people
tejas bhakta, the reported co-founder who isn't on the yc page, and the three-person team — who they are, what they wrote, and what they claim.
the founders
- tejas bhakta · ceo · bhaktatejas922founder and ceo — verified by the yc directory, his own site ('founder/ceo @morphllm.com (yc s23)'), and every morph hn launch under bhaktatejas922. ex-tesla: 'an ai engineer at tesla in autopilot and energy for 3 years,' preceded by internships on autopilot soc firmware (dojo/hw4) and autopilot ai infra. ucsd b.s. ece (ml specialization), 2019. grew up in his parents' motel and 'built their whole tech stack, including a website that generated $48,000 in yearly revenue when he was 16' — the direct line to the innkeeper pivot.
- vlad lezhnin · co-founderreported co-founder and cto — his linkedin headline reads 'co-founder & cto @ morph' with 15+ years including spotify and specialties in refactoring, ast transforms, and llm-assisted code changes. but the yc page lists only bhakta and no other primary source names him; treat as reported and unverified — early co-founder who departed, informal, or directory lag.
the writing
tejas bhakta writes the thesis in public — the morphllm.com blog corpus, the launch hn threads, and the non-product essays run on the same account that ships the models. quotes are verbatim; dates keep source precision.
- 2024-02-13faster agents · post, morphllm.com/blog · tejas bhaktathe earliest dated post on the morph blog — a year before morphllm.com existed. migrated or backdated; the apply-analysis corpus starts before the brand.
- 2024-02-15how cursor composer and apply work · post, morphllm.com/blog · tejas bhakta'how cursor composer and apply work' — a teardown of cursor's apply internals written while the company was still named something else. the founder was studying the category before he was selling into it.
- 2025-05-02best practices for building coding agents · post, morphllm.com/blog · tejas bhakta'best practices for building coding agents with morph' — the blog's turn toward the agent-builder audience.
- 2025-05-11self hosting · post, morphllm.com/blog · tejas bhakta'self hosting' — the private-deployment offer, posted before the public launch. the enterprise-privacy wedge was there from the start.
- 2025-05-28diffs vs fast apply / morph ❤️ continue · post, morphllm.com/blog · tejas bhaktaa paired drop: 'diffs vs fast apply' argues why a model beats a brittle format; 'morph ❤️ continue' claims the continue integration the same day.
- 2025-06-15morph gets faster · post, morphllm.com/blog · tejas bhakta'morph gets faster' — the speed post. the archived index later displayed the title as '10,500+ tokens per second,' so the page was likely updated after september; the ambiguity is preserved rather than resolved.
- 2025-07-07launch hn: apply ai code edits at 4,500 tokens/sec · thread, hacker news · tejas bhaktathe launch hn thread — morph's most-cited primary document. the pitch, the division-of-labor thesis ('a coding model, not a code merge model'), the bundling defense ('losing 1-2% there is the difference of billions of dollars for them'), and the 'hardcoded snippet → semantic overreach' correction all live here.
We've built a blazing-fast model for applying AI-generated code edits directly into your files at 4,500+ tokens/sec.
This approach was pioneered by Cursor last year, but their models aren't available as APIs—so we built Morph for developers everywhere.
Claude-4 or Gemini still writes the code, we're just the engine that merges it into the original file.
smart diffing is edge case hell and you'll never capture all of them
People love to say Cursor is a Claude wrapper but they miss the reality that Cursor is a fast apply wrapper.
Fast Apply definitely helps with keeping flow state and is a large part of Cursor's success
217 points, 185 comments - 2025-07-14how intelligence self organizes · essay, hacker news · tejas bhakta'how intelligence self organizes' — a non-product essay submitted under bhaktatejas922; the founder's intellectual side-channel next to the launch.
- 2025-07-21we're already post-agi · essay, hacker news · tejas bhakta'we're already post-agi' — the second non-product essay in launch month. the morph founder argues his position in public, on the same account that ships the models.
- 2025-08-22what is morph fast apply? · post, morphllm.com/blog · tejas bhakta'what is morph fast apply?' — the explainer post, written for the audience the launch thread created.
- 2025-09-15we hit 10,500 tokens/sec on b200 · post, morphllm.com/blog · tejas bhakta'we hit 10,500 tokens/sec on b200' — custom cuda kernels plus a specialized speculator, 'roughly 4× faster than the best speeds on cerebras,' under four seconds for complex edits on a 40k-token document. the hn submission drew 2 points.
