Other notable labs

Strategy has shifted toward open weights: Jamba2 Mini, Jamba2 3B and Jamba Reasoning 3B are Apache 2.0, while the legacy Jamba 1.7 Mini/Large sit under the (more restrictive) Jamba Open Model License and are the only paid API SKUs. Reka AI — natively multimodal (text/image/video/audio) small models: Spark, Edge, Flash, Core. Sells a modest API and has open-weighted reka-flash-3 / 3.1 (21B) and reka-edge-2603 (7B multimodal).

Small vendor; check availability before designing around it. Allen Institute for AI (Ai2) — the fully-open reference point: OLMo 3 / OLMo 3.1 (7B and 32B, Base/Instruct/Think/RL-Zero) trained on the ~9.3T-token Dolma 3 corpus, plus Molmo 2 vision-language models (8B, 4B, and Molmo 2-O 7B built on Olmo). Everything — data, code, weights, intermediate checkpoints — is public.

Pick Ai2 when reproducibility or auditability matters more than raw capability. Caveat: Molmo 2's Apache-2.0 card notes some training data is academic/non-commercial-only. NVIDIA — Nemotron is now a full frontier-scale open-weight family: Nemotron 3 Nano / Super (120B-A12B) / Ultra (550B-A55B) plus the Nemotron 3.5 Lightning 30B-A3B, all LatentMoE architectures interleaving Mamba-2 and MoE layers with Multi-Token Prediction, and context up to 1M tokens.

Licensing has been drifting from the NVIDIA Open Model License toward OpenMDW-1.1 on the newest drops. Weights free; the monetisation is GPUs and NIM/AI Enterprise. IBM — Granite 4.2 (3B, 8B, 30B), Apache 2.0, 128K native context extensible to 512K, with switchable thinking modes (full / non-thinking / low-effort).

Deliberately small and enterprise-boring: strong tool calling, 12 languages, cheap to self-host. Also ships Granite Vision, Speech, Guardian, Embedding and Time Series lines. Liquid AI — LFM2 / LFM2.5, on-device-first hybrid conv+GQA models from 230M to 8.3B-A1B, plus VL and audio variants and the DSpark speculative-decoding drafts.

Released under a bespoke "lfm1.0" licence, not Apache. The clear pick for phone/laptop/NPU deployment. Perplexity — not a foundation-model lab in the usual sense; Sonar is a search-grounded answer API with built-in citations.

You are buying retrieval + synthesis, not a raw model. ServiceNow (Apriel 1.6 15B Thinker, MIT, multimodal, 131K context) and Snowflake (Arctic-AWM 4B/8B/14B tool-use agents, Apache 2.0, built on Qwen3 with agentic RL) are both enterprise-platform vendors open-sourcing narrow, agent-shaped models rather than general chat models.

A Access and limits

This entry aggregates eight independent labs, so access differs per family.

PAID FIRST-PARTY APIs (prices below are real, from official pricing pages):

• AI21 Studio sells only two per-token SKUs: "Jamba Mini" ($0.20/$0.40 per 1M) and "Jamba Large" ($2/$8 per 1M). The newer Jamba2 / Jamba Reasoning 3B open-weight releases are NOT on the AI21 pricing table and are therefore priced -1 here.
• Perplexity Sonar sells per-token AND per-request. Every Sonar model carries a search-request fee on top of tokens: sonar $5-$12 per 1K requests, sonar-pro and sonar-reasoning-pro $6-$14 per 1K requests (fee scales with search_context_size low/medium/high). sonar-deep-research instead bills citation tokens at $2/1M, reasoning tokens at $3/1M, and search queries at $5 per 1K. The token prices in this dataset exclude those request fees, so real cost per call is materially higher than input+output alone. IMPORTANT: Perplexity's docs state Sonar Chat Completions is migrating to the Agent API, with Sonar chat-completions support ending 2026-09-27. "sonar-reasoning" (non-Pro) no longer appears on the current pricing or models pages and appears retired.
• Reka publishes a small API price list (Edge, Flash, Core) at docs.reka.ai/pricing; only reka-flash and reka-edge are documented as "always available for public access", other models may require arrangement.
• IBM watsonx.ai does publish per-1M-token prices, but only for older SKUs (e.g. granite-4h-small at $0.0636 in / $0.265 out per 1M). No watsonx per-token price is published for the current Granite 4.2 models, so those are -1 here rather than a guessed carry-over.

OPEN-WEIGHT, NO FIRST-PARTY PER-TOKEN API (all priced -1 by rule):

• Allen Institute for AI (OLMo 3 / 3.1, Molmo 2) — research release, Apache 2.0, weights + data + checkpoints on Hugging Face. Ai2 runs a free playground, not a metered API.
• NVIDIA Nemotron — weights are free on Hugging Face under NVIDIA Open Model License or OpenMDW-1.1. build.nvidia.com offers hosted NIM endpoints on a credit/NVIDIA AI Enterprise basis, not a published per-model per-token rate.
• IBM Granite 4.2 — Apache 2.0 weights.
• Liquid AI LFM2 / LFM2.5 — weights on Hugging Face under the "lfm1.0" licence; Liquid's commercial motion is the LEAP on-device SDK, not a per-token cloud API.
• ServiceNow Apriel — MIT weights on Hugging Face; consumed inside ServiceNow AI Platform, not sold per token.
• Snowflake Arctic (Arctic-AWM agentic models) — Apache 2.0 research weights; Snowflake Cortex bills in Snowflake credits, not published per-token USD.

Do not substitute a third-party host's price (Together, Fireworks, OpenRouter, Bedrock, Vertex) for any of the -1 entries — those are resellers, not the lab.

C Calling the API

Other notable labs accepts requests in the OpenAI Chat Completions format, so any OpenAI-compatible client works by changing the base URL. Example uses reka-edge.

curl Varies by lab: AI21 https://api.ai21.com/studio/v1 | Perplexity https://api.perplexity.ai | Reka https://api.reka.ai/v1 | NVIDIA NIM https://integrate.api.nvidia.com/v1 | IBM watsonx.ai https://{region}.ml.cloud.ibm.com/ml/v1/chat/completions \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "reka-edge",
    "messages": [{"role": "user", "content": "Hello"}]
  }'

from openai import OpenAI

client = OpenAI(
    base_url="Varies by lab: AI21 https://api.ai21.com/studio/v1 | Perplexity https://api.perplexity.ai | Reka https://api.reka.ai/v1 | NVIDIA NIM https://integrate.api.nvidia.com/v1 | IBM watsonx.ai https://{region}.ml.cloud.ibm.com/ml/v1",
    api_key=os.environ["API_KEY"],
)

response = client.chat.completions.create(
    model="reka-edge",
    messages=[{"role": "user", "content": "Hello"}],
)
print(response.choices[0].message.content)

A compatibility layer is not always the provider's full API — caching, strict tool schemas and structured output are commonly unsupported there. Check the official docs before relying on an advanced feature.

D Official references

• Homepage
• API documentation
• Pricing