be swapped to a budget-paid model (like DeepSeek-V4-Flash at $0.13/M input) without code changes.

Code Example: Calling a SiliconFlow Free Model

The integration uses the standard OpenAI SDK. Point base_url at SiliconFlow and set the model to a free model ID:

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_SILICONFLOW_KEY",
    base_url="https://api.siliconflow.cn/v1",
)

# Free model — $0.00 per token
response = client.chat.completions.create(
    model="nex-agi/Nex-N2-Pro",
    messages=[
        {"role": "user", "content": "Refactor this function to use async/await."}
    ],
    stream=True,
)

for chunk in response:
    if chunk.choices and chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="", flush=True)

For the global endpoint (lower latency outside China), use https://api.siliconflow.com/v1 instead.

// Node.js / TypeScript
import OpenAI from "openai";

const client = new OpenAI({
  apiKey: process.env.SILICONFLOW_KEY,
  baseURL: "https://api.siliconflow.cn/v1",
});

const stream = await client.chat.completions.create({
  model: "nex-agi/Nex-N2-Pro", // free
  messages: [{ role: "user", content: "Explain the builder pattern." }],
  stream: true,
});

for await (const chunk of stream) {
  process.stdout.write(chunk.choices[0]?.delta?.content ?? "");
}

Zero-Cost Fallback Architecture

The highest-value use of free models is as a cost-floor fallback in a multi-model routing setup. The pattern:

Primary (paid, high throughput)
  └─ on 429 / 5xx / timeout
      └─ Fallback (free, best-effort)

Primary model: A paid model with scaled rate limits — for example, DeepSeek-V4-Pro at $1.60/M input on SiliconFlow, or a DashScope model through Alibaba Cloud Bailian.

Fallback model: nex-agi/Nex-N2-Pro (free, 262K context, strong coding). If Nex-N2-Pro pricing changes, swap to Qwen/Qwen3-8B (also free) or deepseek-ai/DeepSeek-V4-Flash ($0.13/M — near-free).

Here is a minimal Python implementation of this pattern:

from openai import OpenAI
import httpx

primary = OpenAI(
    api_key="YOUR_KEY",
    base_url="https://api.siliconflow.cn/v1",
)
fallback = OpenAI(
    api_key="YOUR_KEY",
    base_url="https://api.siliconflow.cn/v1",
)

def chat_with_fallback(messages: list[dict]) -> str:
    try:
        r = primary.chat.completions.create(
            model="Pro/deepseek-ai/DeepSeek-V4-Pro",
            messages=messages,
            timeout=30,
        )
        return r.choices[0].message.content
    except (httpx.HTTPStatusError, Exception) as e:
        # Fallback to free model on 429, 5xx, or timeout
        r = fallback.chat.completions.create(
            model="nex-agi/Nex-N2-Pro",  # $0.00
            messages=messages,
            timeout=60,
        )
        return r.choices[0].message.content

For production, you would add retry logic, circuit breakers, structured logging, and latency tracking. The point is that the fallback slot costs nothing — every request it absorbs is pure savings.

TheRouter supports provider/model routing and fallback where the live product paths support it. If your stack uses TheRouter, you can configure the free model as a fallback target in your routing rules without writing the retry logic yourself.

Budget-Tier Alternatives When Free Isn't Enough

If free model rate limits are too restrictive but you still want minimal cost, these SiliconFlow models are close to free:

Prices from siliconflow.com/pricing, retrieved 2026-06-19.

A practical three-tier routing chain:

1. Tier 1 (quality): DeepSeek-V4-Pro at $1.60/M input — for production traffic needing the best output.
2. Tier 2 (budget): DeepSeek-V4-Flash at $0.13/M input — for cost-sensitive production traffic.
3. Tier 3 (free): Nex-N2-Pro at $0.00 — for overflow, dev, and non-critical paths.

Nex-N2-Pro: What You Need to Know

Since Nex-N2-Pro is the most capable free model in the SiliconFlow catalog, here is a closer look:

• Architecture: MoE (Mixture of Experts), 397B total parameters, 17B active per token. Built on Qwen3.5-397B-A17B.
• Context window: 262K tokens input, 256K tokens max output.
• Modality: Text + image input, text output.
• License: Apache 2.0 (open weights on Hugging Face).
• Release: June 2, 2026, by Nex AGI.
• SiliconFlow model ID: nex-agi/Nex-N2-Pro

Reported benchmarks (from the Nex AGI model card — company-reported, not independently verified):

These numbers place Nex-N2-Pro in the same range as GPT-5.5 and Claude Opus 4.7 on coding and agentic tasks. The model is particularly strong for agentic coding workflows — terminal execution, multi-step debugging, tool calling. For general chat or creative writing, it may not be the best fit; a model like Qwen3.5-27B or DeepSeek-V3.2 would likely produce more natural prose.

Truthful caveat: Nex-N2-Pro is not currently in TheRouter's model catalog. We cannot verify routing support for this specific model. Check the live models page for current availability.

Production Caveats

Before putting free models in production:

1. Rate limits are hard caps. You cannot raise them by spending more. If you hit the ceiling, you need a paid model or the Pro/ variant.
2. No SLA. Downtime on a free model is not a billing event — it is expected.
3. Pricing may change. Nex-N2-Pro's $0.00 pricing is current as of June 2026 but not guaranteed to persist. Treat it as promotional until confirmed otherwise.
4. Benchmark claims are vendor-reported. We have not independently verified Nex-N2-Pro's SWE-Bench or Terminal-Bench scores. Use your own evaluation suite.
5. Identity verification required. SiliconFlow requires identity verification (SMS or equivalent) before you can use free models. This is per-account, not per-key.
6. The free model list changes. Models enter and exit the free tier. Check siliconflow.com/pricing periodically.

FAQ

Can I use free SiliconFlow models for commercial applications?

Yes, based on SiliconFlow's terms of service and the models' individual licenses. Nex-N2-Pro uses Apache 2.0. However, verify the specific license for each model you use, as they come from different vendors.

What happens when I hit the rate limit on a free model?

You receive an HTTP 429 response. Wait until the rate-limit window resets and retry. SiliconFlow provides a rate limit handling guide with backoff strategies.

Is there a difference between siliconflow.cn and siliconflow.com?

Both point to the same platform. The .cn domain and api.siliconflow.cn endpoint is optimized for China-based traffic. The .com domain and api.siliconflow.com endpoint is the global alternative. Free models are available on both.

Can I use SiliconFlow free models with Cursor, Claude Code, or other coding tools?

Yes, if the tool supports custom OpenAI-compatible endpoints. Set the base URL to https://api.siliconflow.cn/v1, provide your API key, and set the model to nex-agi/Nex-N2-Pro. SiliconFlow has published integration guides for Continue in VS Code and Roo Code.

How does this compare to other free LLM APIs (OpenRouter free tier, Google AI Studio)?

SiliconFlow's free tier stands out because it includes Nex-N2-Pro, a 397B-parameter model — far larger and more capable than the typical free-tier models on other platforms. OpenRouter also offers a free tier for some models, but the model selection and rate limits differ. Google AI Studio offers free Gemini access but with different terms and region availability.

---

Sources: SiliconFlow pricing (retrieved 2026-06-19), SiliconFlow rate limits docs (retrieved 2026-06-19), SiliconFlow quickstart (retrieved 2026-06-19), Nex-N2-Pro model page on SiliconFlow (retrieved 2026-06-19), Nex-N2-Pro on Puter Developer (retrieved 2026-06-19), Nex-N2-Pro analysis on FelloAI (retrieved 2026-06-19).