← TinyGPT · docs · devlog · roadmap · speedup
source: docs/learn/competitive-landscape.md · view on GitHub ↗

Competitive landscape (2026)

Factual map of the players a Mac-first SLM toolkit competes with or positions against. The strategy derived from this lives in docs/sessions/2026-06-13-market-landscape-mac-first.md; this page is the citable evidence. Researched 2026-06-13; numbers move, re-verify before quoting externally.

Fine-tuning platforms

PlayerHostingWedgeMac/MLX?
OpenAI fine-tuningcloudbest base models, zero infra; SFT/DPO + RFTno
Together AIcloudbroad open-model menu, cheap LoRA/full-FTno
Fireworks AIcloudmanaged RFT, PyTorch pedigreeno
Predibase → Rubrikcloud/VPCfirst to productize RFT (acquired Jun 2025)no
Laminicloud + on-prem/air-gapMemory Tuning; enterprise privacyno
Modal / Replicatecloudserverless GPU infra (where FT jobs run)no
CastformcloudRL on agent-trace/RAG envs; export weightsno
Tinker (Thinking Machines)cloudlow-level distributed-LoRA APIno
OpenPipe / ART → CoreWeaveOSS + cloudagent RL trainer (acquired Sep 2025)no
UnslothOSSfastest QLoRA; MPS 3–5× slower, native MLX “coming”partial
Axolotl / TorchTune / LLaMA-FactoryOSSconfig-driven FTCUDA-centric (no 4/8-bit on Mac)
Kiln AIOSS, local UXlocal-first workbench — but delegates training outorchestrates, doesn’t train on-device
MLX-LM (Apple)OSS, localthe native Apple-Silicon LoRA/QLoRA/DoRA pathyes (library/CLI)

Read: the commercial market is one shape — rent our GPUs, send us your data, pay per token or per GPU-hour. Mac-native training as a product is unowned; only Apple’s MLX-LM library + thin wrappers serve it, and Kiln’s local UX delegates the actual training elsewhere.

Agent eval / observability

PlayerHostingWedgeLocal story
Braintrustcloud (+ enterprise self-host)integrated eval + experiment + monitorminimal
LangSmithhybridLangChain-native tracing + evalenterprise tier only
Langfuse → ClickHouseOSS + cloudmost-adopted OSS observability (acquired Jan 2026)strong (self-host)
Arize PhoenixOSS + SaaSOTel-based, self-hostablestrong
Galileocloudguardrail models + “Insights” root-causeno
Patronus AIcloudproprietary eval models (Lynx/GLIDER/Percival)no
Humanloop → Anthropicdead as standalone (acqui-hire Aug 2025)
Promptfoo → OpenAIOSS CLIeval + red-team, local by default (acquired Mar 2026)strong
Comet OpikOSS + cloudApache-2.0 tracing/evalsolid (self-host)
W&B Weave → CoreWeavecloudone-line tracing + eval dashboardslimited
DeepEval (Confident AI)OSS + cloud50+ metrics, local-firststrong (VPC self-host)
RagasOSS libreference-free RAG metricsruns anywhere

Read: “self-host” here means your K8s/VPC, not your Mac. True on-device eval is a gap, but local-eval alone is commoditizing (Promptfoo, DeepEval, Langfuse, Phoenix all do it). The bigger gap is mechanistic “why did it do that” — every “root cause” feature is just an LLM summarizing traces, not model internals.

Interpretability tooling

PlayerState
Goodfire (Ember)public API → partnership-only (Feb 2026); not self-serve
NeuronpediaOSS, research/safety-funded; closest to “productized” interp
TransformerLens / SAELensactively-maintained research libraries; no product
Anthropic interpretabilitystays research; not a sold feature

Read: nobody sells activation patching / logit lens / SAEs as a paid agent-debugging feature. The eval community and the interp community barely overlap.

The whitespace (one line each)

  1. Mac-first training as a product — owned by a library (MLX-LM), not a product. → TinyGPT’s B6 + B31.
  2. Eval + interp + local, fused — category-of-one; nobody combines all three. → TinyGPT already ships the fusion.
  3. Academic agent benchmarks as a local CI gate — BFCL/τ-bench are leaderboards, not products. → TinyGPT wrapped both (E1/E2); reframe as a workflow primitive (B32).

Consolidation (the market is being rolled up)

Predibase→Rubrik · W&B Weave→CoreWeave · OpenPipe→CoreWeave · Humanloop→Anthropic · Langfuse→ClickHouse · Promptfoo→OpenAI · Goodfire→partnership-only. The buyers are GPU-infra companies and frontier labs. A $0-marginal-cost + data-stays-on-device + OSS-inspectable tool is structurally outside that roll-up — no GPU meter to acquire, no ingestion revenue to absorb.