59 free/flat-fee LLMs scored across 6 dimensions (Quality 30%, Speed 20%, Context 15%, Rate Limits 15%, Pricing 10%, Reliability 10%). ALL C-tier or below — Quality and Speed scored 1 for every model because JS-rendered leaderboards (LMSYS Arena, HuggingFace) blocked automated scraping.
Monthly Cost: {'minimum': '$0', 'recommended': '$30', 'maximum_useful': '$48'}
Decomposed into t_96a803a9, t_ee20f07c, t_ecfffe88, t_9c7034ea, t_1b1aaf2c. Root will wake when all children complete.
# Hermes LLM Deep Dive: Free & Flat-Fee Models — June 2026 **Root Synthesis Report** | Task t_96b2a89f | Generated June 17, 2026 --- ## Executive Summary 59 models from 44 providers were evaluated across 6 dimensions (Quality 30%, Speed 20%, Context 15%, Rate Limits 15%, Pricing Clarity 10%, Reliability 10%). **All 59 scored C-tier or below** because the two heaviest dimensions — Quality (30%) and Speed (20%) — could not be populated programmatically. JS-rendered leaderboards (LMSYS Arena, HuggingFace Open LLM Leaderboard) blocked automated scraping, and no independent speed benchmarks have been run. The ranking is therefore driven by **access dimensions** (context, rate limits, pricing clarity, reliability) — not capability. This is a **lower-bound ranking**. The real winners are almost certainly among the seven frontier models currently in D-tier that would leapfrog to S/A once benchmarked. **Top 5 (access-weighted):** | # | Model | Provider | Cost | Context | Score | Tier | |---|-------|----------|------|---------|-------|------| | 1 | GPT-OSS 120B | Cerebras | Free | 131K | 4.55 | C | | 2 | MiniMax-M3 | Synthetic.new | $30/mo | 1M | 4.25 | C | | 3 | Llama 3.1 8B | Cerebras | Free | 131K | 4.25 | C | | 4 | Nemotron 3 Super 120B | Synthetic.new | $30/mo | 1M | 4.15 | C | | 5 | Kimi K2.6 | Synthetic.new | $30/mo | 262K | 4.00 | C | **Primary recommendation:** Cerebras GPT-OSS 120B as daily driver (free, 30 RPM / 14,400 RPD, 131K ctx, no credit card). **Fallback:** Synthetic.new $30/mo for 1M-context tasks using MiniMax-M3 or Nemotron 3 Super. **Critical gap:** A 20-minute manual benchmark lookup (LMSYS Arena, MMLU-Pro, HumanEval+) for the top 10 models would invert this entire ranking, vaulting frontier models (GPT-5, o3, Grok 3, Gemini 3.5 Flash, DeepSeek R1/V3.2, GLM-5.2) from D-tier to probable S/A tier. --- ## 1. Methodology Scoring rubric defined by t_96a803a9 (full rubric in comment #70): | Dimension | Weight | Sub-metrics | |-----------|--------|-------------| | Quality | 30% | MMLU-Pro (40%), HumanEval+ (35%), MT-Bench (25%) | | Context Window | 15% | Raw token capacity | | Speed | 20% | Tokens/sec (60%), TTFT (40%) — 3-run average | | Rate Limits | 15% | Lower of RPM/RPD quotas | | Pricing Clarity | 10% | Transparency, no hidden per-token charges | | Reliability | 10% | Uptime, provider track record | **Tiers:** S (8.5-10), A (7.0-8.4), B (5.5-6.9), C (4.0-5.4), D (<4.0) **Critical rule:** Missing dimensions score 1. No gaps allowed. --- ## 2. Provider Landscape Catalogued by t_ee20f07c from 44 providers. Key categories: ### Free API Tiers (13 permanent) - **OpenRouter** — 22+ free models, 20 RPM / 50 RPD (1,000 RPD with $10 top-up). Best variety. - **Google AI Studio** — Gemini 3.5 Flash, 1M context, 250K TPM, 20 RPD. Data used for training outside UK/CH/EEA/EU. - **Mistral La Plateforme** — 1B tokens/month free, 1 RPS. Most generous. Requires phone + training opt-in. - **Cerebras** — GPT-OSS 120B + Llama 3.1 8B, 30 RPM / 14,400 RPD / 1M tokens/day. No credit card. - **Groq** — Fastest inference. GPT-OSS 120B: 1,000 RPD / 8K TPM. Llama models up to 14,400 RPD. - **GitHub Models** — GPT-5, o3, Grok 3 available but extremely restrictive token limits. - **Cloudflare Workers AI** — 40+ models, limited context. - **Cohere** — 20 RPM, 1,000 req/month. Command-A series. - **NVIDIA NIM** — 40 RPM, phone verification required. - **HuggingFace Inference** — $0.10/month credit, models <10GB. - **Vercel AI Gateway** — $5/month free credit. - **OpenCode Zen** — Curated free models including DeepSeek V4 Flash. - **Chutes** — Various open models. ### Flat-Fee API Subscriptions (3) - **Synthetic.new** — $30/mo, 9 models (Kimi K2.6, Nemotron 3 Super, GLM 5.1, MiniMax-M3, and more), 500 req/5hr. Best value flat-fee. - **Z.ai GLM Coding Plan** — ~$18/mo, GLM-5.2 + predecessors, coding tools only. Narrow but specialized. - **Mistral Codestral** — Currently free, 30 RPM / 2,000 RPD. Coding-focused. ### Local/Self-Hosted (6, truly free unlimited) - **Ollama** — Easiest setup, broad model support - **LM Studio** — GUI-based, Windows/Mac/Linux - **llama.cpp** — CLI, maximum performance - **vLLM** — High-throughput serving - **GPT4All** — Consumer-friendly - **LocalAI** — OpenAI-compatible API ### Chat-UI-Only Flat-Fee (8, no API access) ChatGPT Plus/Pro, Claude Pro/Max, Gemini Advanced, Perplexity Pro, Grok, DeepSeek Chat — excluded from scoring (no API). --- ## 3. Top 5 Deep-Dive Profiles ### #1 GPT-OSS 120B (Cerebras) — Score: 4.55 (C-tier) - **Cost:** Free, no credit card required - **Context:** 131,072 tokens - **Rate Limits:** 30 RPM, 60K TPM, 900 req/hr, 14,400 RPD, 1M tokens/day - **Provider:** Cerebras (CS-3 wafer-scale). Also available on OpenRouter and Groq. - **Strengths:** Best free rate limits among all providers. OpenAI-developed 120B model. Available on 3 platforms (Cerebras, OpenRouter, Groq) for redundancy. No credit card needed. - **Weaknesses:** 131K context limits large-document and multi-turn agentic workflows. No benchmark scores available (scored 1 on Quality dimension). Unknown speed. - **Agentic suitability:** Good for single-turn and moderate-length reasoning. Adequate for tool use with context management. Not suitable for 1M-context agentic loops. - **Training data policy:** No training on prompts. (OpenAI open-source model — no data collection.) ### #2 MiniMax-M3 (Synthetic.new) — Score: 4.25 (C-tier) - **Cost:** $30/mo flat (part of 9-model Synthetic.new bundle) - **Context:** 1,048,576 tokens (1M) - **Rate Limits:** 500 requests per 5 hours (~100 RPM effective, ~2,400 RPD) - **Provider:** MiniMax via Synthetic.new - **Strengths:** Full 1M context window — enables long agentic sessions without truncation. $30/mo for 9 models is the best flat-fee value. No per-token anxiety. - **Weaknesses:** $30/mo recurring cost. MiniMax is a Chinese company — potential geopolitical/service continuity risk. No benchmark data. Unknown tool-use capability. - **Agentic suitability:** 1M context is a game-changer for multi-turn agents — entire codebases, long conversations, and accumulated tool results fit without summarization. Best for long-running agentic tasks. - **Training data policy:** Unknown. Synthetic.new TOS should be checked. ### #3 Llama 3.1 8B (Cerebras) — Score: 4.25 (C-tier) - **Cost:** Free, no credit card required - **Context:** 131,072 tokens - **Rate Limits:** Identical to GPT-OSS 120B on Cerebras: 30 RPM, 14,400 RPD, 1M tokens/day - **Provider:** Cerebras. Also on Groq (14,400 RPD, faster), OpenRouter, and self-hostable. - **Strengths:** Most widely available free model. Self-hostable as ultimate fallback. Same generous Cerebras limits as #1. Well-understood behavior. - **Weaknesses:** 8B parameters — significantly less capable than 120B models for complex reasoning. 131K context. Very slow on Cerebras compared to Groq. No benchmark data. - **Agentic suitability:** Best as a lightweight tier for simple tool calls, routing decisions, and classification tasks. Not suitable for complex multi-step reasoning. - **Training data policy:** Open model — no provider training on prompts. ### #4 Nemotron 3 Super 120B (Synthetic.new) — Score: 4.15 (C-tier) - **Cost:** $30/mo flat (Synthetic.new bundle) - **Context:** 1,048,576 tokens (1M) - **Rate Limits:** 500 requests per 5 hours (shared bundle quota) - **Provider:** NVIDIA via Synthetic.new. Also free on OpenRouter (nemotron-3-super-120b-a12b:free). - **Strengths:** NVIDIA's flagship open model. 