We just doubled our speculative code edit throughput and hit 10k tok/sec per request!
hn: 2 points - 2025-10everything is models / llms are bad at being forced · essay, morphllm.com/blog · tejas bhakta'everything is models' paired with 'llms are bad at being forced' — the platform thesis: apply is one reflex; agents need a stack of small specialized models. the hn submission drew 1 point.hn: 1 point
- 2026-01-23learning to test code changes with rl · post, morphllm.com/blog · tejas bhaktathe rl-testing pair — 'learning to test code changes with rl' and the browser-verification post — the written preview of glance.
- 2026-03-31warpgrep v2 · post, morphllm.com/blog · tejas bhaktathe warpgrep v2 launch post — '#1 on swe-bench pro'-style retrieval claims, the subagent framing made explicit. (the coverage lane dates the post 2026-02-23; the conflict is preserved.)
- 2026-06optimizing models to be fast at codegen · essay, morphllm.com/blog · tejas bhakta'optimizing models to be fast at codegen' — the research post: train a custom speculator per model, and the serving business (qwen, glm, deepseek, minimax) becomes the point. the most substantive public statement of the inference thesis.
A generic draft gets 1.93x, a specilized one got 100x faster
Chinchilla says ~20 tokens per parameter is compute-optimal, but that assumes training is the cost. For a model you train once and serve billions of times, it isn't, and the optimum slides hard toward small and overtrained
hn: 2 points - 2026-06-23one backbone, many reflexes · post, morphllm.com/blog · tejas bhakta'one backbone, many reflexes' — the multi-head classifier engine behind reflexes: bhakta's tesla data-sampling instincts applied to petabyte-scale agent-trace signals.
the team
- the teamyc lists team size 3 and primary partner tom blomfield. composio's podcast notes describe 'a two-person team' with a gpu bill ~8× salary spend. github contributors observed on morph repos include stayblue, dhruvbhatia0, and tracycam — affiliations unverified. no other named employees found.
- the backersy combinator (s23, standard ~$500k deal per linkedin's listing) is the only verified institutional money — neuronfeed adds transpose platform as a january-2025 co-investor, conflicting with the june-2023 date. no form d exists; tracxn lists the company as unfunded; the $19m seed in aggregators is morph l2's round misattributed.
interviews and talks
- 2026-02-09the infra pod · the infra podthe infra pod — 'coding agents need infra to apply code changes.' the full origin account: the three-or-four yc pivots, the february-2025 morph start, output ~70–80% identical to input, lazy-sketch prompting alone ~30–40% faster end-to-end, and the unverified ~4×-cerebras claim.
- 2026-03-27the start podcast · startthe start podcast — the positioning in one line, plus the chapter title 'you can't scale human attention 100x.' cites apply ~10k tok/s, ~100k context, and compact at 33k tok/s.
- 2025-12the composio interview · composiothe composio interview notes — 'a two-person team… gpu bill that's 8x their salary spend' and the claim that 'doubling speed roughly doubles conversions.'
- undatedthe second spotify episode · spotifya second spotify episode exists in the catalog; its details are unverified beyond the listing.
the philosophy
themes tracked across the writing: agents over bigger models, the apply thesis, edge-case models, evals at tesla scale, latency as ux, small models as tools, speculative decoding as ip.
- 2023–2026the co-founder who isn't on the yc pagevlad lezhnin's linkedin reads 'co-founder & cto @ morph' with 15+ years including spotify; the yc page lists only bhakta — duplicated twice, a directory artifact — and no other primary source names him.unresolved — an early co-founder who departed, an informal title, or a directory lag. the dossier carries him as reported and keeps him off the verified founder line.
- 2023–2025four names, one companythe entity trail — lingtual → innkeeper → autoinfra, inc. → morph — rests on yc's aka list, third-party trackers, and a copyright footer; no incorporation records were obtained. founding-date fields conflict: the s23 batch says 2023, the yc directory field says 2025.the dossier reads entity versus brand: the company dates to the s23 batch under earlier names; morph the product dates to ~february 2025. the official answer stays open.
- 2019–2026the single-sourced origin storythe founder lore is thin-sourced: the motel upbringing and the '$48,000 in yearly revenue when he was 16' website rest on a single beamstart profile; the tesla tenure is corroborated across linkedin and his own site, but the details (dojo/hw4 firmware, ai infra) come from the same one page.plausible and internally consistent — the motel line explains innkeeper — but the corpus keeps the lore at reported confidence, not confirmed.
AI-drafted at Ben Guo's direct request and credited to Hraness; every claim links to its cataloged source.