CrofAI

CrofAI (ai.nahcrof.com) is a hosted LLM inference provider that markets itself as "the cheapest inference provider for large models." It serves a curated catalog of open-weights models — DeepSeek V4, GLM 5 / 5.1, Kimi K2.5 / K2.6, Mimo V2.5, Qwen 3.5 / 3.6, Gemma 4, MiniMax M2.5, and a model called greg — over what looks like an OpenAI-compatible API (the public pricing page is the only documented surface at ingest time; auth-gated endpoints handle requests). Three things make it interesting versus the bigger providers:

1. Quantization is disclosed. Every model in the table lists its weight format: Q4_0, Q4_K_M, Q6_K, Q8_0, int4, fp8, awq. Most large providers hide this detail. CrofAI exposes it as a first-class column. See llm-quantization for why Q4_0 vs Q4_K_M vs UD-Q4_K_XL matters.
2. Two billing modes, not stacked. The Free tier is pure pay-as-you-go at the listed per-token rates ("Pay only for usage. No recurring monthly charge"). The paid tiers ($5–$100/mo) are flat subscriptions that buy a daily request quota (500–15,000 requests/day), with no further token charge implied — the per-token rates appear to be the PAYG / overage price, not an additional charge to subscribers within their quota. Some "precision" variants count as 2 requests against the daily quota. The page doesn't explicitly say what happens past the daily cap.
3. Q4_0 is the baseline. Most models use the more aggressive Q4_0 quantization rather than Q4_K_M. That's measurably worse quality on MoE architectures (see llm-quantization). The "precision" variants — deepseek-v4-pro-precision, mimo-v2.5-pro-precision, glm-5.1-precision, kimi-k2.6-precision — bump to Q8_0 or int4 at 2-3× the price.

Pricing snapshot (2026-05-13)

Full table in the source page. Some entries flag vision (multimodal image input), BETA, or 2x requests.

kimi-k2.5-lightning reporting ~1890 t/s is the standout — that's roughly 10–25× the throughput of comparable Kimi K2 deployments at other providers, almost certainly via heavy speculative-decoding or batched draft tokens. The price ($1.00 / $3.00) reflects the speed tier.

Plans

The subscription tier is a daily-request quota at flat monthly cost — unusual for an LLM API provider, closer to a ChatGPT-style "$X/mo for N messages a day" model. PAYG sits alongside as the per-token alternative. Cost predictability is the obvious upside for high-volume small-prompt traffic (chatbots, classification); long-context workflows are worse off, since each request burns one quota slot regardless of size. The "precision" 2x requests flag is the lever they use to keep heavyweight runs from exhausting low-tier quotas. What the page doesn't explain: whether token billing still applies inside the quota, what happens at overage, and whether the per-token rates change for subscribers versus PAYG.

How it compares

CrofAI is closer to DeepInfra / Together / Fireworks (a direct host of open-weights models) than to routing.run / gomodel / claude-code-router (which route to other providers). It does not aggregate — the catalog is whatever ai.nahcrof.com physically runs.

The Q4_0 baseline is the catch. For coding-agent workloads where Q4_K_M is already considered the floor (see local-llm-16gb-vram-tests and habr-local-llm-quantization-deep-dive), Q4_0 may not match the quality you'd get from DeepSeek's own API at similar cost. For experimentation, vision, or low-cost batch work, the prices are genuinely low.

Limitations / open questions

• Closed-source backend; quantization is disclosed but routing, batching, and KV-cache handling are not
• The /pricing_api endpoint requires auth — public pricing requires browser scrape
• No published track record, customer references, or uptime numbers on the marketing page
• Q4_0 quants on production models (especially MoE) may underperform what's advertised — see llm-quantization
• "Speed" numbers are throughput, not first-token latency
• No public docs page about API shape, auth, rate-limit headers (the pricing page links to a "Documentation" tab requiring sign-in)
• Single-domain operation (ai.nahcrof.com) — no public team, no SOC2 / privacy posture, no SLA

Watchlist reason

On watchlist: small provider with no observable track record, aggressive baseline quantization, single-domain operation. Re-check in 90 days for: third-party usage reports, sustained model catalog updates, whether quality complaints surface on Q4_0 baselines, and whether the "cheapest" pricing holds against DeepInfra / Fireworks / Together repricing rounds.

Cross-references

• llm-quantization — why Q4_0 vs Q4_K_M is a real quality difference
• deepseek — three CrofAI models are DeepSeek-line (V4 Pro, V4 Flash, V3.2)
• mixture-of-experts — most catalog entries are MoE; quant sensitivity is worse for MoE
• speculative-decoding — likely mechanism behind the kimi-k2.5-lightning ~1890 t/s
• routing.run — different shape (router, not host)
• gomodel — gateway over many providers
• claude-code-router — could plausibly target CrofAI as one provider
• crofai-pricing — source snapshot