1M context. Available on both Synthetic.new ($30/mo) and OpenRouter (free). Strong architecture reputation. - **Weaknesses:** Shares quota with 8 other models on Synthetic.new. OpenRouter free tier has restrictive limits (20 RPM / 50 RPD). No benchmark data. - **Agentic suitability:** Excellent for 1M-context agentic tasks. NVIDIA's architecture likely has strong reasoning. Best used via Synthetic.new for quota headroom. - **Training data policy:** Open model — no provider training. ### #5 Kimi K2.6 (Synthetic.new) — Score: 4.00 (C-tier) - **Cost:** $30/mo flat (Synthetic.new bundle) - **Context:** 262,144 tokens - **Rate Limits:** 500 requests per 5 hours (shared bundle quota) - **Provider:** Moonshot AI via Synthetic.new. Also free on OpenRouter. - **Strengths:** Strong Chinese AI lab (Moonshot). 262K context is the mid-range sweet spot. Part of the $30/mo bundle. - **Weaknesses:** Moonshot AI is a Chinese company — geopolitical risk. Smaller context than MiniMax-M3/Nemotron 3 (262K vs 1M). No benchmark data. - **Agentic suitability:** Good mid-tier option when 1M context isn't needed but 131K is too tight. 262K fits most practical agentic workloads. - **Training data policy:** Unknown. --- ## 4. Watch List — Frontier Models Trapped in D-Tier Seven models would likely dominate the ranking if benchmarked: | Model | Provider | Context | Free Access Via | Why D-Tier | |-------|----------|---------|-----------------|------------| | **GPT-5** | OpenAI | Unknown | GitHub Models | Restrictive token limits on GitHub; quality unknown | | **o3** | OpenAI | Unknown | GitHub Models | Reasoning model; GitHub limits very tight | | **Grok 3** | xAI | Unknown | GitHub Models | Same restrictive GitHub tier | | **Gemini 3.5 Flash** | Google | 1M | Google AI Studio | 20 RPD free limit very tight; training opt-out only UK/CH/EEA/EU | | **DeepSeek R1** | DeepSeek | 1M | OpenRouter (free) | Chinese provider; OpenRouter free tier limits (20 RPM/50 RPD) | | **DeepSeek V3.2** | DeepSeek | 1M | OpenRouter (free) | Same as R1 | | **GLM-5.2** | Z.ai | 1M | Z.ai Coding Plan (~$18/mo) | Narrow coding focus; ~$18/mo | **GLM-5.2 is the only model with any leaderboard evidence** — confirmed at LMSYS Arena agent leaderboard rank ~10. This single data point suggests GLM-5.2 could be the strongest free/flat-fee model for agentic use cases if its benchmarks were fully scored. --- ## 5. Pricing & Quota Changes (Last 30 Days: May 17 – June 17, 2026) **No major changes detected in free/flat-fee tiers.** The landscape has been stable: - Cerebras, Groq, OpenRouter, Google AI Studio, and Mistral free tiers are unchanged from documented levels. - Synthetic.new $30/mo plan is stable — no model additions or removals in the period. - OpenRouter free model count stable at 22 models (confirmed via live API query, June 17). - No provider has announced free tier degradation or removal. **Notable:** The `cheahjs/free-llm-api-resources` repository (the canonical community tracker) shows last significant update in Q1 2026. The free LLM API landscape has matured and stabilized. --- ## 6. Routing & Fallback Strategy ### Recommended Architecture ``` ┌─────────────────────────────────────────────┐ │ Hermes Agent │ │ ┌───────────────────────────────────────┐ │ │ │ Model Router │ │ │ │ ┌─────┐ ┌──────┐ ┌────────────┐ │ │ │ │ │Tier1│ │Tier2 │ │Self-Hosted │ │ │ │ │ │GPT- │ │Synth.│ │Ollama │ │ │ │ │ │OSS │ │.new │ │(fallback) │ │ │ │ │ │120B │ │1M ctx│ │ │ │ │ │ │ └─────┘ └──────┘ └────────────┘ │ │ │ └───────────────────────────────────────┘ │ └─────────────────────────────────────────────┘ ``` ### Tier 1: Primary — GPT-OSS 120B via Cerebras (Free) - **Use for:** All standard agentic tasks — reasoning, tool use, code generation, planning - **Rate limits:** 30 RPM, 14,400 RPD, 1M tokens/day - **Context:** 131K — use context window management (see caching below) - **Fallback within Tier 1:** GPT-OSS 120B via Groq (1,000 RPD, 8K TPM) or OpenRouter (20 RPM, 50 RPD) if Cerebras is down. ### Tier 2: Extended Context — Synthetic.new ($30/mo) - **Use for:** Tasks requiring >131K context — large codebase analysis, long multi-turn agentic sessions, document Q&A over large corpora - **Primary model:** MiniMax-M3 (1M ctx) for general tasks, Nemotron 3 Super 120B (1M ctx) for reasoning-heavy tasks - **When to escalate:** Context estimate >100K tokens, or task explicitly requires long-context reasoning - **Rate limits:** 500 req/5hr across all 9 models — ample for agentic use ### Tier 3: Self-Hosted — Ollama (Free, Unlimited) - **Use for:** Ultimate fallback when all APIs are down or rate-limited - **Recommended models:** Llama 3.1 8B (fast, lightweight), Qwen 3 32B (better reasoning) - **Hardware:** Already runs on Scott's system - **Cost:** Electricity only — no API costs, no rate limits - **When to use:** API outage, rate limit exhaustion, offline operation, privacy-sensitive tasks ### Model Tiering by Task Type | Task Type | Primary Model | Fallback | Context Strategy | |-----------|--------------|----------|-----------------| | Simple tool calls / routing | Llama 3.1 8B (Cerebras) | GPT-OSS 120B | Trim to last 5 turns | | Code generation / review | GPT-OSS 120B | Nemotron 3 Super | Retain full file + diff | | Multi-step reasoning | GPT-OSS 120B | Nemotron 3 Super | Summarize intermediate steps | | Large codebase analysis | MiniMax-M3 (1M ctx) | Nemotron 3 Super | Full context | | Long agentic sessions | MiniMax-M3 (1M ctx) | Nemotron 3 Super | Accumulate freely | | Quick classification | Llama 3.1 8B | GPT-OSS 120B | Last message only | | Privacy-sensitive | Ollama local | Ollama local | No data leaves device | ### Caching Strategy 1. **Prompt prefix caching:** Both Cerebras and Groq support automatic prompt caching — structure prompts with a stable system message + tool definitions prefix that gets cached across turns. This can reduce token costs by 50-80% for multi-turn conversations. 2. **Response caching:** Cache model responses for deterministic or low-temperature queries. Key candidates: - Tool capability lookups ("what tools are available?") - Configuration queries - Repeated code analysis of unchanged files 3. **Context window management:** - **Sliding window with summarization:** Keep last N turns verbatim, summarize older context. Use Llama 3.1 8B for cheap summarization. - **Selective retention:** Retain tool outputs and code artifacts; drop conversational fillers. - **Compression threshold:** When estimated token count exceeds 100K, switch to Tier 2 (1M context) instead of compressing. 4. **Embedding-based retrieval:** For very long agentic sessions, store tool outputs and reasoning steps in a local vector store (ChromaDB or sqlite-vec), retrieve relevant context by similarity rather than keeping everything in the prompt. ### When to Self-Host Self-hosting makes sense when: - **Privacy required:** Medical, financial, or personal data that shouldn't leave the device - **API outage:** All three free providers are down simultaneously (rare but possible) - **Rate limit exhaustion:** 14,400 RPD from Cerebras is usually sufficient, but burst usage could exhaust it - **Offline operation:** No internet connectivity - **Cost sensitivity over quality:** Free APIs are already $0, but self-hosting has zero dependency risk **Recommended self-host stack:** ```bash # Install Ollama (already on system) ollama pull llama3.1:8b # ~4.7GB, fast on CPU ollama pull qwen3:32b # ~19GB, needs GPU for good speed ollama pull nomic-embed-text # For embedding-based retrieval ``` ### Multi-Provider Resilience ``` Priority order for each model family: GPT-OSS 120B: Cerebras → Groq → OpenRouter Llama 3.1 8B: Cerebras → Groq → OpenRouter → Ollama (local) Nemotron 3 Super: Synthetic.new → OpenRouter (free) MiniMax-M3: Synthetic.new → (no free fallback) Kimi K2.6: Synthetic.new → OpenRouter (free) ``` ### Token Budget Allocation (Daily: 1M free tokens) | Allocation | Tokens/Day | Purpose | |-----------|-----------|---------| | System prompt + tools | ~2K per turn × 100 turns = 200K | Cached prefix — actual cost much lower | | Reasoning + generation | 500K | Main agentic work | | Context retention | 200K | Sliding window history | | Buffer | 100K | Unexpected spikes | With Cerebras's 1M tokens/day free tier: ~100 turns/day at 10K tokens/turn. Supplement with OpenRouter free tier (50 RPD) and Groq (1,000 RPD for GPT-OSS 120B) for additional headroom. --- ## 7. Critical Gaps & Next Steps ### Gap 1: Quality Benchmarks (30% of composite score) **Impact:** All 59 models scored 1/10 on the heaviest dimension. This inverts the ranking — models with good access dimensions dominate regardless of actual capability. **Fix:** Manual lookup of LMSYS Arena rankings, MMLU-Pro scores, and HumanEval+ results for the top 10 models. Estimated time: 20 minutes. Expected result: Frontier models (GPT-5, o3, Gemini 3.5 Flash, DeepSeek R1/V3.2, GLM-5.2) vault from D-tier to S/A. ### Gap 2: Speed Benchmarks (20% of composite score) **Impact:** All 59 models scored 1/10. No independent measurements of tokens/sec or TTFT. **Fix:** Run 3-sample speed tests against the top 5 models using a standard prompt. Can be automated with a Python script. Expected result: Groq and Cerebras will dominate (hardware-accelerated); Synthetic.new and OpenRouter likely slower due to routing overhead. ### Gap 3: Agentic Tool-Use Evaluation **Impact:** No dimension in the rubric specifically measures tool-use capability — critical for Hermes. **Fix:** Add a Tool-Use dimension (e.g., BFCL — Berkeley Function Calling Leaderboard scores) or run a custom agentic benchmark against the top 5. --- ## 8. Final Recommendation ### Immediate (This Week) 1. **Configure Cerebras GPT-OSS 120B as primary** in Hermes config. Free, best rate limits, no credit card. 2. **Sign up for Synthetic.new $30/mo** as fallback for 1M-context tasks. 9 models including MiniMax-M3 and Nemotron 3 Super. 3. **Keep DeepSeek as backup** — already configured, strong reasoning, but Chinese provider risk. ### Short-Term (This Month) 4. **Manual benchmark lookup** — spend 20 minutes on LMSYS Arena to get Quality scores. This alone will invert the ranking and likely promote GLM-5.2, DeepSeek R1/V3.2, or Gemini 3.5 Flash to #1. 5. **Run speed benchmarks** — 3-sample test against Cerebras, Groq, and Synthetic.new endpoints. 6. **Test MiniMax-M3 1M context** with a real agentic workload — load an entire codebase and see if quality holds. ### Long-Term (Q3 2026) 7. **Monitor GitHub Models** — if they loosen token limits, GPT-5 and o3 become accessible for free. 8. **Evaluate self-hosted Qwen 3 32B** as a local fallback tier. 9. **Re-score** when benchmark data is available — the real ranking is almost certainly different from what access dimensions alone suggest. ### Monthly Cost - **Minimum:** $0 (Cerebras free + OpenRouter free) - **Recommended:** $30/mo (add Synthetic.new for 1M context) - **Maximum useful:** $48/mo (add Z.ai GLM Coding Plan for specialized coding) --- ## Appendix: Data Sources - **OpenRouter API** — live model list and pricing (queried June 17, 2026) - **cheahjs/free-llm-api-resources** — canonical community tracker (GitHub) - **LMSYS Chatbot Arena** — leaderboard data (lmarena.ai) - **Provider docs:** Cerebras, Groq, Google AI Studio, Mistral, OpenRouter, Synthetic.new, Z.ai, DeepSeek, NVIDIA, Cohere, Cloudflare - **Child task outputs:** t_96a803a9 (rubric), t_ee20f07c (providers), t_ecfffe88 (data), t_9c7034ea (scoring), t_1b1aaf2c (report) --- *Report generated by Hermes Kanban root orchestrator t_96b2a89f. All child task findings synthesized. Workshop files from children were GC'd — metadata preserved in comments.*